Claude Skill

self-improve

Extract lessons from the current session, or sweep the project's past sessions when asked, and route them to the appropriate knowledge layer (project AGENTS.md, auto memory, existing skills, or new skills). Use when the user asks to "self-improve", "distill this session", "distil

LLM Mart · 0 points · 16 views 0 listing impressions 0 install-command copies
Virus-scanned Reviewed automatically before listing.

Full trust report

Download tobihagemann-turbo-codex_skills_self-improve-13e177e.zip · 13 KB
tobihagemann/turbo 403 31 forks MIT Updated 1d ago
Part of tobihagemann/turbo — 147 skills

Install

skills CLI npx skills add https://github.com/tobihagemann/turbo/tree/main/codex/skills/self-improve
Claude Code claude plugin marketplace add https://llmmart.ai/marketplace.json && claude plugin install tobihagemann-turbo@llmmart
Git git clone https://github.com/tobihagemann/turbo.git

The skills CLI installs just this skill, for any of its supported agents. Claude Code installs the whole tobihagemann/turbo collection as a plugin from our marketplace. Git is the plain clone.

Skill manifest

Self-Improve

Review the current conversation, or the project's past sessions when asked, to extract durable lessons and route each one to the right knowledge layer.

Step 1: Detect Context

Available destinations:

  • Project AGENTS.md — the instruction file in the project root, plus any nested ones in subdirectories. At each level use AGENTS.override.md when one is present, otherwise AGENTS.md: an override replaces that directory's AGENTS.md rather than adding to it, so a lesson written to a shadowed file never loads. A nested file scopes its guidance to that subtree, so a lesson scoped to one subtree belongs in the nearest enclosing file, with the root reserved for project-wide rules.
  • Auto memory — one writable memory target resolved for the canonical project root. Prefer a synced Claude Code project-memory source, then a memory-update target explicitly named by the active harness. If neither resolves, skip this destination.
  • Skills — Project skills at .agents/skills/ (walked from project root down to cwd) and user-installed skills at ~/.agents/skills/ (resolve symlinks)

Resolve the canonical project root as the Git top-level containing the working directory, or the physical working directory outside a repository. Resolve Auto memory in this order:

  1. Synced Claude Code source memory — Read ~/.turbo/config.json. When codex.sharedClaudeAutoMemory is exactly true, treat shared-memory setup as complete for every project and use this branch. Resolve autoMemoryDirectory from Claude Code's managed settings, then project-local .claude/settings.local.json, then the user <Claude config home>/settings.json; exclude checked-in .claude/settings.json, where Claude Code ignores this security-sensitive field. Use the expanded override when set. Otherwise locate the resource directory under ~/.codex/memories/extensions/external_agent_import/resources/ whose scope.json matches the canonical project root, take its directory name as the exact Claude project key, and use projects/<project key>/memory/ under CLAUDE_CONFIG_DIR when set or ~/.claude/. Treat false, a missing or invalid key, and an unreadable ~/.turbo/config.json as not opted in and continue to the harness target.
  2. Harness memory update target — Otherwise use a project-specific memory file or additive note intake that the active harness instructions explicitly name for memory updates.
  3. No target — Otherwise skip Auto memory and route lessons through the remaining destinations.

Treat imported extension resources and the harness's consolidated memory files as read-only. For the resolved target, list the directory when applicable and read its index before routing lessons.

Discover the project instruction files (the root file and any nested ones in subdirectories, resolved through the override rule above) and read them. When those files or the resolved Auto memory target point at a knowledge base the repo maintains, read its index too; it is a documentation source for Step 3 rather than a routing destination. List all skill directories but do not read them yet — Step 2 needs to run first so you know what to look for.

Skill Ownership Detection

Classify every skill this session touched:

  • Skills that live in the project are user/project skills
  • If ~/.turbo/repo/ exists, list directories in ~/.turbo/repo/codex/skills/; any skill in ~/.agents/skills/ with a matching directory there is a turbo skill
  • For every other skill in ~/.agents/skills/, read ~/.agents/.skill-lock.json. Its top-level skills object is keyed by skill name, and a skill listed there was installed from a source that replaces it wholesale on its next update. Match on the resolved path rather than the name alone, against the entry's skillPath, so a local fork that replaced the installed copy is not mistaken for it. A skill that matches is package-managed. The signal runs one way: absence from the file leaves ownership genuinely open, so carry an unlisted skill into Step 4 as ownership-unresolved

Verification rule (mandatory before routing in Step 4): For every candidate skill that is about to be routed as turbo, confirm with a fresh test -d ~/.turbo/repo/codex/skills/<name> check that the skill actually lives in the turbo repo. Do not rely on remembered listings from earlier in the session, filename hits in grep output, or assumptions based on where a SKILL.md was read from. A miss here mislabels a user/project skill as turbo, triggers the contribution flow unnecessarily, and can introduce session-specific content into a shared skill — so the check is not optional.

Exception: If the current project IS the turbo repo (i.e., the working directory contains this skill collection), route turbo skill lessons through the Existing user/project skill destination in Step 4 — edits go directly to codex/skills/<name>/ in the project, with no installed-copy indirection and no contribution flow.

Step 2: Gather Session Evidence and Scan for Lessons

Recover Pre-Compaction Evidence

Skip when the conversation is visible in full from the user's own first message.

When it starts from a summary of earlier work instead, recover the compacted turns from the on-disk rollout file. Spawn a single sub-agent (inherited model defaults). The sub-agent's prompt must include:

  1. The absolute path of the project root
  2. A distinctive phrase from the visible conversation, for confirming which rollout file belongs to this session
  3. An instruction to read references/transcript-miner.md for rollout file location, extraction, and output format

Treat the returned items as raw evidence for the scan below.

Sweep Past Sessions

Run when asked to distill sessions beyond the current one. Skip otherwise.

Propose a cutoff first. When Step 1 resolved a harness memory update target, take its newest modification time, state it, then use request_user_input to confirm sweeping from it or sweeping the whole history; when that gate cannot reach the user, sweep from the proposed cutoff and say so in the report. A memory file's timestamp records a write rather than a completed sweep, so it bounds the work without settling what a previous run covered. When Step 1 resolved the synced Claude Code source, take its newest modification time and state that Claude Code writes also advance it, so it may omit unswept Codex sessions; use request_user_input with whole Codex history as the recommended option and that timestamp as the bounded option. When that gate cannot reach the user, sweep the whole Codex history and say so in the report. When no memory location is available, sweep the whole history without asking.

Spawn a single sub-agent (inherited model defaults). The sub-agent's prompt must include:

  1. The absolute path of the project root
  2. The confirmed cutoff as an ISO-8601 timestamp, or that there is none
  3. An instruction to read references/transcript-miner.md and follow its sweep process

Treat the returned items as raw evidence for the scan below.

Identify Session Skills

Before scanning for lessons, identify which skills were loaded during this session:

  • Scan the conversation for Turbo skill invocations and SKILL.md reads from ~/.agents/skills/
  • Build a list of session skills, marking each as turbo, user/project, package-managed, or ownership-unresolved (using the detection from Step 1)
  • This list informs routing in Step 4: when a lesson clearly arose from a specific skill's workflow, that skill is the natural routing target

Scan for Lessons

Scan the full conversation with this priority:

  1. Corrections — Where the user interrupted, said "no", "actually", "stop", "not like that", redirected, or manually fixed something Codex did wrong. Highest-value lessons.
  2. Repeated guidance — Instructions the user gave more than once. Across separate sessions this counts even where each instance reads as ordinary steering on its own.
  3. Skill-shaped knowledge — Domain expertise that was needed repeatedly, tool/API integration details that had to be looked up, decision frameworks that emerged for evaluating options, content templates or writing conventions that were refined, and multi-step workflows where ordering mattered (as reusable domain knowledge, not the workflow itself — see #4).
  4. New workflows — Did this session establish a novel multi-step procedure, coordination pattern, or automation that worked? A successful workflow that would need to be repeated is a prime skill candidate — even if it ran fine this time. Distinct from #3: this captures the procedure itself as a repeatable artifact, not knowledge about how to do it. Flag it.
  5. Preferences — Formatting, naming, style, or tool choices the user expressed.
  6. Failure modes — Approaches that failed, with what worked instead. For tool or script call failures, trace back to the information source that led to the error and route the fix there (e.g., clarify a reference file, update skill instructions, add missing documentation).
  7. Domain knowledge — Facts or conventions Codex needed but did not have.
  8. Improvement opportunities — Out-of-scope improvements noticed during work: code that could be refactored, missing tests, performance issues, readability concerns, or feature ideas that were intentionally skipped to stay focused. Skipped findings count here: when code simplification or code review identified a genuine improvement or issue but it was skipped for this session, route it as a project improvement so it isn't lost.
  9. Trusted reviewer feedback — Human PR review comments that reveal project conventions, patterns, or corrections. Trusted reviewers are repo collaborators with admin or maintain roles (determine via gh api repos/{owner}/{repo}/collaborators --jq '.[] | select(.role_name == "admin" or .role_name == "maintain") | .login'). Their feedback takes precedence over other reviewers and AI bots when there are contradictions.

After scanning, read all skill SKILL.md files (they are small). This gives Step 4 full context for routing.

Step 3: Filter

Keep only lessons that are:

  • Stable — likely to remain true across future sessions
  • Non-obvious — Codex would not already know this
  • Actionable — can be expressed as a rule or instruction
  • Not already documented — absent from the files read in Steps 1 and 2, from each skill's references/ and other supporting files, and from the project's own knowledge stores. Search all of those for each candidate's keywords rather than assuming. A lesson documented only in an unrelated subtree's AGENTS.md still counts as undocumented for the subtree it actually applies to.
  • Still a concern — the issue is not already fixed by changes made in this session. If a bug was found and fixed, or a missing feature was added, future sessions will see the corrected code — they don't need a reminder about the old problem. Exception: successful workflows and procedures are not "resolved" — they're skill candidates precisely because they worked and will need to be repeated. When sweeping past sessions, judge this against the current state of the code and docs rather than against this session's changes.

Discard anything session-specific, speculative, one-off, or already resolved by code changes in this session (but not successful workflows — see exception above). If no lessons survive filtering, tell the user and stop.

Step 4: Route Each Lesson

Assign each surviving lesson to exactly one destination.

Skill-first rule (mandatory): Before consulting the table below, check whether the lesson corrects, refines, or adds a guardrail to any existing skill's behavior — turbo or user/project. This includes lessons about skipping steps, wrong defaults, missing edge cases, or any "don't do X when running $skill-name" correction. If yes, route to that skill. Do not route skill corrections to auto memory or AGENTS.md — they belong in the skill they correct. This rule is not a preference; it is a hard constraint that takes precedence over the table rows below.

Package-managed skills (mandatory): A skill Step 1 classified as package-managed from the lock file routes its lesson straight to Auto memory, recording the skill it applies to, with no question put to the user: the lock file has already settled ownership. A skill under ~/.agents/skills/ that Step 1 left ownership-unresolved may still be one a package manager replaces wholesale on its next update, discarding any edit made here. Before routing a lesson to such a skill, say plainly that an edit to it survives only while the user maintains it themselves. Then use request_user_input to ask which is the case, with the options phrased as that effect: edits to this skill stick, or the next update overwrites them. When that gate cannot reach the user, treat the skill as package-managed and say so in the report. Route the lesson to Auto memory when a package manager maintains the skill, recording the skill it applies to so the knowledge survives the next update. Whichever branch sent it there, when Step 1 resolved no Auto memory target, leave the skill unedited and report the lesson as unrouted. For those skills this rule outranks the skill-first rule, the routing table rows, and the tiebreakers below.

Destination Criteria
Project improvements Actionable improvement to existing code: refactoring, performance, reliability, readability, testing, or DX. Not for documentation fixes — factual errors in AGENTS.md belong in the Project AGENTS.md row. Route to .turbo/improvements.md via the $note-improvement skill.
Auto memory Discovered knowledge with no skill home: API quirks, debugging workarounds, compiler gotchas, tool pitfalls, user preferences. Must not overlap with any existing skill's domain — if it does, route to the skill instead (see skill-first rule above). A lesson the package-managed skills rule sends here stays here, whatever domain it overlaps.
Project AGENTS.md Intentional project decisions: conventions, architecture, stack choices, build setup, module boundaries. Also factual corrections to AGENTS.md content (wrong commands, outdated paths, incorrect conventions) — fix these directly, do not defer to Project improvements. When the lesson applies only to one subtree, route it to the nearest enclosing AGENTS.md; reserve the root file for project-wide decisions.
Existing user/project skill Lesson would improve a skill's instructions, supporting files, or reference materials, add a missing edge case, correct its workflow, or refine its trigger conditions. Route to any skill whose domain covers the lesson — not just the skill worked on in this session. Changes go to the skill file directly. No contribution flow.
New skill A cohesive body of knowledge emerged that deserves its own on-demand context. The test: would this knowledge be too large for an AGENTS.md section, and should it only be loaded when relevant? See the skill categories table below.
Existing turbo skill Same criteria as Existing user/project skill above, but for turbo skills. Before routing here, run test -d ~/.turbo/repo/codex/skills/<name>; if the directory does not exist, route to the Existing user/project skill destination instead, subject to the package-managed skills rule above. Changes go to the installed copy at ~/.agents/skills/, and are flagged for contribution (see Step 6).
No destination Does not clearly fit any destination. Drop it. Routing a weak lesson is worse than losing it.

Skill categories:

Category What it encodes Example
Domain expertise Best practices, patterns, API preferences SwiftUI expert, Core Data guide
Tool/Service integration API references, operations, ID formats Paddle, Stripe, Keycloak
Decision framework Judgment criteria, confidence levels, triage Evaluate findings, performance audit
Content template Writing conventions, tone, structure Drafting, blog post, changelog
Knowledge/Research Information discovery, schema definitions Knowledge base, research process
Orchestrated workflow Stateful multi-step procedures Process ticket, process income

Splitting heuristic: When a session creates scripts or multi-step procedures, split the lesson: a brief pointer goes to AGENTS.md (script names, purpose), and the full workflow goes to a skill. Don't collapse them into a single AGENTS.md entry.

Tiebreakers (in priority order):

  1. Skill correction → skill (hard rule). Any lesson that corrects, constrains, or refines a skill's behavior MUST route to that skill. Never to auto memory, never to AGENTS.md. This is the highest-priority routing rule, ahead of every tiebreaker below and yielding only to the package-managed skills rule above.
  2. Turbo skill vs. AGENTS.md → always the turbo skill. Broader impact (benefits all turbo users), better scoped, loaded only when relevant.
  3. Skill vs. AGENTS.md → always the skill. Skills are more discoverable, better scoped, and loaded only when relevant.
  4. Skill vs. auto memory → always the skill. If a lesson falls within the domain of an existing skill, it goes to the skill. Auto memory is for knowledge that has no skill home, and for the domain of a skill a package manager maintains.
  5. AGENTS.md vs. auto memory — intentional decisions go to AGENTS.md. Discovered knowledge (gotchas, workarounds, quirks) goes to auto memory.
  6. Lesson vs. improvement — if the item is knowledge to remember, it's a lesson. If it's work to do later, it's an improvement. They don't compete — the same session can produce both.

Step 5: Present Routing Plan

Output a table as text before making any changes:

| # | Lesson | Destination | Action |
|---|--------|-------------|--------|
| 1 | Always use X for... | Project AGENTS.md | Append to ## Conventions |
| 2 | The $create-pr skill should... | ~/.agents/skills/create-pr | Update Step 2 |
| 3 | Multi-step deploy workflow | New project skill | Create new skill |
| 4 | User prefers short commit msgs | Auto memory | Update <resolved memory target> |

For each lesson, show: concise summary, exact target file/skill, and whether it's an append, update-in-place, or new creation. For Auto memory, name the resolved source or harness update target and state when a source write still requires import and consolidation before Codex can recall it. The approval covers the durable memory write at that displayed target.

Then use request_user_input with these options: Approve or Reject.

Step 6: Execute

Apply approved changes in order:

  1. Improvements — For items routed to project improvements, run the $note-improvement skill with the summary, location, and rationale for each.

  2. Updates to auto memory — Re-read the approved target immediately before writing. For a synced Claude Code source, match the existing topic-file and index conventions, leave .consolidate-lock untouched, and report propagation as pending; the lesson becomes available to Codex only after successful import and consolidation. Claim completed recall only after verifying it. For a harness memory update target, follow the active memory instructions; when they name an additive note intake, create a new note rather than editing consolidated memory files. Keep imported extension resources read-only in every branch.

  3. Updates to AGENTS.md — Read the target file selected in Step 4 (the root file or a nested subtree file, resolved through the override rule in Step 1), find the right section, append or update in place. Match the tone and format already present.

  4. Updates to user/project skills — Run the $create-skill skill to apply changes to any file inside the skill directory (SKILL.md, references, scripts, assets).

  5. New skills — Run the $create-skill skill for each new skill. Provide the trigger conditions and relevant context from the session.

  6. Updates to turbo skills — For each lesson routed to a turbo skill:

    1. Read ~/.turbo/repo/codex/SKILL-CONVENTIONS.md so turbo-specific conventions are in context before any editing.
    2. Run the $create-skill skill to update the installed copy at ~/.agents/skills/<name>/.

    Once every turbo skill edit is in place and reviewed, use request_user_input to ask "These turbo skill improvements could benefit other users. Propose them upstream?" When the user confirms, run the $contribute-turbo skill once for all of them.

Then call update_plan to mark this step completed and continue with the next step of the active workflow.

Writing Guidelines

  • Match the tone and format of the target file
  • Use imperative mood and short declarative sentences
  • Group related insights under a descriptive heading
  • Omit rationale unless the rule would seem arbitrary without it
  • Never include temporary state, in-progress work, or task-specific details
  • Keep lessons generic—avoid overly concrete examples; state the rule, not the instance
  • For AGENTS.md: write as agent documentation — project rules any AI agent on this repo should follow
  • For auto memory: write as personal project notes — concise, operational, organized by topic
  • For skills: follow the conventions in the existing skill collection
  • For files that live in the repo (AGENTS.md and project skills): name only mechanisms that live in the repo too; describe an installed skill's behavior generically
Files (turbo)
  • references
    • transcript-miner.md 13 KB 
      # Transcript Miner Guidelines

Recover session evidence that context compaction dropped, from the rollout file Codex persists on disk. Read and report; change nothing.

## Mining Process

### 1. Locate the Rollout File

Codex writes every session to `~/.codex/sessions/<yyyy>/<mm>/<dd>/rollout-<timestamp>-<session-id>.jsonl`. The path carries no working directory, and sub-agent threads write their own rollout files under the same directories, carrying a copy of the parent's turns from the fork point. Match on the first record of each file, which is `session_meta`: keep the files whose payload `cwd` is the project root or a directory beneath it, and which carry no parent link. The parent link takes three shapes, and a file carrying any of them is a copy: `parent_thread_id`, `forked_from_id`, and a `source` object holding a `subagent` key.

```bash
python3 - "<project root>" <<'PY'
import glob, json, os, sys

hits = []
for path in glob.glob(os.path.expanduser("~/.codex/sessions/*/*/*/rollout-*.jsonl")):
    try:
        with open(path, encoding="utf-8") as fh:
            meta = json.loads(fh.readline()).get("payload", {})
    except Exception:
        continue
    cwd = meta.get("cwd") or ""
    root = sys.argv[1].rstrip("/")
    if not (cwd == root or cwd.startswith(root + "/")):
        continue
    if meta.get("parent_thread_id") or meta.get("forked_from_id"):
        continue
    src = meta.get("source")
    if isinstance(src, dict) and "subagent" in src:
        continue
    hits.append(path)
for path in sorted(hits, key=os.path.getmtime, reverse=True)[:5]:
    print(path)
PY
```

The current session is the most recently modified match. Confirm it by grepping for the distinctive phrase you were given, using a short fragment that sits on one line and contains no quote characters: the stored text is JSON-escaped. When the phrase is absent, try the next file down.

If no matching rollout file can be found, report that in one line and stop.

### 2. Extract the User Side

Each line is a record with `type` and `payload`. Submitted user turns are `event_msg` records taking one of two shapes, and a file carries only one of them: a payload with `type: "user_message"` holding the text in `payload.message`, or a payload with `type: "item_completed"` whose `payload.item.type` is `UserMessage`, holding the text in the `text` content parts of `payload.item.content` (other part types are attachments carrying none). Match both shapes, since an extraction matching one returns nothing at all on a file written in the other. The model-facing copies live in `response_item` records, wrapped in harness blocks that read like user speech, so extract from `event_msg` instead. The `compacted` record type marks where a compaction cut the thread; anything before it is what this run needs to recover.

A submitted turn can open with a harness-injected block, such as an attachment manifest or captured browser context, followed by the user's own text in the same record. Where the turn opens with one of those blocks, the user's text follows a `## My request:` marker, so keep what follows the last one, and keep the record whole when nothing follows it. Strip a leading block only when it is a known harness heading; a block that can itself carry the user's words, such as a response-annotation block quoting their comment on an earlier turn, is kept whole rather than dropped.

An answer to a `request_user_input` gate produces no ordinary user turn and arrives in one of two shapes. A synchronous answer is a `response_item` whose payload is a `function_call_output`, holding a JSON string in `payload.output` shaped `{"answers": {"<question id>": {"answers": ["<answer>"]}}}`. Its question text sits in the preceding `function_call` record carrying the same `call_id`, whose `arguments` is a JSON string listing `questions` by `id`, so pair the two. An asynchronous answer leaves that output empty (`{"answers": {}}`) and arrives in a later user turn as a `<send_user_message_question_reply>` block. That block holds a JSON array of objects carrying their own `question` and `answer`, so it needs no pairing. Each entry holds the user's own words: the option label they selected, or the text they supplied.

Print the real user turns:

```bash
python3 - "<rollout path>" <<'PY'
import json, sys

WRAPPERS = ("# Files mentioned by the user:", "# In app browser:",
            "# Context from my IDE setup:", "<in-app-browser-context")
REQUEST = "## My request:"
GATE = "request_user_input"
REPLY = "<send_user_message_question_reply>"
REPLY_END = "</send_user_message_question_reply>"


def strip_wrapper(text):
    """Drop leading harness blocks, keeping whatever the user typed after them."""
    if not text.startswith(WRAPPERS):
        return text
    if REQUEST in text:
        return text.rsplit(REQUEST, 1)[1].strip() or text
    lines = text.splitlines()
    while lines and (not lines[0].strip() or lines[0].startswith("#")):
        lines.pop(0)
    return "\n".join(lines).strip()


def elide(text):
    """Keep the opening and the operative ending of an unusually long turn."""
    if len(text) <= 6000:
        return text
    return text[:4000] + "\n[...]\n" + text[-2000:]


def questions(arguments):
    """Map question id to question text for one request_user_input call."""
    try:
        asked = json.loads(arguments or "{}")
    except ValueError:
        return {}
    return {q.get("id"): q.get("question", "") for q in asked.get("questions") or []
            if isinstance(q, dict)}


def rendered(pairs):
    """Render question/answer pairs as one user turn."""
    return "The user answered: " + ", ".join(pairs) if pairs else ""


def answers(output, asked):
    """Render the answers of a synchronous request_user_input call."""
    try:
        given = json.loads(output or "{}")
    except ValueError:
        return ""
    pairs = []
    for qid, val in (given.get("answers") or {}).items():
        for answer in (val or {}).get("answers") or []:
            pairs.append(f'"{asked.get(qid) or qid}"="{answer}"')
    return rendered(pairs)


def replies(text):
    """Render asynchronous reply blocks, keeping whatever the user typed around them."""
    pairs, rest, remainder = [], [], text
    while REPLY in remainder:
        before, _, tail = remainder.partition(REPLY)
        chunk, _, remainder = tail.partition(REPLY_END)
        rest.append(before)
        try:
            given = json.loads(chunk.strip())
        except ValueError:
            continue
        for entry in given if isinstance(given, list) else []:
            if isinstance(entry, dict):
                pairs.append(f'"{entry.get("question", "")}"="{entry.get("answer", "")}"')
    rest.append(remainder)
    kept = "\n".join(part.strip() for part in rest if part.strip())
    return "\n".join(part for part in (rendered(pairs), kept) if part)


def typed(item):
    """Join the text parts of a completed user message, skipping attachments."""
    return "\n".join(part.get("text") or "" for part in item.get("content") or []
                     if isinstance(part, dict) and part.get("type") == "text").strip()


asked = {}
seen = ()
for line in open(sys.argv[1], encoding="utf-8"):
    try:
        rec = json.loads(line)
    except ValueError:
        continue
    payload = rec.get("payload") or {}
    kind = payload.get("type")
    item = payload.get("item") if isinstance(payload.get("item"), dict) else {}
    if rec.get("type") == "event_msg" and kind == "user_message":
        text = strip_wrapper((payload.get("message") or "").strip())
    elif (rec.get("type") == "event_msg" and kind == "item_completed"
            and item.get("type") == "UserMessage"):
        text = typed(item)
        text = replies(text) if REPLY in text else strip_wrapper(text)
    elif kind == "function_call" and payload.get("name") == GATE:
        asked[payload.get("call_id")] = questions(payload.get("arguments"))
        continue
    elif kind == "function_call_output" and payload.get("call_id") in asked:
        text = answers(payload.get("output"), asked.pop(payload.get("call_id")))
    else:
        continue
    stamp = rec.get("timestamp", "")
    if not text or (stamp, text) == seen:
        continue
    seen = (stamp, text)
    print(f"--- {stamp}\n{elide(text)}\n")
PY
```

A session driven entirely by skill pipelines can yield no typed user turns at all. Then the evidence lives on the assistant side: extract `response_item` records whose payload has `type: "message"` and `role: "assistant"`, joining the `text` of their content parts. Those payloads carry a `phase` of `commentary` or `final_answer`; label each extracted turn with its phase, so a preamble is not read as a conclusion.

### 3. Identify Evidence

Read the extracted turns in order and collect:

- **Corrections** — the user interrupted, said no, redirected, or fixed something by hand
- **Repeated guidance** — the same instruction given more than once
- **Preferences** — formatting, naming, style, or tool choices the user expressed
- **Failure modes** — an approach that failed, with what replaced it
- **Other** — anything else that stays true beyond this session

A correction is ambiguous without the thing it corrected. For each one, locate its record with `grep -n "<timestamp>" "<rollout path>"` and read the preceding records with a line-range slice such as `sed -n '<start>,<end>p'` to see what prompted it, then state that in one line.

Keep items that would still hold in a future session. Drop one-off instructions that only steer the task at hand.

## Sweep Process

Follow this instead of the Mining Process when mining the project's whole history rather than one session.

### 1. List Every Rollout File

Keep every Mining Process filter except the distinctive-phrase match. Take every match for the project, oldest first, applying the cutoff you were given:

```bash
python3 - "<project root>" "<cutoff ISO-8601, or empty>" <<'PY'
import datetime, glob, json, os, sys

root = sys.argv[1].rstrip("/")
cutoff = 0.0
if len(sys.argv) > 2 and sys.argv[2]:
    cutoff = datetime.datetime.fromisoformat(sys.argv[2]).timestamp()

hits = []
for path in glob.glob(os.path.expanduser("~/.codex/sessions/*/*/*/rollout-*.jsonl")):
    try:
        with open(path, encoding="utf-8") as fh:
            meta = json.loads(fh.readline()).get("payload", {})
    except Exception:
        continue
    cwd = meta.get("cwd") or ""
    if not (cwd == root or cwd.startswith(root + "/")):
        continue
    if meta.get("parent_thread_id") or meta.get("forked_from_id"):
        continue
    src = meta.get("source")
    if isinstance(src, dict) and "subagent" in src:
        continue
    if os.path.getmtime(path) < cutoff:
        continue
    hits.append(path)
for path in sorted(hits, key=os.path.getmtime):
    print(path)
PY
```

Drop the most recently modified match: that is the live session, already covered by the scan that dispatched you. When no file matches, or none remains after that drop, report it in one line and stop.

Drop what a previous run already distilled: the records preceding the last `$self-improve` invocation in any file that holds one. Codex injects a loaded skill as a `<skill><name>…</name>` block inside a `response_item`, which the extraction below skips in favor of `event_msg`, so find it in the raw file and slice from there, resolving each rollout file to the path the next step reads:

```bash
LINE=$(grep -n '<name>self-improve</name>' "<rollout path>" | tail -1 | cut -d: -f1)
INPUT="<rollout path>"
if [ -n "$LINE" ]; then
  INPUT="<scratch>/<session-id>.jsonl"
  sed -n "$((LINE + 1)),\$p" "<rollout path>" > "$INPUT"
fi
```

Match that block rather than the bare string `$self-improve`, which also matches file reads, diffs, and prose that merely names the skill.

### 2. Extract Each One

Run the Mining Process extraction script unchanged over each resolved input path rather than writing a fresh one. Its `event_msg` targeting is what keeps the output readable, since the `response_item` copies are wrapped in plugin and skill-injection blocks that bury the typed turns, and its gate handling recovers answers that produce no ordinary user turn at all.

Loop over those paths in a single command, printing each rollout file's own path as a header before its extraction and appending both to one scratch file outside the repo. Read that file rather than the extraction output, in oldest-first slices when it is large, carrying the candidate items from each slice forward into the next.

### 3. Identify Evidence Across Sessions

Collect per the Mining Process categories, tracing the context of a correction only for items that survive as candidates. Note where the same guidance appears in more than one session: repetition across sessions is the signal that separates a documentation gap from one-off steering.

Report per the Output Format below, adding a `**Sessions**` line to each entry naming the rollout files it came from.

## Output Format

```
## Recovered Evidence

### <category>: <one-line claim>
- **Quote**: "<verbatim user words>"
- **Context**: <what prompted it>
- **When**: <timestamp>
- **Sessions**: <rollout paths — sweep only>
```

Order the entries by the category order above. When nothing durable survives, say so in one line.
  • SKILL.md 21.9 KB 
    ---
name: self-improve
description: "Extract lessons from the current session, or sweep the project's past sessions when asked, and route them to the appropriate knowledge layer (project AGENTS.md, auto memory, existing skills, or new skills). Use when the user asks to \"self-improve\", \"distill this session\", \"distill past sessions\", \"sweep past sessions\", \"extract lessons from all sessions\", \"save learnings\", \"update AGENTS.md with what we learned\", \"capture session insights\", \"remember this for next time\", \"extract lessons\", \"update skills from session\", or \"what did we learn\"."
---

# Self-Improve

Review the current conversation, or the project's past sessions when asked, to extract durable lessons and route each one to the right knowledge layer.

## Step 1: Detect Context

Available destinations:

- **Project AGENTS.md** — the instruction file in the project root, plus any nested ones in subdirectories. At each level use `AGENTS.override.md` when one is present, otherwise `AGENTS.md`: an override replaces that directory's `AGENTS.md` rather than adding to it, so a lesson written to a shadowed file never loads. A nested file scopes its guidance to that subtree, so a lesson scoped to one subtree belongs in the nearest enclosing file, with the root reserved for project-wide rules.
- **Auto memory** — one writable memory target resolved for the canonical project root. Prefer a synced Claude Code project-memory source, then a memory-update target explicitly named by the active harness. If neither resolves, skip this destination.
- **Skills** — Project skills at `.agents/skills/` (walked from project root down to cwd) and user-installed skills at `~/.agents/skills/` (resolve symlinks)

Resolve the canonical project root as the Git top-level containing the working directory, or the physical working directory outside a repository. Resolve Auto memory in this order:

1. **Synced Claude Code source memory** — Read `~/.turbo/config.json`. When `codex.sharedClaudeAutoMemory` is exactly `true`, treat shared-memory setup as complete for every project and use this branch. Resolve `autoMemoryDirectory` from Claude Code's managed settings, then project-local `.claude/settings.local.json`, then the user `<Claude config home>/settings.json`; exclude checked-in `.claude/settings.json`, where Claude Code ignores this security-sensitive field. Use the expanded override when set. Otherwise locate the resource directory under `~/.codex/memories/extensions/external_agent_import/resources/` whose `scope.json` matches the canonical project root, take its directory name as the exact Claude project key, and use `projects/<project key>/memory/` under `CLAUDE_CONFIG_DIR` when set or `~/.claude/`. Treat `false`, a missing or invalid key, and an unreadable `~/.turbo/config.json` as not opted in and continue to the harness target.
2. **Harness memory update target** — Otherwise use a project-specific memory file or additive note intake that the active harness instructions explicitly name for memory updates.
3. **No target** — Otherwise skip Auto memory and route lessons through the remaining destinations.

Treat imported extension resources and the harness's consolidated memory files as read-only. For the resolved target, list the directory when applicable and read its index before routing lessons.

Discover the project instruction files (the root file and any nested ones in subdirectories, resolved through the override rule above) and read them. When those files or the resolved Auto memory target point at a knowledge base the repo maintains, read its index too; it is a documentation source for Step 3 rather than a routing destination. List all skill directories but do not read them yet — Step 2 needs to run first so you know what to look for.

### Skill Ownership Detection

Classify every skill this session touched:

- Skills that live in the project are user/project skills
- If `~/.turbo/repo/` exists, list directories in `~/.turbo/repo/codex/skills/`; any skill in `~/.agents/skills/` with a matching directory there is a turbo skill
- For every other skill in `~/.agents/skills/`, read `~/.agents/.skill-lock.json`. Its top-level `skills` object is keyed by skill name, and a skill listed there was installed from a source that replaces it wholesale on its next update. Match on the resolved path rather than the name alone, against the entry's `skillPath`, so a local fork that replaced the installed copy is not mistaken for it. A skill that matches is package-managed. The signal runs one way: absence from the file leaves ownership genuinely open, so carry an unlisted skill into Step 4 as ownership-unresolved

**Verification rule (mandatory before routing in Step 4):** For every candidate skill that is about to be routed as turbo, confirm with a fresh `test -d ~/.turbo/repo/codex/skills/<name>` check that the skill actually lives in the turbo repo. Do not rely on remembered listings from earlier in the session, filename hits in grep output, or assumptions based on where a SKILL.md was read from. A miss here mislabels a user/project skill as turbo, triggers the contribution flow unnecessarily, and can introduce session-specific content into a shared skill — so the check is not optional.

**Exception:** If the current project IS the turbo repo (i.e., the working directory contains this skill collection), route turbo skill lessons through the **Existing user/project skill** destination in Step 4 — edits go directly to `codex/skills/<name>/` in the project, with no installed-copy indirection and no contribution flow.

## Step 2: Gather Session Evidence and Scan for Lessons

### Recover Pre-Compaction Evidence

**Skip** when the conversation is visible in full from the user's own first message.

When it starts from a summary of earlier work instead, recover the compacted turns from the on-disk rollout file. Spawn a single sub-agent (inherited model defaults). The sub-agent's prompt must include:

1. The absolute path of the project root
2. A distinctive phrase from the visible conversation, for confirming which rollout file belongs to this session
3. An instruction to read [references/transcript-miner.md](references/transcript-miner.md) for rollout file location, extraction, and output format

Treat the returned items as raw evidence for the scan below.

### Sweep Past Sessions

**Run** when asked to distill sessions beyond the current one. **Skip** otherwise.

Propose a cutoff first. When Step 1 resolved a harness memory update target, take its newest modification time, state it, then use `request_user_input` to confirm sweeping from it or sweeping the whole history; when that gate cannot reach the user, sweep from the proposed cutoff and say so in the report. A memory file's timestamp records a write rather than a completed sweep, so it bounds the work without settling what a previous run covered. When Step 1 resolved the synced Claude Code source, take its newest modification time and state that Claude Code writes also advance it, so it may omit unswept Codex sessions; use `request_user_input` with whole Codex history as the recommended option and that timestamp as the bounded option. When that gate cannot reach the user, sweep the whole Codex history and say so in the report. When no memory location is available, sweep the whole history without asking.

Spawn a single sub-agent (inherited model defaults). The sub-agent's prompt must include:

1. The absolute path of the project root
2. The confirmed cutoff as an ISO-8601 timestamp, or that there is none
3. An instruction to read [references/transcript-miner.md](references/transcript-miner.md) and follow its sweep process

Treat the returned items as raw evidence for the scan below.

### Identify Session Skills

Before scanning for lessons, identify which skills were loaded during this session:

- Scan the conversation for Turbo skill invocations and `SKILL.md` reads from `~/.agents/skills/`
- Build a list of session skills, marking each as turbo, user/project, package-managed, or ownership-unresolved (using the detection from Step 1)
- This list informs routing in Step 4: when a lesson clearly arose from a specific skill's workflow, that skill is the natural routing target

### Scan for Lessons

Scan the full conversation with this priority:

1. **Corrections** — Where the user interrupted, said "no", "actually", "stop", "not like that", redirected, or manually fixed something Codex did wrong. Highest-value lessons.
2. **Repeated guidance** — Instructions the user gave more than once. Across separate sessions this counts even where each instance reads as ordinary steering on its own.
3. **Skill-shaped knowledge** — Domain expertise that was needed repeatedly, tool/API integration details that had to be looked up, decision frameworks that emerged for evaluating options, content templates or writing conventions that were refined, and multi-step workflows where ordering mattered (as reusable domain knowledge, not the workflow itself — see #4).
4. **New workflows** — Did this session establish a novel multi-step procedure, coordination pattern, or automation that worked? A successful workflow that would need to be repeated is a prime skill candidate — even if it ran fine this time. Distinct from #3: this captures the procedure itself as a repeatable artifact, not knowledge about how to do it. Flag it.
5. **Preferences** — Formatting, naming, style, or tool choices the user expressed.
6. **Failure modes** — Approaches that failed, with what worked instead. For tool or script call failures, trace back to the information source that led to the error and route the fix there (e.g., clarify a reference file, update skill instructions, add missing documentation).
7. **Domain knowledge** — Facts or conventions Codex needed but did not have.
8. **Improvement opportunities** — Out-of-scope improvements noticed during work: code that could be refactored, missing tests, performance issues, readability concerns, or feature ideas that were intentionally skipped to stay focused. **Skipped findings count here**: when code simplification or code review identified a genuine improvement or issue but it was skipped for this session, route it as a project improvement so it isn't lost.
9. **Trusted reviewer feedback** — Human PR review comments that reveal project conventions, patterns, or corrections. Trusted reviewers are repo collaborators with `admin` or `maintain` roles (determine via `gh api repos/{owner}/{repo}/collaborators --jq '.[] | select(.role_name == "admin" or .role_name == "maintain") | .login'`). Their feedback takes precedence over other reviewers and AI bots when there are contradictions.

After scanning, read all skill SKILL.md files (they are small). This gives Step 4 full context for routing.

## Step 3: Filter

Keep only lessons that are:
- **Stable** — likely to remain true across future sessions
- **Non-obvious** — Codex would not already know this
- **Actionable** — can be expressed as a rule or instruction
- **Not already documented** — absent from the files read in Steps 1 and 2, from each skill's `references/` and other supporting files, and from the project's own knowledge stores. Search all of those for each candidate's keywords rather than assuming. A lesson documented only in an unrelated subtree's AGENTS.md still counts as undocumented for the subtree it actually applies to.
- **Still a concern** — the issue is not already fixed by changes made in this session. If a bug was found and fixed, or a missing feature was added, future sessions will see the corrected code — they don't need a reminder about the old problem. **Exception: successful workflows and procedures are not "resolved" — they're skill candidates precisely because they worked and will need to be repeated.** When sweeping past sessions, judge this against the current state of the code and docs rather than against this session's changes.

Discard anything session-specific, speculative, one-off, or already resolved by code changes in this session (but not successful workflows — see exception above). If no lessons survive filtering, tell the user and stop.

## Step 4: Route Each Lesson

Assign each surviving lesson to exactly one destination.

**Skill-first rule (mandatory):** Before consulting the table below, check whether the lesson corrects, refines, or adds a guardrail to any existing skill's behavior — turbo or user/project. This includes lessons about skipping steps, wrong defaults, missing edge cases, or any "don't do X when running $skill-name" correction. If yes, route to that skill. Do not route skill corrections to auto memory or AGENTS.md — they belong in the skill they correct. This rule is not a preference; it is a hard constraint that takes precedence over the table rows below.

**Package-managed skills (mandatory):** A skill Step 1 classified as package-managed from the lock file routes its lesson straight to **Auto memory**, recording the skill it applies to, with no question put to the user: the lock file has already settled ownership. A skill under `~/.agents/skills/` that Step 1 left ownership-unresolved may still be one a package manager replaces wholesale on its next update, discarding any edit made here. Before routing a lesson to such a skill, say plainly that an edit to it survives only while the user maintains it themselves. Then use `request_user_input` to ask which is the case, with the options phrased as that effect: edits to this skill stick, or the next update overwrites them. When that gate cannot reach the user, treat the skill as package-managed and say so in the report. Route the lesson to **Auto memory** when a package manager maintains the skill, recording the skill it applies to so the knowledge survives the next update. Whichever branch sent it there, when Step 1 resolved no Auto memory target, leave the skill unedited and report the lesson as unrouted. For those skills this rule outranks the skill-first rule, the routing table rows, and the tiebreakers below.

| Destination | Criteria |
|---|---|
| **Project improvements** | Actionable improvement to existing **code**: refactoring, performance, reliability, readability, testing, or DX. Not for documentation fixes — factual errors in AGENTS.md belong in the **Project AGENTS.md** row. Route to `.turbo/improvements.md` via the `$note-improvement` skill. |
| **Auto memory** | Discovered knowledge with no skill home: API quirks, debugging workarounds, compiler gotchas, tool pitfalls, user preferences. Must not overlap with any existing skill's domain — if it does, route to the skill instead (see skill-first rule above). A lesson the package-managed skills rule sends here stays here, whatever domain it overlaps. |
| **Project AGENTS.md** | Intentional project decisions: conventions, architecture, stack choices, build setup, module boundaries. Also factual corrections to AGENTS.md content (wrong commands, outdated paths, incorrect conventions) — fix these directly, do not defer to Project improvements. When the lesson applies only to one subtree, route it to the nearest enclosing AGENTS.md; reserve the root file for project-wide decisions. |
| **Existing user/project skill** | Lesson would improve a skill's instructions, supporting files, or reference materials, add a missing edge case, correct its workflow, or refine its trigger conditions. Route to any skill whose *domain* covers the lesson — not just the skill worked on in this session. Changes go to the skill file directly. No contribution flow. |
| **New skill** | A cohesive body of knowledge emerged that deserves its own on-demand context. The test: would this knowledge be too large for an AGENTS.md section, and should it only be loaded when relevant? See the skill categories table below. |
| **Existing turbo skill** | Same criteria as **Existing user/project skill** above, but for turbo skills. **Before routing here, run `test -d ~/.turbo/repo/codex/skills/<name>`; if the directory does not exist, route to the Existing user/project skill destination instead, subject to the package-managed skills rule above.** Changes go to the installed copy at `~/.agents/skills/`, and are flagged for contribution (see Step 6). |
| **No destination** | Does not clearly fit any destination. Drop it. Routing a weak lesson is worse than losing it. |

**Skill categories:**

| Category | What it encodes | Example |
|---|---|---|
| Domain expertise | Best practices, patterns, API preferences | SwiftUI expert, Core Data guide |
| Tool/Service integration | API references, operations, ID formats | Paddle, Stripe, Keycloak |
| Decision framework | Judgment criteria, confidence levels, triage | Evaluate findings, performance audit |
| Content template | Writing conventions, tone, structure | Drafting, blog post, changelog |
| Knowledge/Research | Information discovery, schema definitions | Knowledge base, research process |
| Orchestrated workflow | Stateful multi-step procedures | Process ticket, process income |

**Splitting heuristic:** When a session creates scripts or multi-step procedures, split the lesson: a brief pointer goes to AGENTS.md (script names, purpose), and the full workflow goes to a skill. Don't collapse them into a single AGENTS.md entry.

**Tiebreakers (in priority order):**
1. **Skill correction → skill (hard rule).** Any lesson that corrects, constrains, or refines a skill's behavior MUST route to that skill. Never to auto memory, never to AGENTS.md. This is the highest-priority routing rule, ahead of every tiebreaker below and yielding only to the package-managed skills rule above.
2. **Turbo skill vs. AGENTS.md → always the turbo skill.** Broader impact (benefits all turbo users), better scoped, loaded only when relevant.
3. **Skill vs. AGENTS.md → always the skill.** Skills are more discoverable, better scoped, and loaded only when relevant.
4. **Skill vs. auto memory → always the skill.** If a lesson falls within the domain of an existing skill, it goes to the skill. Auto memory is for knowledge that has no skill home, and for the domain of a skill a package manager maintains.
5. **AGENTS.md vs. auto memory** — intentional decisions go to AGENTS.md. Discovered knowledge (gotchas, workarounds, quirks) goes to auto memory.
6. **Lesson vs. improvement** — if the item is *knowledge to remember*, it's a lesson. If it's *work to do later*, it's an improvement. They don't compete — the same session can produce both.

## Step 5: Present Routing Plan

Output a table as text before making any changes:

```
| # | Lesson | Destination | Action |
|---|--------|-------------|--------|
| 1 | Always use X for... | Project AGENTS.md | Append to ## Conventions |
| 2 | The $create-pr skill should... | ~/.agents/skills/create-pr | Update Step 2 |
| 3 | Multi-step deploy workflow | New project skill | Create new skill |
| 4 | User prefers short commit msgs | Auto memory | Update <resolved memory target> |
```

For each lesson, show: concise summary, exact target file/skill, and whether it's an append, update-in-place, or new creation. For Auto memory, name the resolved source or harness update target and state when a source write still requires import and consolidation before Codex can recall it. The approval covers the durable memory write at that displayed target.

Then use `request_user_input` with these options: **Approve** or **Reject**.

## Step 6: Execute

Apply approved changes in order:

1. **Improvements** — For items routed to project improvements, run the `$note-improvement` skill with the summary, location, and rationale for each.
2. **Updates to auto memory** — Re-read the approved target immediately before writing. For a synced Claude Code source, match the existing topic-file and index conventions, leave `.consolidate-lock` untouched, and report propagation as pending; the lesson becomes available to Codex only after successful import and consolidation. Claim completed recall only after verifying it. For a harness memory update target, follow the active memory instructions; when they name an additive note intake, create a new note rather than editing consolidated memory files. Keep imported extension resources read-only in every branch.
3. **Updates to AGENTS.md** — Read the target file selected in Step 4 (the root file or a nested subtree file, resolved through the override rule in Step 1), find the right section, append or update in place. Match the tone and format already present.
4. **Updates to user/project skills** — Run the `$create-skill` skill to apply changes to any file inside the skill directory (SKILL.md, references, scripts, assets).
5. **New skills** — Run the `$create-skill` skill for each new skill. Provide the trigger conditions and relevant context from the session.
6. **Updates to turbo skills** — For each lesson routed to a turbo skill:
   1. Read `~/.turbo/repo/codex/SKILL-CONVENTIONS.md` so turbo-specific conventions are in context before any editing.
   2. Run the `$create-skill` skill to update the installed copy at `~/.agents/skills/<name>/`.

   Once every turbo skill edit is in place and reviewed, use `request_user_input` to ask "These turbo skill improvements could benefit other users. Propose them upstream?" When the user confirms, run the `$contribute-turbo` skill once for all of them.

Then call `update_plan` to mark this step completed and continue with the next step of the active workflow.

## Writing Guidelines

- Match the tone and format of the target file
- Use imperative mood and short declarative sentences
- Group related insights under a descriptive heading
- Omit rationale unless the rule would seem arbitrary without it
- Never include temporary state, in-progress work, or task-specific details
- Keep lessons generic—avoid overly concrete examples; state the rule, not the instance
- For AGENTS.md: write as agent documentation — project rules any AI agent on this repo should follow
- For auto memory: write as personal project notes — concise, operational, organized by topic
- For skills: follow the conventions in the existing skill collection
- For files that live in the repo (AGENTS.md and project skills): name only mechanisms that live in the repo too; describe an installed skill's behavior generically

Comments (0)

Sign in to join the conversation.

No comments yet.

Reviews (0)

No reviews yet.

Related

Claude Prompt
Tighten-my-prose editing prompt
Claude Skill
Adversarial code-review skill for Claude
Claude Workflow
Deep-research workflow (fan-out + verify)
Claude Skill
Conventional-commit writer for Claude