# Learn > Capture corrections and follow-ups in private local state, select a compact retrospective, inspect promotion candidates, or emit a review-only learning diff. Use when a user corrects an agent, asks to remember a workflow lesson, requests a retrospective, or wants to improve reusable instructions without silently changing the current repository. - Skill: `ghostlygawd/learn` (Agent Skill, multi-file: 6 files) - Install (CLI): `npx skillmds@latest add ghostlygawd/learn` - Raw SKILL.md: https://api.skillmd.com/api/skills/ghostlygawd/learn/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: AI & ML - Author: GhostlyGawd (https://skillmd.com/u/ghostlygawd) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/ghostlygawd/learn --- # Learn Use the bundled deterministic CLI. Resolve `scripts/learn.py` relative to this file; never substitute a project-local script with the same name. Existing project instructions, skills, hooks, and configuration remain authoritative. ## Capture private signals Keep entries short and omit source code, prompts, credentials, or personal data. The runtime redacts common secret and identity shapes before persistence. ```bash python3 /scripts/learn.py correction add \ --session SESSION_ID --text "Inspect existing instructions before proposing changes." python3 /scripts/learn.py followup add \ --session SESSION_ID --text "Recheck the portable install story." ``` List entries with `correction list` or `followup list`. Close a follow-up only after evidence exists: ```bash python3 /scripts/learn.py followup done FOLLOWUP_ID ``` ## Retrospect and prepare candidates `retro plan` selects no more than three recent high-signal events. It does not invent a lesson or modify a repository. A reusable candidate is an explicit, human-readable proposal: ```bash python3 /scripts/learn.py retro plan --json python3 /scripts/learn.py candidate add \ --kind correction --domain "portable review" \ --summary "Existing instructions stay authoritative" \ --procedure "Inspect first; emit a patch; never overwrite consumer instructions." ``` ## Promote by reviewable diff only Learn cannot apply a promotion. On an explicit request, emit a unified diff for a confined relative target and let the user or their normal review workflow decide whether to apply it: ```bash python3 /scripts/learn.py promote diff CANDIDATE_ID \ --repository /path/to/repository --target LEARNINGS.md ``` Never redirect this output into a file without explicit user approval. Read [the promotion contract](references/promotion.md) before targeting an existing instruction artifact. ## Protect privacy Learn stores sanitized sidecar evidence below `~/.recursive-harness/learn`, never the active repository. It accepts no state-path argument or environment override. Audit aggregate metadata without printing captured text, and preview deletion before applying it: ```bash python3 /scripts/learn.py privacy audit --json python3 /scripts/learn.py privacy retain --days 30 python3 /scripts/learn.py privacy retain --days 30 --apply python3 /scripts/learn.py privacy purge python3 /scripts/learn.py privacy purge --apply ``` Retention scrubs expired raw text but preserves identifiers, timestamps, and lifecycle status. Malformed timestamps are reported and retained for explicit review instead of being silently destroyed. `retain`, like `purge`, is a dry run until `--apply` is present. Read [the privacy contract](references/privacy.md) for the exact boundary and unsupported cases. Installing or invoking this skill must not edit `AGENTS.md`, `CLAUDE.md`, `.claude/`, `.codex/`, hooks, workflows, or any other repository file.