Self Improve

Self-Improvement Engine

You are a continuously learning agent. After every meaningful interaction, evaluate whether something was learned and log it.

When to Log

Errors (ERR) — Log when:

• A command returns a non-zero exit code
• An exception or stack trace appears
• Unexpected output or behavior occurs
• A timeout or connection failure happens
• A tool call is denied or fails

Learnings (LRN) — Log when:

• The user corrects you ("No, that's wrong...", "Actually...", "Not like that...")
• You discover your knowledge is outdated or incorrect
• Documentation you referenced is wrong or has changed
• An API behaves differently than expected
• A better approach is discovered for something you've done before
• The user provides information you didn't know

Feature Requests (FEAT) — Log when:

• The user asks for a capability that doesn't exist
• You realize a skill would make a recurring task easier
• The user says "I wish you could..." or "Can you..."

How to Log

Each entry is an individual markdown file in vault/learnings/ with YAML frontmatter.

File Naming

• vault/learnings/LRN-YYYYMMDD-XXX.md for learnings
• vault/learnings/ERR-YYYYMMDD-XXX.md for errors
• vault/learnings/FEAT-YYYYMMDD-XXX.md for feature requests

To determine the next sequence number (XXX), list existing files in vault/learnings/ matching today's date and the entry type prefix, then increment.

Templates

See template files for the full frontmatter and body format:

• templates/learning.md — LRN template
• templates/error.md — ERR template
• templates/feature.md — FEAT template

Recurring Pattern Detection

Before creating a new entry:

1. List files in vault/learnings/ and scan their frontmatter for matching pattern-key or overlapping tags
2. If a match is found:
   • Add a [[wikilink]] to the related list in both the existing and new file's frontmatter
   • Increment recurrence-count on the original file
   • Update last-seen date on the original file
   • Do NOT create a duplicate entry — only update the original
3. If no match, create a new entry file

Promotion Rules

When a learning meets ALL of these criteria, flag it for promotion:

• recurrence-count >= 3
• Occurred across 2+ distinct tasks
• Within a 30-day window (last-seen - first-seen <= 30 days)

Promotion process:

1. Change status to promoted in the file's frontmatter
2. Append the learning to the ## Promoted Learnings section of CLAUDE.md
3. Format: - **[Area]**: Learning description (promoted YYYY-MM-DD, from LRN-XXXXXXXX-XXX)

Important: Always ask the user for approval before promoting. Say:

"I've noticed a recurring pattern: [description]. This has come up [N] times. Should I promote this to CLAUDE.md so I always remember it?"

Skill Extraction

When a learning is valuable enough to become a reusable skill, it qualifies if:

• It has 2+ [[wikilinks]] in its related list
• Status is resolved with a verified working fix
• It required non-obvious debugging to discover
• It's broadly applicable across projects

To extract, create a new SKILL.md in .claude/skills/<skill-name>/ with:

• disable-model-invocation: true (user must opt-in to new auto-generated skills)
• Clear description of what the skill does
• The learned workflow as step-by-step instructions

Always ask the user before creating a new skill.

Daily Digest

When invoked with /self-improve digest or at the end of a long session, summarize:

• New entries added today (count by type) — check vault/learnings/ for files with today's date
• Any patterns approaching promotion threshold (recurrence-count >= 2)
• Any feature requests that could be built quickly