菜单
[Beta] Session-start scan that surfaces relevant learnings, recent errors, and eval status before work begins. Bridges the outer loop back into the inner loop by making accumulated knowledge visible at task start. Activated via SessionStart hook or manually before major tasks.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Captures learnings, errors, corrections, and feature requests to enable continuous improvement. Use when: (1) User corrects Claude ('No, that's wrong...', 'Actually...'), (2) User requests a capability that doesn't exist, (3) Claude realizes its knowledge is outdated or incorrect, (4) A better approach is discovered for a recurring task, (5) Receiving a Handoff block from self-healing (a recurring verified heal at Recurrence-Count >= 3) to distill into a memory file or new skill. For ACTIVE runtime failures where the agent needs to apply and verify a fix mid-task, use `self-healing` instead (it files HEAL- entries with proof; self-improvement promotes accumulated patterns). Also review learnings before major tasks. For CI-only/headless learning capture, use self-improvement-ci.
Implementation + audit loop using parallel agent teams with structured simplify, harden, and document passes. Spawns implementation agents to do the work, then audit agents to find complexity, security gaps, and spec deviations, then loops until code compiles cleanly, all tests pass, and auditors find zero issues or the loop cap is reached. Use when: implementing features from a spec or plan, hardening existing code, fixing a batch of issues, or any multi-file task that benefits from a build-verify-fix cycle.
Monitors context window health throughout a session and rides peak context quality for maximum output fidelity. Activates automatically after plan-interview and intent-framed-agent. Stays active through execution and hands off cleanly to simplify-and-harden and self-improvement when the wave completes naturally or exits via handoff. Use this skill whenever a multi-step agent task is underway and session continuity or context drift is a concern. Especially important for long-running tasks, complex refactors, or any work where degraded context would silently corrupt the output. Trigger even if the user doesn't say "context surfing" — if an agent task is running across multiple steps with intent and a plan already established, this skill is live.
[Beta] Creates permanent eval cases from promoted learnings and runs regression checks against them. Turns failures into test cases that prevent silent regression. This is the outer loop's regress-test step. Use when a learning is promoted and has a clear pass/fail condition, or on cadence to verify promoted rules still hold.
[Beta] CI-only learning aggregation workflow using gh-aw (GitHub Agentic Workflows). Scans .learnings/ files on a schedule, groups entries by pattern_key, identifies promotion-ready patterns, and posts a gap report as a PR or issue comment. Use when: you want automated cross-session pattern detection in CI/headless pipelines without interactive prompts. For interactive use, use learning-aggregator.
[Beta] Cross-session analysis of accumulated .learnings/ files. Reads all entries, groups by pattern_key, computes recurrence across sessions, and outputs ranked promotion candidates. This is the outer loop's inspect step — it turns raw learning data into actionable gap reports. Use on a regular cadence (weekly, before major tasks, or at session start for critical projects). Can be invoked manually or scheduled.
| name | pre-flight-check |
| description | [Beta] Session-start scan that surfaces relevant learnings, recent errors, and eval status before work begins. Bridges the outer loop back into the inner loop by making accumulated knowledge visible at task start. Activated via SessionStart hook or manually before major tasks. |
Surfaces relevant accumulated knowledge at the start of a session. This is the bridge that connects the outer loop back into the inner loop — it makes prior learnings visible before the agent starts work.
Without this, accumulated .learnings/ are invisible to new sessions. The agent repeats mistakes that were already captured because nobody told it to look.
The SessionStart hook (scripts/pre-flight.sh) does a fast scan and outputs a brief reminder if there are relevant signals:
<pre-flight-check>
Active learnings: N entries in .learnings/
Recent errors (last 7 days): N
Promotion-ready patterns: N
Failed evals: N
High-priority items:
- [Pattern-Key]: [one-line summary] (seen N times)
- [Pattern-Key]: [one-line summary] (seen N times)
Consider running /learning-aggregator if promotion-ready count > 0.
</pre-flight-check>
If there are no signals (empty .learnings/, no failed evals), the hook outputs nothing — zero overhead.
When invoked explicitly, the pre-flight check does a deeper analysis:
Read .learnings/LEARNINGS.md, .learnings/ERRORS.md, .learnings/FEATURE_REQUESTS.md, and .learnings/HEALS.md (the last from self-healing — verified runtime fixes filed during prior sessions; surface these prominently so the agent applies known fixes before reinventing them).
For each entry, extract:
Read .evals/EVAL_INDEX.md for any failed or stale evals.
Look for unread files in .context-surfing/ (same as handoff-checker.sh but integrated).
If the user described the task area, filter learnings to:
Area matches the taskRelated Files overlap with likely-touched filesPriority: high/critical regardless of areaRecurrence-Count >= 3 (promotion-ready by recurrence threshold — need attention)## Pre-Flight Check
### Task Area: [inferred or stated]
### Relevant Learnings
| ID | Summary | Recurrence | Priority | Status |
|----|---------|-----------|----------|--------|
| LRN-... | ... | 3 | high | pending |
| ERR-... | ... | 2 | medium | pending |
### Key Warnings
- [Pattern-Key]: "Concise warning based on learning" — seen N times, last on YYYY-MM-DD
- [Pattern-Key]: "Concise warning based on learning" — seen N times, last on YYYY-MM-DD
### Failed Evals
| Eval ID | Pattern-Key | Last Failed | Recovery Action |
|---------|------------|-------------|-----------------|
| eval-... | ... | YYYY-MM-DD | ... |
### Handoff Files
- [filename] — from session on YYYY-MM-DD
### Recommendations
- [ ] Read handoff files before starting
- [ ] Run learning-aggregator (N promotion-ready patterns)
- [ ] Fix failed evals before starting new work
- [ ] Watch for [specific pattern] in [area]
.learnings/*.md — accumulated learning entries from self-improvement.evals/EVAL_INDEX.md — eval results from eval-creator.context-surfing/ — handoff files from context-surfingThis is where the blog's compounding happens:
Outer loop improves harness → pre-flight surfaces improvements → inner loop starts stronger
Every learning captured, every rule promoted, every eval created becomes visible at the next session start. The knowledge gaps get smaller with every cycle.
The hook script can be extended to use a local cache file (.pre-flight-cache.json) storing last-known state — entry counts, scan date, high-priority items — so the next session start only re-scans entries newer than the cached state. This would enable delta reporting ("since your last session, 2 new errors were logged and 1 pattern crossed the promotion threshold") and keep the hook near-instant regardless of how large .learnings/ grows. Not implemented today — the current hook scans directly on every session start.
.learnings/ files (read-only)