devlog
Use when finishing a feature, fixing a notable bug, wrapping up a work session, or discovering something non-obvious (a gotcha, undocumented behavior, a workaround that took several tries, a constraint you only noticed after hitting it). Write the entry immediately, in the same t
Install
npx skills add https://github.com/technicalpickles/pickled-claude-plugins/tree/main/plugins/second-brain/skills/devlog
claude plugin marketplace add https://llmmart.ai/marketplace.json && claude plugin install technicalpickles-pickled-claude-plugins@llmmart
git clone https://github.com/technicalpickles/pickled-claude-plugins.git
The skills CLI installs just this skill, for any of its supported agents. Claude Code installs the whole technicalpickles/pickled-claude-plugins collection as a plugin from our marketplace. Git is the plain clone.
Skill manifest
Devlog
Overview
Devlog is continuous, point-in-time work commentary: what happened, why, and anything surprising, captured as it happens instead of reconstructed from memory at the end of a session. One narrative session note accumulates every entry from the current session — no separate "learnings" system, no til/cursed split, everything worth remembering goes in the same place.
Processing (routing, connecting, daily-linking) happens later via
/second-brain:process-inbox. This skill only ever appends to the inbox.
Write the entry now, in the same turn as whatever triggered it. Don't
ask "should I log this?" and don't offer to — just write it. Asking is
exactly the failure mode that made the old distill-conversation flow get
skipped: by the time a session ends, you're mentally done, and a skip-able
step gets skipped.
When to write
- Finished a feature or fixed a notable bug
- Wrapping up a work session
- Discovered something non-obvious: a gotcha, undocumented behavior, a workaround that took several tries, a constraint you only noticed after hitting it
When not to write
- Routing, connecting, or moving existing notes — that's
/second-brain:process-inbox - Reading an external source and writing it up — that's
second-brain:capture - Trivially obvious or one-off details you'd never reach for again
- A bare project-specific fact (a config value, a repo convention) with no story behind it — that belongs in Claude Code's own memory system alone, not the vault
Step 1: Load Configuration
Don't preflight-check sb availability — just run it:
npx @techpickles/sb config default && npx @techpickles/sb config vaults
If this fails, run ${CLAUDE_PLUGIN_ROOT}/scripts/diagnose-sb.sh and follow
its guidance rather than re-deriving the checks by hand.
Then pick the vault for this session's note:
${CLAUDE_PLUGIN_ROOT}/scripts/resolve-vault.sh
It prints a vault name, or nothing. Use that name as --vault <name> on every
sb call below that takes it (note create, vault obsidian, inbox list).
When it prints nothing, omit --vault and let sb use its default. The name comes
from SB_NOTE_VAULT if set (a per-project override), else from whichever
registered vault contains the working directory. Never write a session's note
to a different vault than the one the session is working in or for: a vault can
have rules about what may be written to it.
If no default vault is configured:
Second brain not configured. Run /second-brain:setup first.
Step 2: Find or create this session's devlog note
Every entry from the current session lands in ONE note. Never create a second devlog note for a session that already has one.
- Collect session context:
- Current repo and branch, if in a worktree
- Current work item, if any (see Work items below)
- Short session id (first 8 characters of the session UUID)
- Source string format:
claude-code session {short-id} ({repo}, branch {branch}, {tracker} {id})(omit fields that don't apply)
Work items
This plugin doesn't know or assume an issue tracker. If your context names a tracked work item for this session (a ticket, issue, or task ID), record it; otherwise leave it out. It can arrive two ways:
- Your instructions (CLAUDE.md, the user's own words) say what tracker they use and which item is current.
- A hook injects candidates when this skill loads, in this shape:
AWork items this session ({tracker}): - {id} ({status}): {title}PreToolUsehook on theSkilltool can emit this asadditionalContextforsecond-brain:devlog.
If there are several candidates, pick the one(s) this entry is actually about. A follow-up item filed during the session is usually not what the session was about. If none fit, or nothing names a work item, omit it. Don't guess one from the branch name.
Check your own memory first. If you already created or appended to this session's devlog note earlier in this same conversation, you already have its path in context — go straight to Step 2a and append. No CLI search needed: this is the same mechanism
second-brain:insightuses for its own accumulating session note ("if the session file already exists, was created in an earlier call in this same conversation, append").If this is the first devlog entry this session, go to Step 2b to create the note, and keep its path in mind for the rest of the session.
Fallback, only if you've genuinely lost track of the note's path (e.g. a context compaction wiped your memory of this session) — resolve a real path instead of guessing:
npx @techpickles/sb vault obsidian(add--vault <name>when step 1 picked one) — parse the JSON for the top-levelinboxfield (the inbox subfolder name, e.g.📫 Inbox). This is the precomputed fieldsb note createitself resolves the inbox folder from — don't parseapp.newFileFolderPathdirectly, since that skips the Zettelkasten Prefixer plugin's folder override (inboxiszkPrefixer?.folder ?? app?.newFileFolderPath) and can point at the wrong folder.npx @techpickles/sb inbox list --detail— parse the JSON, look for the entry withtype: session-noteswhosesourcefield contains the current short session id, and take itsfilename- Combine:
~/.claude/vaults/{vault-name}/{inbox}/{filename}is the real, readable/editable path.inbox list --detailonly ever returns a bare filename, never a resolvable path on its own. - With the resolved path in hand, go to Step 2a and append. If the
inbox list --detailsearch finds no entry matching the current session id (genuinely nothing to recover), fall through to Step 2b and create the note instead.
Step 2a: Append to the existing note
Read the note, append one new bullet to the end of its body (format below) using the Edit tool. Don't touch its frontmatter.
Step 2b: Create the note
Staged and written in one Bash call:
VAULT={vault name from resolve-vault.sh, or empty}
STAGE=$(mktemp)
cat > "$STAGE" <<'EOF'
- {HH:MM} {entry}
EOF
npx @techpickles/sb note create \
${VAULT:+--vault "$VAULT"} \
--title "Session: {topic}" \
--source "claude-code session {short-id} ({repo}, branch {branch}, {tracker} {id})" \
--content "$(cat "$STAGE")"
rm -f "$STAGE"
Run it sandboxed first. It works sandboxed when the vault's real path is in
the sandbox's allowWrite (for example ~/Vaults/<vault>/**/*.md). Name
the real directory, not a symlink to it: the sandbox matches the resolved
path. If the call fails with Operation not permitted on the vault path,
the vault isn't allowlisted. Re-run the whole call with the Bash tool's
dangerouslyDisableSandbox: true, not just the sb line: $TMPDIR
resolves to a different path unsandboxed, so staging and reading have to
happen in the same call under the same mode or the read can miss the file.
Either way, the staged file must be a fresh mktemp path,
not a fixed filename: $TMPDIR is shared across every Claude Code
session on the machine, and a fixed-name file can be overwritten or read
by a stale write from a different session.
sb note create prints JSON with a path field. If stdout has anything
before the { (e.g. a stray npm warn ... line from the caller's own
.npmrc), extract the JSON substring starting at the first { rather
than parsing the whole stdout blob.
sb note create only ever writes captured, source, repo, branch,
commit to frontmatter — it has no --type/--status flag, and repo/
branch come out as the literal none/none (devlog's --source format
doesn't match sb's own conversation:repo=...,branch=... parsing
convention). Read the file back at the path just parsed to preserve the
exact frontmatter sb wrote, then Edit it:
- add
status: raw - add
type: session-notes - replace the
repo: noneline sb wrote withrepo: {repo-name}(from git context if available) — don't add a secondrepo:line - replace the
branch: noneline sb wrote withbranch: {branch-name}(from git context if available) — don't add a secondbranch:line - add
issue: {id}if there's a current work item (see Work items); a YAML list if the entry covers several. Omit the line entirely otherwise - keep
sourceas sb wrote it — theclaude-code session ...string already covers provenance for devlog's format - keep
capturedas sb wrote it
Without this fixup, the note is invisible both to this skill's own
"check memory / find existing note" logic (Step 2, above) and to
process-inbox's stage 3a filter, which requires exactly status: raw +
type: session-notes to ever pick the note up.
Entry format
One bullet per entry, each with a wall-clock timestamp:
- {HH:MM} {what happened, why, anything surprising — one to three
sentences, standalone enough to make sense without the surrounding
conversation}
Prose, not a title. Write it like a short journal line, not a commit message.
Also saved somewhere else?
A record somewhere else doesn't replace the devlog entry. Claude Code's memory is what future sessions act on, an issue or ticket is the tracked work, and the devlog is the vault's record of what happened. When the discovery was also saved to one of those, write the entry anyway and say where it went:
- 14:05 The VM's guest DNS drifts off its DHCP lease after the host
sleeps, so image pulls time out until it's reset. Saved as a memory
(`~/.claude/projects/<project>/memory/reference_vm_dns_drift.md`).
Point at it, don't paste it: one or two sentences of what happened, then the pointer. The other record holds the details. Which trackers count beyond memory (and how to cite them) is up to your own instructions or the hook that nudged you here; this skill doesn't assume any.
Constraints
- One note per session — append, never create a second one for the same session.
- Never ask before writing, never present a summary and wait for a selection.
- Don't route, connect, move the note, or touch the daily note. That's
entirely
process-inbox's job, later. - Frontmatter stays
status: raw,type: session-notesfor the life of the session.process-inboxowns every status transition after that.
Files (pickled-claude-plugins)
-
evals
-
trigger-evals.json 1.7 KB
Loading…
-
-
SKILL.md 11.2 KB
--- name: devlog description: Use when finishing a feature, fixing a notable bug, wrapping up a work session, or discovering something non-obvious (a gotcha, undocumented behavior, a workaround that took several tries, a constraint you only noticed after hitting it). Write the entry immediately, in the same turn - don't ask permission, don't offer, just do it. Do NOT use for routing, connecting, or moving notes (that's process-inbox), and do NOT use for one-off external-source capture (that's second-brain:capture). allowed-tools: - Read(~/.claude/vaults/**/CLAUDE.md) - Read(~/.claude/vaults/**/*.md) - Write(~/.claude/vaults/**/*.md) - Edit(~/.claude/vaults/**/*.md) - Bash(npx @techpickles/sb:*) - Bash(${CLAUDE_PLUGIN_ROOT}/scripts/diagnose-sb.sh) - Bash(${CLAUDE_PLUGIN_ROOT}/scripts/resolve-vault.sh:*) hooks: PreToolUse: - matcher: "Bash" hooks: - type: command command: "python3 \"${CLAUDE_PLUGIN_ROOT}\"/hooks/check-sb-before-call.py" --- # Devlog ## Overview Devlog is continuous, point-in-time work commentary: what happened, why, and anything surprising, captured as it happens instead of reconstructed from memory at the end of a session. One narrative session note accumulates every entry from the current session — no separate "learnings" system, no til/cursed split, everything worth remembering goes in the same place. Processing (routing, connecting, daily-linking) happens later via `/second-brain:process-inbox`. This skill only ever appends to the inbox. **Write the entry now, in the same turn as whatever triggered it.** Don't ask "should I log this?" and don't offer to — just write it. Asking is exactly the failure mode that made the old `distill-conversation` flow get skipped: by the time a session ends, you're mentally done, and a skip-able step gets skipped. ## When to write - Finished a feature or fixed a notable bug - Wrapping up a work session - Discovered something non-obvious: a gotcha, undocumented behavior, a workaround that took several tries, a constraint you only noticed after hitting it ## When not to write - Routing, connecting, or moving existing notes — that's `/second-brain:process-inbox` - Reading an external source and writing it up — that's `second-brain:capture` - Trivially obvious or one-off details you'd never reach for again - A bare project-specific fact (a config value, a repo convention) with no story behind it — that belongs in Claude Code's own memory system alone, not the vault ## Step 1: Load Configuration Don't preflight-check sb availability — just run it: ```bash npx @techpickles/sb config default && npx @techpickles/sb config vaults ``` If this fails, run `${CLAUDE_PLUGIN_ROOT}/scripts/diagnose-sb.sh` and follow its guidance rather than re-deriving the checks by hand. Then pick the vault for this session's note: ```bash ${CLAUDE_PLUGIN_ROOT}/scripts/resolve-vault.sh ``` It prints a vault name, or nothing. Use that name as `--vault <name>` on every `sb` call below that takes it (`note create`, `vault obsidian`, `inbox list`). When it prints nothing, omit `--vault` and let sb use its default. The name comes from `SB_NOTE_VAULT` if set (a per-project override), else from whichever registered vault contains the working directory. Never write a session's note to a different vault than the one the session is working in or for: a vault can have rules about what may be written to it. If no default vault is configured: ``` Second brain not configured. Run /second-brain:setup first. ``` ## Step 2: Find or create this session's devlog note Every entry from the current session lands in ONE note. Never create a second devlog note for a session that already has one. 1. Collect session context: - Current repo and branch, if in a worktree - Current work item, if any (see [Work items](#work-items) below) - Short session id (first 8 characters of the session UUID) - Source string format: `claude-code session {short-id} ({repo}, branch {branch}, {tracker} {id})` (omit fields that don't apply) ### Work items This plugin doesn't know or assume an issue tracker. If your context names a tracked work item for this session (a ticket, issue, or task ID), record it; otherwise leave it out. It can arrive two ways: - **Your instructions** (CLAUDE.md, the user's own words) say what tracker they use and which item is current. - **A hook** injects candidates when this skill loads, in this shape: ``` Work items this session ({tracker}): - {id} ({status}): {title} ``` A `PreToolUse` hook on the `Skill` tool can emit this as `additionalContext` for `second-brain:devlog`. If there are several candidates, pick the one(s) this entry is actually about. A follow-up item filed during the session is usually not what the session was about. If none fit, or nothing names a work item, omit it. Don't guess one from the branch name. 2. **Check your own memory first.** If you already created or appended to this session's devlog note earlier in this same conversation, you already have its path in context — go straight to Step 2a and append. No CLI search needed: this is the same mechanism `second-brain:insight` uses for its own accumulating session note ("if the session file already exists, was created in an earlier call in this same conversation, append"). 3. **If this is the first devlog entry this session,** go to Step 2b to create the note, and keep its path in mind for the rest of the session. 4. **Fallback, only if you've genuinely lost track of the note's path** (e.g. a context compaction wiped your memory of this session) — resolve a real path instead of guessing: - `npx @techpickles/sb vault obsidian` (add `--vault <name>` when step 1 picked one) — parse the JSON for the top-level `inbox` field (the inbox subfolder name, e.g. `📫 Inbox`). This is the precomputed field `sb note create` itself resolves the inbox folder from — don't parse `app.newFileFolderPath` directly, since that skips the Zettelkasten Prefixer plugin's folder override (`inbox` is `zkPrefixer?.folder ?? app?.newFileFolderPath`) and can point at the wrong folder. - `npx @techpickles/sb inbox list --detail` — parse the JSON, look for the entry with `type: session-notes` whose `source` field contains the current short session id, and take its `filename` - Combine: `~/.claude/vaults/{vault-name}/{inbox}/{filename}` is the real, readable/editable path. `inbox list --detail` only ever returns a bare filename, never a resolvable path on its own. - With the resolved path in hand, go to Step 2a and append. If the `inbox list --detail` search finds no entry matching the current session id (genuinely nothing to recover), fall through to Step 2b and create the note instead. ### Step 2a: Append to the existing note Read the note, append one new bullet to the end of its body (format below) using the Edit tool. Don't touch its frontmatter. ### Step 2b: Create the note Staged and written in one Bash call: ```bash VAULT={vault name from resolve-vault.sh, or empty} STAGE=$(mktemp) cat > "$STAGE" <<'EOF' - {HH:MM} {entry} EOF npx @techpickles/sb note create \ ${VAULT:+--vault "$VAULT"} \ --title "Session: {topic}" \ --source "claude-code session {short-id} ({repo}, branch {branch}, {tracker} {id})" \ --content "$(cat "$STAGE")" rm -f "$STAGE" ``` Run it sandboxed first. It works sandboxed when the vault's real path is in the sandbox's `allowWrite` (for example `~/Vaults/<vault>/**/*.md`). Name the real directory, not a symlink to it: the sandbox matches the resolved path. If the call fails with `Operation not permitted` on the vault path, the vault isn't allowlisted. Re-run the *whole* call with the Bash tool's `dangerouslyDisableSandbox: true`, not just the `sb` line: `$TMPDIR` resolves to a *different path* unsandboxed, so staging and reading have to happen in the same call under the same mode or the read can miss the file. Either way, the staged file must be a fresh `mktemp` path, not a fixed filename: `$TMPDIR` is shared across every Claude Code session on the machine, and a fixed-name file can be overwritten or read by a stale write from a different session. `sb note create` prints JSON with a `path` field. If stdout has anything before the `{` (e.g. a stray `npm warn ...` line from the caller's own `.npmrc`), extract the JSON substring starting at the first `{` rather than parsing the whole stdout blob. `sb note create` only ever writes `captured`, `source`, `repo`, `branch`, `commit` to frontmatter — it has no `--type`/`--status` flag, and `repo`/ `branch` come out as the literal `none`/`none` (devlog's `--source` format doesn't match `sb`'s own `conversation:repo=...,branch=...` parsing convention). Read the file back at the path just parsed to preserve the exact frontmatter sb wrote, then Edit it: - add `status: raw` - add `type: session-notes` - replace the `repo: none` line sb wrote with `repo: {repo-name}` (from git context if available) — don't add a second `repo:` line - replace the `branch: none` line sb wrote with `branch: {branch-name}` (from git context if available) — don't add a second `branch:` line - add `issue: {id}` if there's a current work item (see [Work items](#work-items)); a YAML list if the entry covers several. Omit the line entirely otherwise - keep `source` as sb wrote it — the `claude-code session ...` string already covers provenance for devlog's format - keep `captured` as sb wrote it Without this fixup, the note is invisible both to this skill's own "check memory / find existing note" logic (Step 2, above) and to `process-inbox`'s stage 3a filter, which requires exactly `status: raw` + `type: session-notes` to ever pick the note up. ## Entry format One bullet per entry, each with a wall-clock timestamp: ```markdown - {HH:MM} {what happened, why, anything surprising — one to three sentences, standalone enough to make sense without the surrounding conversation} ``` Prose, not a title. Write it like a short journal line, not a commit message. ### Also saved somewhere else? A record somewhere else doesn't replace the devlog entry. Claude Code's memory is what future sessions act on, an issue or ticket is the tracked work, and the devlog is the vault's record of what happened. When the discovery was also saved to one of those, write the entry anyway and say where it went: ```markdown - 14:05 The VM's guest DNS drifts off its DHCP lease after the host sleeps, so image pulls time out until it's reset. Saved as a memory (`~/.claude/projects/<project>/memory/reference_vm_dns_drift.md`). ``` Point at it, don't paste it: one or two sentences of what happened, then the pointer. The other record holds the details. Which trackers count beyond memory (and how to cite them) is up to your own instructions or the hook that nudged you here; this skill doesn't assume any. ## Constraints - One note per session — append, never create a second one for the same session. - Never ask before writing, never present a summary and wait for a selection. - Don't route, connect, move the note, or touch the daily note. That's entirely `process-inbox`'s job, later. - Frontmatter stays `status: raw`, `type: session-notes` for the life of the session. `process-inbox` owns every status transition after that.
Comments (0)
Sign in to join the conversation.
Reviews (0)
No reviews yet.
No comments yet.