Stats
Actions
Tags
From ClaudeHut
Use in the Learn phase at the end of every task, before declaring done - dispatches the learner agent to record what was learned (conventions, pitfalls, reuse points, decisions) to the cross-session store and refresh the committed memory index, then closes the phase. Runs inline on the main thread (it owns the state write).
How this skill is triggered — by the user, by Claude, or both
Slash command
/claudehut:capture-learningsThis skill is limited to the following tools:
The summary Claude sees in its skill listing — used to decide when to auto-load this skill
```
NO TASK ENDS WITHOUT A LEARN PASS
If you learned a project pattern, a pitfall, or a reuse point, record it before stopping. The Stop gate
blocks "done" until this runs. Runs inline on the main thread — the learner agent does the recording in
isolation; this skill owns the state write (the learner has no Bash).
Small-tier inline learn (when nothing novel surfaced). If the tier is small AND the task produced
no new pitfall, convention, or reuse point (it only confirmed existing patterns), skip the learner
dispatch and append ONE line inline to
${CLAUDE_PROJECT_DIR}/.claude/claudehut/learnings.jsonl (1 tool call — the Stop gate checks content,
not author):
{"id":"<uuid>","ts":"<iso>","project":"<name>","phase":"learn","category":"convention","trigger":"<file-pattern keywords>","learning":"<one sentence: what was confirmed>","evidence":"<file:line>","confidence":0.6,"hits":1}
Then go to step 4 (skip the learner dispatch AND the merge script — one confirmed line needs neither). Dispatch the full learner ONLY when something genuinely new was found — a learner round-trip to record "nothing new" is pure latency. (Full tier always dispatches: a task that went through Brainstorm/Spec/Plan has decisions worth distilling.)
Dispatch claudehut:claudehut-learner (Agent tool) with a short task summary: the task dir
(tasks/NNNN-<slug>/), the decisions made, surprises hit, reuse points created, and Review findings.
The learner does the judgment only:
${task_dir}/learn-candidates.jsonl — one JSON object per line
({category, trigger, learning, evidence, confidence?}). It does not dedup, assign ids, promote, or
prune (that is the script's job in step 2).reuse-index.json with anything newly built.MEMORY.md (the committed index) when a new topic/category/artifact appears.Run the deterministic merge on the candidates the learner produced — this is what actually writes the cross-session store, exactly and in milliseconds (the learner must NOT do this math by reasoning):
"${CLAUDE_PLUGIN_ROOT}/scripts/merge-learnings.sh" --candidates "${CLAUDE_PROJECT_DIR}/.claude/claudehut/tasks/NNNN-<slug>/learn-candidates.jsonl"
It dedups against the canonical ${CLAUDE_PROJECT_DIR}/.claude/claudehut/learnings.jsonl (NOT
.claudehut/memory.jsonl — only the canonical file is injected at the next session's SessionStart) by
category + normalized trigger → merge (hits++, confidence = min(+0.05, 1.0), ts=now) or append a
new L-#### line; promotes proven pitfalls (hits≥5 ∧ confidence≥0.85) into the matching
.claude/rules/ file; prunes decayed noise. It prints a {added, merged, promoted, dropped} report.
If native auto-memory is enabled, mirror a short narrative there — convenience only, not the source of truth.
Main thread closes the phase after the merge runs:
claudehut-state --session ${CLAUDE_SESSION_ID} set-phase learn
REQUIRED NEXT: the task may now end (the Stop gate is satisfied). The next session's SessionStart will inject the top of what you recorded.
npx claudepluginhub taipt1504/claudehut --plugin claudehutGuides creation, editing, and verification of skills for AI coding agents using test-driven development with subagent scenarios. Use when authoring or debugging skills.