Claude Cursor Skill

meta-learn

Captures a completed skill session in depth: three extraction questions (what surprised you, what was wrong, what was missing) become falsifiable pattern statements, logged to /context/skill-sessions.md for meta-synthesis. Trigger on: "capture what we learned", "log this session"

LLM Mart · 0 points · 0 views 0 listing impressions 0 install-command copies
Download stefanoskarakasis-product-marketing-skills-pmm-meta_skills_meta-learn-5de42dc.zip · 4 KB

Install

skills CLI npx skills add https://github.com/stefanoskarakasis/Product-Marketing-Skills/tree/main/pmm-meta/skills/meta-learn
Claude Code claude plugin marketplace add https://llmmart.ai/marketplace.json && claude plugin install stefanoskarakasis-product-marketing-skills@llmmart
Git git clone https://github.com/stefanoskarakasis/Product-Marketing-Skills.git

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

Skill manifest

meta-learn

The deliberate, on-demand deep-dive into a completed skill session — for when the automatic one-line Learning Close every T1/T2 skill already writes isn't enough. Asks what the session actually taught the user — not what the skill's output said, but what surprised them, what they disagreed with, or what was missing. Turns real answers into a specific, falsifiable statement and logs it to the same file, in the same format, every T1/T2 skill's own Learning Close writes to — so meta-synthesis reads one consistent log regardless of which produced any given row.

Trigger

  • When: The user wants a deeper capture of a completed skill session than the automatic one-line row every T1/T2 skill already logs at its own close (see that skill's Learning Close step). Use this for a session that deserves real reflection — something surprised the user, a recommendation felt wrong, or important context was missing — and a structured three-question interview is worth the extra few minutes.

  • Not for: Auditing a SKILL.md's structure → use meta-review. Checking whether a skill's output itself is correct → use meta-verify. Detecting patterns across multiple sessions or proposing guardrails → use meta-synthesis, which reads what this skill logs.

  • Example prompts:

    • "Capture what we learned from that retro"
    • "Log this session"
    • "What surprised you about that positioning run?"
    • "Save the learnings before we close"

Inputs

  • Args: The completed session's skill name and a short description of its output. N/A if invoked immediately after a skill session in the same conversation — infer both from context.
  • Defaults: If the skill name isn't obvious, ask for it before proceeding.
  • Context keys: /context/skill-sessions.md — appended to, created if it doesn't exist yet.

Pre-flight

  • This skill is context-agnostic. Do not load /foundation/brain.md and do not let prior company context shape what counts as a pattern — extraction comes from the user's answers alone.
  • If there's no completed session to close out (skill hasn't produced output yet), stop and say so — this isn't a mid-session check-in.

Steps

Step 1: Confirm the Session Closed

Confirm which skill just ran and what it produced, in one line:

"Closing out [skill-name] — produced [one-line description of output]. Three quick questions before we log this."

Step 2: Ask the Three Extraction Questions

Ask all three in one message. Wait for complete answers before proceeding.

"1. What surprised you? Something unexpected the skill surfaced that you hadn't considered.

  1. What was wrong or off? Any recommendation you pushed back on or felt missed the mark. Be specific: what, and why?

  2. What was missing? Context the skill didn't have, a question it didn't ask, an output it should have produced but didn't."

If all three come back as "nothing," skip to Step 5 — log a clean close, no pattern extraction needed.

Step 3: Make Each Answer Falsifiable

For any answer with real content, restate it as a specific, testable statement — not a vague impression.

  • ❌ Vague: "communication could be better"
  • ✅ Falsifiable: "when the launch timeline slips more than 10 days, the brief didn't flag which stakeholders needed to be renotified"

If an answer can't be sharpened into something falsifiable, it isn't a pattern yet — note it as a loose observation rather than forcing it into a false-precision statement.

Step 4: Show What Will Be Logged

Before writing, show the user the exact log entry:

🔁 SESSION LEARNING

Skill: [name]
Date: [date]
Pattern: [falsifiable statement, or "none — clean session"]
Source: [surprised / wrong / missing]

Log this? (yes / edit / skip)

Never log a pattern without this confirmation. A clean close (no pattern) logs automatically without this gate — there's nothing to approve.

Step 5: Append to the Session Log

Append one entry to /context/skill-sessions.md, in the format defined in product-marketing-context/.claude-plugin/skill-sessions-format.md (Type A — execution session):

type: execution
skill: [name]
session_date: [YYYY-MM-DD]
pattern: [falsifiable statement, or "none"]
source: [surprised / wrong / missing / N/A]

This is the only file this skill writes to. meta-synthesis is what reads across many of these rows later to find what repeats — this skill's only job is making sure each individual row is worth reading.

Outputs

  • Files written: /context/skill-sessions.md — one appended row per session, per Step 5.
  • Chat output format: The three questions (Step 2), the falsifiable restatement if applicable (Step 3), and the logged-entry confirmation (Step 4).
  • External side effects: N/A

Verification

  • The session log entry exists in /context/skill-sessions.md with today's date and the correct skill name.
  • If a pattern was captured, it's stated as a specific, falsifiable claim — not a vague impression.
  • Nothing was written without the Step 4 confirmation, except a clean close with no pattern.

Do Not Use For

  • meta-review — auditing a SKILL.md's own structure against the spec
  • meta-verify — checking whether a skill's output itself is correct before it goes out
  • meta-synthesis — detecting patterns across 2+ logged sessions and proposing guardrails; this skill only produces the raw log rows meta-synthesis later reads

Operating Rules

  • Extract from the user's real answer, not from the skill's own output. The skill already said what it produced. This is about what the user noticed that it didn't say.
  • Falsifiable or nothing. A pattern statement that can't be proven wrong isn't useful to meta-synthesis later — sharpen it or log it as a loose observation, don't force false precision.
  • Never force a pattern. If the user says "nothing surprised me," accept that and log a clean close. Don't manufacture insight to seem thorough.
  • One log, one format. Every session appends to /context/skill-sessions.md in the same YAML shape — meta-synthesis depends on consistent structure to scan across sessions.
  • Nothing writes without the confirmation gate, except the automatic clean-close log entry when there's genuinely nothing to approve.
  • This skill doesn't decide what's a repeating pattern. That's meta-synthesis's job, reading across many logged sessions. This skill's job ends at logging one session well.

Quality Gate

Check Standard Pass =
All three questions asked Surprised / wrong / missing, in one message Yes
Pattern is falsifiable Specific claim, not a vague impression Yes or N/A (clean close)
Confirmation shown before write Exact log entry shown, user confirmed Yes or N/A (clean close)
Log entry appended /context/skill-sessions.md has the new row Yes
Clean closes handled "Nothing" answers don't force a fake pattern Yes
Files (product-marketing-skills)
  • evals
    • meta-learn.eval.md 3.5 KB
      Loading…
  • SKILL.md 7.5 KB
    ---
    name: meta-learn
    version: 3.1.0
    description: >
      Captures a completed skill session in depth: three extraction questions (what surprised
      you, what was wrong, what was missing) become falsifiable pattern statements, logged to
      /context/skill-sessions.md for meta-synthesis. Trigger on: "capture what we learned",
      "log this session", "save the learnings", "what did we learn", or any request to encode
      deeper insights from a completed skill session.
    metadata:
      author: Stefanos Karakasis
      context: context-agnostic
      quality_gate: true
    last_updated: 2026-08-22
    ---
    
    # meta-learn
    
    The deliberate, on-demand deep-dive into a completed skill session — for
    when the automatic one-line Learning Close every T1/T2 skill already writes
    isn't enough. Asks what the session actually taught the user — not what the
    skill's output said, but what surprised them, what they disagreed with, or
    what was missing. Turns real answers into a specific, falsifiable statement
    and logs it to the same file, in the same format, every T1/T2 skill's own
    Learning Close writes to — so `meta-synthesis` reads one consistent log
    regardless of which produced any given row.
    
    ## Trigger
    
    - **When:** The user wants a deeper capture of a completed skill session
      than the automatic one-line row every T1/T2 skill already logs at its own
      close (see that skill's Learning Close step). Use this for a session that
      deserves real reflection — something surprised the user, a recommendation
      felt wrong, or important context was missing — and a structured
      three-question interview is worth the extra few minutes.
    
    - **Not for:** Auditing a `SKILL.md`'s structure → use `meta-review`.
      Checking whether a skill's output itself is correct → use `meta-verify`.
      Detecting patterns across multiple sessions or proposing guardrails → use
      `meta-synthesis`, which reads what this skill logs.
    
    - **Example prompts:**
      - "Capture what we learned from that retro"
      - "Log this session"
      - "What surprised you about that positioning run?"
      - "Save the learnings before we close"
    
    ## Inputs
    
    - **Args:** The completed session's skill name and a short description of
      its output. `N/A` if invoked immediately after a skill session in the
      same conversation — infer both from context.
    - **Defaults:** If the skill name isn't obvious, ask for it before proceeding.
    - **Context keys:** `/context/skill-sessions.md` — appended to, created if
      it doesn't exist yet.
    
    ## Pre-flight
    
    - This skill is context-agnostic. Do not load `/foundation/brain.md` and do
      not let prior company context shape what counts as a pattern — extraction
      comes from the user's answers alone.
    - If there's no completed session to close out (skill hasn't produced output
      yet), stop and say so — this isn't a mid-session check-in.
    
    ## Steps
    
    ### Step 1: Confirm the Session Closed
    
    Confirm which skill just ran and what it produced, in one line:
    
    > "Closing out [skill-name] — produced [one-line description of output].
    > Three quick questions before we log this."
    
    ### Step 2: Ask the Three Extraction Questions
    
    Ask all three in one message. Wait for complete answers before proceeding.
    
    > "1. **What surprised you?** Something unexpected the skill surfaced that
    > you hadn't considered.
    >
    > 2. **What was wrong or off?** Any recommendation you pushed back on or
    > felt missed the mark. Be specific: what, and why?
    >
    > 3. **What was missing?** Context the skill didn't have, a question it
    > didn't ask, an output it should have produced but didn't."
    
    If all three come back as "nothing," skip to Step 5 — log a clean close, no
    pattern extraction needed.
    
    ### Step 3: Make Each Answer Falsifiable
    
    For any answer with real content, restate it as a specific, testable
    statement — not a vague impression.
    
    - ❌ Vague: "communication could be better"
    - ✅ Falsifiable: "when the launch timeline slips more than 10 days, the
      brief didn't flag which stakeholders needed to be renotified"
    
    If an answer can't be sharpened into something falsifiable, it isn't a
    pattern yet — note it as a loose observation rather than forcing it into a
    false-precision statement.
    
    ### Step 4: Show What Will Be Logged
    
    Before writing, show the user the exact log entry:
    
    ````
    🔁 SESSION LEARNING
    
    Skill: [name]
    Date: [date]
    Pattern: [falsifiable statement, or "none — clean session"]
    Source: [surprised / wrong / missing]
    
    Log this? (yes / edit / skip)
    ````
    
    Never log a pattern without this confirmation. A clean close (no pattern)
    logs automatically without this gate — there's nothing to approve.
    
    ### Step 5: Append to the Session Log
    
    Append one entry to `/context/skill-sessions.md`, in the format defined in
    `product-marketing-context/.claude-plugin/skill-sessions-format.md`
    (Type A — execution session):
    
    ````yaml
    type: execution
    skill: [name]
    session_date: [YYYY-MM-DD]
    pattern: [falsifiable statement, or "none"]
    source: [surprised / wrong / missing / N/A]
    ````
    
    This is the only file this skill writes to. `meta-synthesis` is what reads
    across many of these rows later to find what repeats — this skill's only
    job is making sure each individual row is worth reading.
    
    ## Outputs
    
    - **Files written:** `/context/skill-sessions.md` — one appended row per
      session, per Step 5.
    - **Chat output format:** The three questions (Step 2), the falsifiable
      restatement if applicable (Step 3), and the logged-entry confirmation
      (Step 4).
    - **External side effects:** N/A
    
    ## Verification
    
    - The session log entry exists in `/context/skill-sessions.md` with today's
      date and the correct skill name.
    - If a pattern was captured, it's stated as a specific, falsifiable claim —
      not a vague impression.
    - Nothing was written without the Step 4 confirmation, except a clean close
      with no pattern.
    
    ## Do Not Use For
    
    - **meta-review** — auditing a `SKILL.md`'s own structure against the spec
    - **meta-verify** — checking whether a skill's output itself is correct
      before it goes out
    - **meta-synthesis** — detecting patterns across 2+ logged sessions and
      proposing guardrails; this skill only produces the raw log rows
      meta-synthesis later reads
    
    ## Operating Rules
    
    - **Extract from the user's real answer, not from the skill's own output.**
      The skill already said what it produced. This is about what the user
      noticed that it didn't say.
    - **Falsifiable or nothing.** A pattern statement that can't be proven wrong
      isn't useful to `meta-synthesis` later — sharpen it or log it as a loose
      observation, don't force false precision.
    - **Never force a pattern.** If the user says "nothing surprised me," accept
      that and log a clean close. Don't manufacture insight to seem thorough.
    - **One log, one format.** Every session appends to
      `/context/skill-sessions.md` in the same YAML shape — `meta-synthesis`
      depends on consistent structure to scan across sessions.
    - **Nothing writes without the confirmation gate**, except the automatic
      clean-close log entry when there's genuinely nothing to approve.
    - **This skill doesn't decide what's a repeating pattern.** That's
      `meta-synthesis`'s job, reading across many logged sessions. This skill's
      job ends at logging one session well.
    
    ## Quality Gate
    
    | Check | Standard | Pass = |
    |---|---|---|
    | All three questions asked | Surprised / wrong / missing, in one message | Yes |
    | Pattern is falsifiable | Specific claim, not a vague impression | Yes or N/A (clean close) |
    | Confirmation shown before write | Exact log entry shown, user confirmed | Yes or N/A (clean close) |
    | Log entry appended | `/context/skill-sessions.md` has the new row | Yes |
    | Clean closes handled | "Nothing" answers don't force a fake pattern | Yes |
    

Comments (0)

Sign in to join the conversation.

No comments yet.

Reviews (0)

No reviews yet.