🧩 Skills
Self-Improving Agent

Self-Improving Agent

Self-Improving Agent is a skill that critiques and rewrites other skills. Give it a skill file and a goal — it runs the skill, evaluates the output against the goal, rewrites the instructions to perform better, and iterates until the output meets your criteria.

Install:

npx clawhub@latest install self-improving-agent

clawhub.ai/pskoett/self-improving-agent (opens in a new tab)

What it does

  1. Reads the target skill file
  2. Runs the skill against a test input you provide
  3. Evaluates the output against your stated goal
  4. Identifies the weakest parts of the skill instructions
  5. Rewrites those sections to better achieve the goal
  6. Re-runs and re-evaluates
  7. Repeats until the output meets your criteria or a maximum iteration count is hit

The result is an improved SKILL.md saved back to your ~/skills/ folder.

Basic usage

Use the self-improving-agent skill to improve my morning-brief skill.
Goal: the brief should feel like a human wrote it, not an AI summary. The weather section should be conversational.
Test input: today is Monday, May 25, 2026. Use dummy data for everything.
Max iterations: 3.
Use the self-improving-agent skill on my humanizer skill.
Goal: it should catch more AI patterns in short paragraphs (under 100 words).
Test input: [paste a short AI-sounding paragraph]

What to specify

For best results, give the self-improving agent three things:

  • Goal — what "good" looks like. Be specific. "More human" is vague. "Vary sentence length, cut AI vocabulary, and add at least one concrete example per section" is actionable.
  • Test input — what to run the skill against. Use real or representative data.
  • Max iterations — usually 2–3 is enough. More than 5 is rarely worth the token cost.

How it pairs with other skills

Self-Improving Agent works on any skill file. Useful combinations:

  • Improve your morning-brief skill — tighter structure, better section ordering, more useful data
  • Improve your humanizer skill — better pattern detection, more targeted rewrites
  • Improve a research skill — sharper search queries, better synthesis instructions
  • Improve a code-review skill — more specific criteria, cleaner output format

Self-Improving Agent modifies your skill files. Back up any skill before running it: cp ~/skills/morning-brief.md ~/skills/morning-brief.backup.md. If the result is worse, restore from backup and try a more specific goal.

The broader pattern

Self-Improving Agent is an implementation of the autoresearch loop — a general pattern for autonomous improvement of any artifact with a measurable goal. You can apply the same principle manually: run a skill, evaluate the output against a rubric, edit the instructions, repeat.

The skill just automates that loop.

HumanizerOpenAI Whisper