A frontier model did a task well. Can an open-weight model do the same task, given the right instructions?
skill-distill is a skill for Claude Code or Codex (the teacher). You call it at the end of a good
session and name a student: an agent harness plus a model, for example omp with DeepSeek-V4.1-Flash.
The teacher then loops:
- reads the session and writes a skill for the student;
- starts the student in tmux, on a fresh copy of the starting state, with the original prompt and the skill;
- compares what the student produced with what the reference produced;
- finds the first place the student went wrong and fixes the skill there;
- runs again, three trials per version, until the student matches the reference.
You make the final call: the teacher shows you the reference and the student's best results side by side, and the skill counts as done only when you accept it.
The result is two things: the reusable operator skill in this repo, and a produced skill for one task, one model and one harness.
git clone https://github.com/sybil-solutions/skill-distill
ln -s "$PWD/skill-distill/skills/skill-distill" ~/.claude/skills/skill-distill # Claude Code
ln -s "$PWD/skill-distill/skills/skill-distill" ~/.agents/skills/skill-distill # Codex
Requirements: tmux, git, Python 3.10+, and the student harness installed and pointed at its model.
For web results, pip install playwright pillow && playwright install chromium.
At the end of a session that went well:
distill this session into a skill for omp with deepseek-v4.1-flash on my local server
The teacher follows skills/skill-distill/SKILL.md. You can watch the
students work with tmux attach -t distill-<task>.
| path | what |
|---|---|
skills/skill-distill/SKILL.md |
the loop the teacher follows |
skills/skill-distill/scripts/distill.py |
run bookkeeping: init, trial, status, wait, evidence, verdict, ledger, report, accept |
skills/skill-distill/scripts/extract_session.py |
Claude Code / Codex / omp session log → readable reference (drops system and AGENTS.md text) |
skills/skill-distill/scripts/capture_web.py |
serve a web app, drive it in headless Chromium, save screenshots and errors |
skills/skill-distill/adapters/ |
one script per student harness: omp, pi, Codex, Claude Code, opencode |
skills/skill-distill/templates/student-skill/ |
the starting shape of a produced skill |
examples/ |
finished runs: the prompt, the produced skill, the ledger, screenshots |
- The student gets the reference's prompt verbatim, a clean starting state, and the skill. Nothing else: the student harness's own global context files and skills are turned off where the harness allows it.
- The skill may not contain the reference's finished output or a script that writes it. It teaches the steps, decisions, traps and checks; short snippets only where the student can't get a part right otherwise.
- Three trials per skill version. One lucky run proves nothing.
- Checks are tested on the reference first. A check that fails on the reference is a broken check.
- Raw session logs stay in
runs/, which is git-ignored: they contain your machine's paths and context.
| harness | status |
|---|---|
| omp (oh-my-pi) | tested |
| Codex CLI | tested as the teacher; as a student needs a provider in STUDENT_ARGS |
| Claude Code | written; open models need ANTHROPIC_BASE_URL to an Anthropic-compatible server |
| pi | written, untested |
| opencode | written, untested; its JSON log has no parser in extract_session.py yet |
MIT