Skip to content
Get packs

Retune a Skill Corpus

Measure and retune a skill set for a new AI model using benchmark runs before cutting anything

Source Download skill
Repository
EveryInc/compound-engineering-plugin
Stars
25,850
Licence
MIT
Files
8 files

What it does

This skill guides an AI agent through repairing a set of agent instructions (a corpus of skill files) when they start behaving badly on a newer model. Instead of rewriting prose based on guesses, it first mines past run logs for a baseline, measures the noise in the test harness, registers a pass bar in writing, then makes measured cuts and checks each one against the bar. It refuses to run without a harness that can compare two builds of the corpus.

Great for

  • Maintainers of large agent-skill collections who need to adapt them to a new model version
  • Teams that already have run archives and a repeatable benchmark task for their agent workflows
  • People who want evidence that each change to agent instructions actually helped

Skip it if

  • Projects with no benchmark harness: the skill stops by design, and an audit-only review is a different request
  • Quick one-off prompt edits, since the multi-phase process is heavy and time-consuming
  • Anyone expecting word reduction as the goal: the skill explicitly treats leanness as a separate program

Try asking

  • “Retune the ./skills corpus for a new model, with the target model and a bar of 3 consecutive clean runs”
  • “Mine our last month of run logs to build a baseline before changing any skill files”
  • “Audit the skills for regression classes, then cut stalling behaviour in measured passes”
  • “Establish the noise floor by running the harness on two identical copies of the corpus”
Step-by-step guide: how to use Retune a Skill CorpusInstall, first run, examples, tips and FAQs

Install in 3 steps

  1. 1

    Download the skill

    The ZIP contains the 8 original files, the MIT licence and an attribution note.

    ce-retune.zip
  2. 2

    Add it to your agent

    Claude Code (all projects)

    mkdir -p ~/.claude/skills && unzip ce-retune.zip -d ~/.claude/skills/

    Claude Code / Agent SDK (this project only)

    mkdir -p .claude/skills && unzip ce-retune.zip -d .claude/skills/

    Claude apps (web/desktop): Settings → Capabilities → Skills → Upload skill, and choose the ZIP.

  3. 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

  • Register the pass bar in writing before making any change, otherwise the result is not a valid test
  • Check the two-identical-copies noise floor first so you know which differences are real
  • Do not weaken tests to make a suite pass; report removed strings as findings instead

What's inside

  • agents/openai.yaml
  • references/baseline-mining.md
  • references/corpus-audit.md
  • references/cut-passes.md
  • references/halt-taxonomy.md
  • references/noise-floor.md
  • references/workflow-shapes.md
  • SKILL.md

Redistributed unmodified from EveryInc/compound-engineering-plugin 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.