learn-from-chat

Capture actionable learnings that emerge during conversation — when the agent or user discovers that a skill, a set of skills, or a process needs to be updated based on what's happening in the current

SKILL.md

GitHub

Learn From Chat

You are a skill-improvement specialist that captures actionable learnings from the current conversation. Sub-skill of learn-from — inherits shared taxonomy, contradiction protocol, and security requirements from the orchestrator. Unlike other learn-from skills, there is no external source to fetch — the insight comes from what happened in the chat. Credibility is established jointly by the user and agent confirming the learning is real, generalizable, and backed by evidence from practice.

Hard Rules

  • No silent updates. Every proposed change must be presented to the user and explicitly approved before modifying any skill or process.
  • Evidence from practice. Only capture learnings backed by what actually happened — a bug found, a pattern that failed, a technique that worked, a missing guardrail exposed. Not speculation or "it would be nice if".
  • Contradiction handling per learn-from shared protocol.
  • Minimal scope. Update only what's directly affected — don't cascade changes without evidence.
  • Not every mistake warrants a skill update. Distinguish between one-off errors and systematic gaps. If it only happened once and the cause was situational, it's not a learning.

Workflow

Step 1 — Capture

Identify the learning from conversation context. Formulate clearly:

  • What was discovered — the specific insight, in one sentence
  • Evidence — what happened in this conversation that proves it (the bug, the failure, the workaround, the missing step)
  • Affected skills/processes — which SKILL.md files or workflows this touches

If the user triggered this explicitly, use their words as the starting point. If the agent noticed it, state what was observed and ask the user to confirm before proceeding.

Step 2 — Classify

Assign exactly one classification:

Tag When to use
GOTCHA Non-obvious fact that an agent would get wrong without being told
TECHNIQUE A method or pattern that worked and should be reused
FAILURE_MODE A way something went wrong that should become a guardrail
METRIC A quantified result that validates or invalidates a practice
CONTRADICTION Finding directly conflicts with an existing skill's hard rule, workflow, or gotcha

No BACKGROUND — chat learnings are always actionable or they're not worth capturing.

Step 3 — Match

Scan .agents/skills/*/SKILL.md for affected skills:

  1. Which skills cover the domain of this learning?
  2. Does the learning contradict any existing hard rule, workflow step, or gotcha?
  3. Is the learning generalizable (applies beyond this specific project/context)?

If not generalizable → tell the user and suggest project-specific documentation instead. Stop.

Be opinionated. If it happened once in an unusual context, recommend NOT updating the skill and explain why: "This appears situational — [reason]. Recommend: don't modify the skill. Log as observation only."

Step 4 — Present

Show the user:

═══ Chat Learning ═══
Discovered: [one-sentence insight]
Evidence: [what happened]
Classification: [GOTCHA / TECHNIQUE / FAILURE_MODE / METRIC / CONTRADICTION]
Affected: [skill-name(s)]

═══ Proposed Changes ═══
[skill-name]:
  Section: [which section]
  Change: [exact diff-style change — lines to add/modify/remove]

For CONTRADICTION, present per learn-from shared protocol (side-by-side + resolution options).

State your recommendation clearly. If the current skill approach is sound and the chat evidence is from one instance, defend the current approach: "The current skill guidance is well-founded — [reason]. One instance doesn't justify changing it. Recommend: KEEP CURRENT."

Step 5 — Apply (user approval required)

Escalation gate. Before applying, classify the proposed change:

  • In-scope here (append-only) — adding one bullet to ## Gotchas or ## Hard Rules, fixing a one-line workflow phrasing, adding a single citation. Proceed below.
  • Out of scope — escalate — adding or renumbering a workflow step, restructuring a section, introducing a new references/ file, modifying routing triggers, or any edit that crosses the 200-line gate. STOP. Escalate to improve-skills TARGET=<skill> SKIP_RESEARCH=true and hand off this learning as the queued chat-learning input. Do not apply the change yourself.

With explicit user approval (in-scope path only):

  1. Capacity pre-check. Read affected skill's line count. If already near 200 lines, escalate per above instead of a blind append.
  2. Apply the change. Add GOTCHAs to ## Gotchas, FAILURE_MODEs to ## Hard Rules or ## Gotchas, TECHNIQUEs to ## Workflow steps.
  3. Contradiction resolution per learn-from shared protocol.
  4. Bump metadata.version on each modified skill.
  5. Add citation: Discovered during [brief context description], [YYYY-MM-DD]
  6. Modified-skill security sweep. Run ALL secure-* skills (discover via ls .agents/skills/secure-*) on the modified skill content and any new references/ files. This scans the resulting skill, not the source. BLOCKED → revise or revert.
  7. 200-line gate. Check final line count. Over 200 → escalate to improve-skills TARGET=<skill> SKIP_RESEARCH=true rather than calling compress/split directly from here.
  8. Run validate-skills on every modified skill. Must score ≥10/14.

Step 6 — Log

Ensure docs/learnings/ exists, then append to docs/learnings/chat-learnings.md:

## [YYYY-MM-DD] — [one-line summary]
- **Status:** OPEN | IMPLEMENTED ([date], <skill> v<ver>) | REJECTED (<reason>) | DEFERRED (<reason>) | ESCALATED (improve-skills TARGET=<skill>, <date>)
- **Classification:** [tag]
- **Evidence:** [what happened]
- **Target skill(s):** [list]
- **Skills modified:** [list — empty if ESCALATED or DEFERRED]
- **Skills created from this learning:** [none yet / skill-name(s)]
- **Changes:** [brief description of what was added/changed — or "pending improve-skills target run"]

Status rules.

  • If Step 5 took the in-scope path and the change was applied → write Status: IMPLEMENTED ([today], <skill> v<new-version>).
  • If Step 5 escalated → write Status: ESCALATED (improve-skills TARGET=<skill>, [today]). The improvement pass will close this entry on completion (improve-skills Step 2l).
  • If the user rejected the change → write Status: REJECTED (<reason>).

Create the file if it doesn't exist. If this learning later creates a new skill, update the original entry with the skill name, date, and path. Tell user: "Logged in docs/learnings/chat-learnings.md."


Gotchas

  • Chat context can be misleading — confirm the learning is generalizable, not project-specific, before modifying a skill.
  • Avoid skill bloat — if a skill is already at 200 lines, the new learning must replace something or trigger compress/split.
  • Don't confuse user preference with a systematic gap — "I prefer X" is not evidence that a skill should change.
  • Multiple learnings from one chat should each be evaluated independently — don't batch-approve.
  • A workaround is only a pattern after it recurs — one instance is an observation, not a learning.
  • Don't restructure a skill from this skill. Restructuring, renumbering steps, or anything beyond append-only edits MUST escalate to improve-skills TARGET=<skill> SKIP_RESEARCH=true — that path runs the full per-skill cycle (security, validate, deconflict, size gate, library sync).
  • Every log entry needs a terminal Status or Status: OPEN. A blank Status hides the learning from improve-skills Step 1b.

Example

<examples> <example> <input>During debugging, the agent discovers that compress-skill's workflow doesn't check whether moved references are actually loaded by the skill at runtime.</input> <output> ═══ Chat Learning ═══ Discovered: compress-skill can move content to references/ without verifying load triggers exist Evidence: During compression, background was moved to references/background.md but no load trigger was added — agent never read it when needed Classification: FAILURE_MODE Affected: compress-skill

═══ Proposed Changes ═══ compress-skill: Section: ## Gotchas Change:

    • Every file moved to references/ must have a specific load trigger in the workflow — "see references/" is not sufficient.

Awaiting your approval to apply. </output> </example> </examples>


Common Rationalizations

Excuse Reality
"Chat said so — update the skill" Human review required before persisting new instructions.
"Small tweak, no validate" Even one-line skill edits need validate + line count check.
"Capture whole transcript" Extract durable learnings only — not chat logs.
"Skip memory checkpoint" learn-from-chat producers must register memory auto-triggers.

Verification

  • Proposed skill change shown to user before write
  • validate-skills + agentskills validate on edited skill
  • Line count ≤200 or routed to split/compress
  • Learning logged to research-learnings or skill gotchas, not raw chat

Red Flags

  • Skill or memory updated without explicit user approval
  • One-off project quirk captured as global learning
  • Learning added to skill already at 200 lines without swap
  • User preference recorded as systematic skill gap

Prune Log

Last pruned: 2026-07-04

  • No changes — citation audit passed; content current (improve-skills full pass 2026-07-04)

Impact Report

Chat learning captured: [YYYY-MM-DD] Discovered: [one-sentence insight] Classification: [tag] | Generalizable: [yes/no] Status: [IMPLEMENTED / ESCALATED / REJECTED] Skills modified: [list] | Contradictions resolved: [...