Self-Improvement Loops
Guides an agent in designing loops that safely let its own harness, prompts, or context evolve.
- Stars
- 18,282
- Licence
- MIT
- Files
- 2 files
What it does
When installed, the agent gets a detailed design playbook for systems where the agent's own scaffolding (its harness, workflow, prompts, or context-generating mechanism) is the thing being automatically improved. It explains how to pick the right level to fix a failure, how to keep the scoring signal out of the optimizer's reach so it cannot game it, and how to accept self-edits only when measured results show no regressions. It is advice and structure for designing such loops, not a tool that runs one by itself.
Great for
- Engineers building recursive self-improvement or meta-harness experiments for AI agents
- Teams designing acceptance gates and held-out evaluation for self-modifying agent systems
- Researchers diagnosing reward hacking, diversity collapse, or silent stagnation in agent optimization loops
Skip it if
- Simply running a single autonomous agent with logging and approval steps; the skill says that belongs to harness-engineering, so it adds little here (it routes you elsewhere)
- Non-technical users looking for a ready-made product; it is a design reference for building optimization loops, and misapplying it can produce loops that quietly game their own metrics
- Projects that need only one-off prompt or token-efficiency tuning without any improvement loop
Try asking
- “Help me design a loop where my coding agent mines its failed test runs and proposes bounded edits to its own prompts.”
- “Review my acceptance gate for a self-modifying agent: should I require both held-in and held-out splits to improve?”
- “Diagnose why my self-improving harness keeps scoring higher while output quality is falling; check for reward hacking.”
- “Which rung of the optimization ladder should I fix this recurring failure at: prompt, context, workflow, or harness code?”
Install in 3 steps
- 1
Download the skill
The ZIP contains the 2 original files, the MIT licence and an attribution note.
self-improvement-loops.zip - 2
Add it to your agent
Claude Code (all projects)
mkdir -p ~/.claude/skills && unzip self-improvement-loops.zip -d ~/.claude/skills/Claude Code / Agent SDK (this project only)
mkdir -p .claude/skills && unzip self-improvement-loops.zip -d .claude/skills/Claude apps (web/desktop): Settings → Capabilities → Skills → Upload skill, and choose the ZIP.
- 3
Ask for the task
No special command needed — the agent reads the skill's description and loads it when your request matches. Try one of the prompts above.
Get better results
- Pair this skill with harness-engineering for governance and evaluation for building the scorer, since it deliberately defers those topics.
- Keep evaluator source hidden from any proposer agent, and enforce budgets and permissions in the runtime rather than in prompts.
- Start with a capability validation over a fixed number of iterations on held-out tasks before enabling recursive self-edits.
What's inside
- references/loop-design-evidence.md
- SKILL.md
Redistributed unmodified from muratcankoylan/Agent-Skills-for-Context-Engineering under the MIT licence. The assessment was drafted with AI from the skill's files on 2026-10-29. Skills can instruct your agent to run code — read SKILL.md and any scripts before installing.