Find your next open-source project

Search by name, purpose, or keyword — or paste a GitHub repository URL

sybil-solutions/skill-distill

AI Engineering

Turn a strong Claude Code or Codex session into a reusable skill that an open-weight model can follow.

This repo is an agent skill pack · 1 skillsSkill list and install →
114
4
Python
MIT
new today

What it does

skill-distill is a Claude Code or Codex skill that turns a successful session into a written skill for a weaker open-weight student model. You name the student as a harness plus a model; the teacher extracts the session, writes a skill for the student, runs the student on a clean copy of the starting state with the original prompt, compares the output with the reference, fixes the first failure, and repeats. Each version is run three times, and the teacher presents the best student result alongside the reference for final human acceptance.

The output has two parts: the reusable operator skill that drives this loop, and a produced skill scoped to one task, one model and one harness. The teacher's loop is defined in SKILL.md.

Install

Requirements are tmux, git, Python 3.10+, and an installed student harness pointed at its model. For web tasks, also install playwright and pillow and the Chromium browser.

Clone the repository and link the skill into the relevant skills directory:

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

Quick start

At the end of a session that went well, invoke the skill with a concrete student, for example:

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 and begins the trial loop. You can watch the student run with tmux attach -t distill-<task>. The run ledger commands (init, trial, status, wait, evidence, verdict, ledger, report, accept) are implemented in skills/skill-distill/scripts/distill.py. Raw session logs are read with extract_session.py, and web results can be captured with capture_web.py.

Constraints

  • The student receives only the reference prompt, a clean starting state and the skill; the harness's global context files and skills are disabled where supported, so results depend on how faithfully the harness allows that.
  • The skill must teach steps, decisions, traps and checks, not embed the reference's finished output or a script that writes it.
  • Three trials per version are required; a single good run is not accepted. Checks must first pass on the reference.
  • Raw session logs remain in runs/ and are git-ignored because they contain local machine paths and context.
  • Harness support varies: omp is tested, Codex CLI is tested as teacher, Claude Code needs an Anthropic-compatible endpoint for open models, pi and opencode are written but untested, and opencode logs currently lack a parser in extract_session.py.
  • MIT license.

HotRepo's take

The repository was recorded on the new-project board on 2026-11-21 with 113 stars gained and 4 forks in its first observed window, ending at 113 total stars by the 12:15 UTC reading. That is a single-snapshot signal, not a trend, and no discussion threads were recorded. The project's practical value depends on the adapter quality and on the three-trial acceptance loop; the README's own status table marks most non-omp harnesses as written but untested, so the tested surface is currently narrow.

Written by AI from the project's public information and HotRepo's records, updated . The project's own docs are the source of truth.