/learn-rule

name: learn-rule
description: Capture a correction or lesson as a persistent learning rule with category, mistake, and correction. Stores, categorises, and retrieves rules for future sessions. Use after mistakes or when the user says "remember this", "don't forget", "note this", or "learn from this".
user-invocable: true

Learn Rule

Capture a lesson from the current session into permanent memory.

Trigger

Use when the user says "remember this", "add to rules", "don't do that again", or after a mistake is identified.

Workflow

1. Identify the lesson — what mistake was made? What should happen instead? 2. Format the rule with full context. 3. Propose the addition and wait for user approval. 4. After approval, persist to LEARNED section or project memory.

Format

[LEARN] Category: One-line rule
Mistake: What went wrong
Correction: How it was fixed

Wiki-scoped rules

Append `Wiki: <slug>` to bind the rule to a single pro-workflow wiki. The rule loads only when that wiki is in scope, avoiding cross-project pollution:

[LEARN] Editing: Cite a sources.md row before adding any wiki claim.
Wiki: agent-memory

The capture hook auto-detects `Wiki: <slug>` and links the learning to that wiki via `learnings_wiki`.

Categories

| Category | Examples | |----------|---------| | Navigation | File paths, finding code, wrong file edited | | Editing | Code changes, patterns, wrong approach | | Testing | Test approaches, coverage gaps, flaky tests | | Git | Commits, branches, merge issues | | Quality | Lint, types, style violations | | Context | When to clarify, missing requirements | | Architecture | Design decisions, wrong abstractions | | Performance | Optimization, O(n^2) loops, memory |

Example

Recent mistake: Edited wrong utils.ts file

[LEARN] Navigation: Confirm full path when multiple files share a name.

Add to LEARNED section? (y/n)

Guardrails

• Always wait for user approval before persisting.
• Keep rules to one line — specific and actionable.
• Bad: "Write good code". Good: "Always use snake_case for database columns".
• Include the mistake context so the rule makes sense later.

Output

• The proposed `[LEARN]` rule with category
• Confirmation after persisting