Guide · Agent skill
How to use the Retune a Skill Corpus agent skill (ce-retune)
Updated · written with AI from the project's own documentation and checked against our catalogue data
The ce-retune skill guides an AI agent through repairing a set of agent instructions (a corpus of skill files) that start behaving badly on a newer model. It is built for people who maintain large skill collections and already have run logs and a repeatable benchmark task. After this guide you will know what the skill needs, how to install it, how to trigger it, and what it will refuse to do.
Retune a Skill CorpusEveryInc/compound-engineering-plugin · MIT Download the skill Quick start
- 1Confirm your harness. Check that you have run logs with tool-call traces, a build selector and a repeatable end-to-end task.
- 2Install the skill. Unzip the downloaded ZIP into ~/.claude/skills/ or a project's .claude/skills/, or upload it under Settings → Capabilities → Skills in the Claude apps.
- 3Ask for a retune with clear inputs. Name the target model or symptom, the corpus path and the bar in consecutive clean runs.
- 4Let the archive be mined first. The agent builds a baseline from existing run logs before any paid runs.
- 5Measure the noise floor and register the bar. Run the harness on two identical copies of the corpus, then record the pass bar in writing before any edits.
- 6Audit and cut in measured passes. Run separate adversarial audit agents, then apply one problem class per pass and measure each against the bar.
What the ce-retune skill makes the agent do
Instead of rewriting prose based on guesses, the skill works measurement-first. It mines past run logs for a baseline, measures the noise in your test harness, registers a pass bar in writing before any change, audits the corpus adversarially, and then makes cuts in measured passes, checking each one against the bar until the bar clears.
The stated outcome is a corpus whose measured behavior on the target model clears the pre-registered bar, with regression classes removed and each removal attributable. Word reduction is explicitly not the goal; leanness is treated as a separate program.
- Phase 0: confirm a run archive or harness, a build selector, and a repeatable task exist.
- Phase 1: mine the archive for a baseline before spending any runs.
- Phase 2: run the harness against two identical copies of the corpus to find the noise floor, then register the bar.
- Phase 3: audit adversarially, with one agent proposing cuts and a separate agent defending the existing prose.
- Phase 4 and 5: apply one problem class per pass and measure until the bar clears.
- Phase 6: ship with commits and a write-up.
Install the ce-retune skill
Download the ZIP for this skill from this page and unzip it into one of two places. For a personal install, use ~/.claude/skills/. For a single project, use the .claude/skills/ folder inside that project.
If you use the Claude apps instead, open Settings, go to Capabilities, and upload the skill under Skills.
- Personal: unzip into ~/.claude/skills/
- Project: unzip into .claude/skills/ inside your project
- Claude apps: Settings → Capabilities → Skills, then upload
Trigger the skill with a natural request
This skill is set to disable model invocation, so describe the job clearly and name the skill so the agent picks it up. The skill's argument hint lists three optional inputs: the target model or symptom, the path to the corpus (defaults to ./skills), and the bar as a number of consecutive clean runs.
Good requests name the new model, the symptom, and where your harness and corpus live. The agent will ask you to confirm the harness it finds before continuing.
- Use ce-retune to adapt our skills folder to the new model; the symptom is runs that stop early and ask for confirmation.
- Use ce-retune on ./skills for the new model. Our harness writes per-run JSONL logs in the runs directory.
- Run ce-retune with a bar of 5 consecutive clean runs on the corpus at ./skills.
Check what you need before starting
The skill will not run without a way to observe behavior. Before you start, make sure you have a benchmark harness that produces per-run logs with tool-call traces, terminal markers, token counts and final messages. You also need a way to point runs at a specific source checkout of the corpus, such as a plugin-dir override or an env var, plus a repeatable end-to-end task the corpus executes.
The audit phase also needs the host to run independent agent contexts. If it cannot, the skill reports that as a blocker and stops the audit rather than arguing both sides in one context.
- Run logs with tool-call traces, terminal markers, token counts and final messages
- A build selector to run two corpus builds under one runner
- A repeatable end-to-end task
- Independent agent contexts for the adversarial audit passes
Understand the safety notes
The skill does not run scripts and does not use the network, according to its listing. It works by reading files and dispatching agents, and it edits the corpus during the cut phases.
Two rules bound every pass. The skill never edits a test to make a suite green; a removed string that a test pins is reported as a finding. Also, not every stop in a workflow is a problem: some workflows exist to stop and ask, and the skill checks who is on the other side before cutting any stop.
- Runs scripts: no
- Uses network: no
- Edits to corpus files happen in Phase 4; the audit phase edits nothing.
- Commit history and the write-up are how changes stay traceable.
Tips for getting useful results
Expect the work to be slow. The multi-phase process is heavy by design, so it is a poor fit for a quick prompt tweak. A green test suite does not count as done, because it shows nothing broke, not that behavior improved.
Do not treat a large number of audit findings as a large effect. Only the harness can say whether a cut helped. Ask the agent to report what stayed unmeasured, since a cleared bar does not imply coverage it lacks.
- Set the bar before any change; a bar chosen after seeing results is not a bar.
- Keep one problem class per pass.
- Read the halt taxonomy before cutting any stop.
What people use it for
- Adapting a large collection of skill files after moving an agent workflow to a newer model.
- Diagnosing agents that stop early or halt and ask when they should keep going.
- Showing a team with run archives that each instruction change measurably helped.
- Checking which parts of a corpus were never exercised by the benchmark task.
Frequently asked questions
How do I install the ce-retune skill?
Download the ZIP from this page and unzip it into ~/.claude/skills/ for a personal install, or into .claude/skills/ inside a project. In the Claude apps, upload it under Settings → Capabilities → Skills.
What does the ce-retune skill require?
It requires a benchmark harness that produces per-run logs with tool-call traces, terminal markers, token counts and final messages. It also needs a way to point runs at a specific corpus checkout, a repeatable end-to-end task, and independent agent contexts for the audit. Without a harness, the skill stops by design.
Is ce-retune free to use?
The skill is listed under the MIT licence. Check the licence file in the source repository for the full terms. This page does not state any price.
Does ce-retune run scripts or use the internet?
According to this listing, it runs no scripts and uses no network access.
Can I use ce-retune without a benchmark harness?
No. The skill refuses to run without one that can A/B two builds of the corpus. If you only want a static review of your skill files, that is a different request and the skill does not cover it.
How is this different from a quick prompt edit or a word-count reduction?
The multi-phase process is heavy and slow, so it is meant for large collections adapted to a new model. The skill treats word reduction as a separate program and does not aim for shorter prose as the result.
New agent skills, repos and AI tools — every week
The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.
Our verdict
Use ce-retune if you maintain a large set of agent skill files, have run archives and a repeatable benchmark task, and need evidence that each change helped on a new model. Pick something else if you have no benchmark harness, want a quick prompt edit, or are aiming for shorter instructions.
Download the skill