Self-Optimization

Use this skill to close the loop after real work. The goal is not just to log what went wrong. The goal is to convert signal from mistakes, corrections, and repeated effort into stronger future behavior.

Core Loop

1. 1. Detect meaningful signal.
2. Capture it in .learnings/.
3. De-duplicate and link related entries.
4. Promote stable patterns into durable guidance.
5. Extract reusable skills when the pattern is broad and proven.

Quick Reference

Situation	Action
Command, tool, or integration fails unexpectedly	Append an entry to INLINECODE1
User corrects the agent or provides missing facts

Append an entry to .learnings/LEARNINGS.md | | A better repeatable approach is discovered | Append an entry to .learnings/LEARNINGS.md | | User asks for a missing capability | Append an entry to .learnings/FEATURE_REQUESTS.md | | Same issue keeps reappearing | Link entries, bump priority, and consider promotion | | Pattern is stable across tasks | Promote to AGENTS.md, CLAUDE.md, TOOLS.md, SOUL.md, or .github/copilot-instructions.md | | Pattern is reusable beyond one repo | Extract a new skill scaffold |

Detection Triggers

Capture a learning when any of these happen:

• - The first attempt was wrong and needed correction.
• A tool or command failed in a non-obvious way.
• The user revealed a project convention that was not documented.
• The agent discovered a stronger pattern than the one it started with.
• The same workaround or warning has appeared more than once.
• The user asked for a capability the current system does not provide.

Skip noisy one-off trivia. Capture things that would realistically save a future session time, confusion, or rework.

Log Files

Create a local .learnings/ directory in the workspace or in the OpenClaw workspace.

CODEBLOCK0

LEARNINGS.md

Use for:

• - corrections
• knowledge gaps
• best practices
• project conventions
• improved workflows

Template:

INLINECODE12

ERRORS.md

Use for:

• - command failures
• exceptions
• bad tool assumptions
• API or integration breakage

Template:

INLINECODE14

FEATURE_REQUESTS.md

Use for:

• - missing tooling
• automation requests
• product gaps
• missing agent behaviors

Template:

INLINECODE16

ID Format

Use TYPE-YYYYMMDD-XXX.

• - LRN for learning
• INLINECODE19 for error
• INLINECODE20 for feature request

Examples:

• - INLINECODE21
• INLINECODE22
• INLINECODE23

Promotion Rules

Promote an entry when it becomes more valuable as guidance than as a historical note.

Target	Promote When
INLINECODE24	Project facts, conventions, or recurring gotchas
INLINECODE25

Workflow rules, delegation patterns, automation steps |
| .github/copilot-instructions.md | Repo guidance that should reach Copilot |
| TOOLS.md | Tool quirks, auth requirements, environment gotchas |
| SOUL.md | Behavioral or communication rules for OpenClaw sessions |

Promotion checklist:

1. 1. Distill the learning into a short prevention rule.
2. Add it to the right target file.
3. Update the original entry status to promoted.
4. Record where it was promoted.

Recurrence And Dedupe

Before creating a new entry for a familiar issue:

1. 1. Search .learnings/ for a related keyword or Pattern-Key.
2. If a related item exists, link it with See Also.
3. Increase Recurrence-Count and refresh Last-Seen.
4. Escalate priority if the pattern is recurring and costly.

Recurring issues often mean one of three things:

• - documentation is missing
• automation is missing
• the architecture or workflow is inviting the same failure

When To Extract A Skill

Extract a reusable skill when the pattern is:

• - resolved and trustworthy
• useful across multiple tasks
• non-obvious enough to justify explicit guidance
• portable beyond a single private incident

Use the helper:

CODEBLOCK5

Then customize the generated SKILL.md and update the original learning entry with:

• - INLINECODE36
• INLINECODE37

Review Rhythm

Review .learnings/ at these checkpoints:

• - before major tasks
• after finishing a feature or bugfix
• when working in an area with previous failures
• during periodic maintenance

Useful checks:

CODEBLOCK6

OpenClaw Setup

OpenClaw works especially well with this skill because workspace files and hooks let the improvement loop stay visible between sessions.

Install

CODEBLOCK7

Manual install:

CODEBLOCK8

This package is an OpenClaw-oriented evolution of the earlier self-learning workflow.

Hook Setup

Optional bootstrap reminder:

CODEBLOCK9

Workspace Layout

CODEBLOCK10

Hook Support For Other Agents

Claude Code / Codex

Use hook scripts in settings:

CODEBLOCK11

GitHub Copilot

Add a reminder to .github/copilot-instructions.md:

CODEBLOCK12

Best Practices

1. 1. Log signal, not noise.
2. Prefer prevention rules over postmortems.
3. Link related incidents instead of duplicating them.
4. Promote broadly useful guidance quickly.
5. Treat repeated friction as a systems problem, not just a note-taking problem.
6. Review learnings before repeating the same class of work.