Why use it: The repo you need isn't in the library of 1,000 skills yet, or the existing skill has drifted from upstream.
What it produces: A self-contained skill graph (SKILL.md + sub-skills/ + references/ + scripts/) that has passed verification, along with a review artifact located outside the runtime tree.
0. Mode boundaries (read first, save yourself wasted turns)
| Mode | Which skills it sees | Task |
|---|---|---|
| Researcher (default) | operating + shared, including user skills not declared metadata.disco-role |
Use operational knowledge to do research work |
| Creator | metadata.disco-role: meta + shared |
Build and maintain skills |
Creator has 15 meta skills (countable: 15 directories declaring disco-role: meta in cli/packages/coding-agent/src/disco/skills/). The remaining two directories: repo-skills-router is operating, workflow-authoring is shared.
Open Creator:
disco --creator # tương tác
disco --creator -p "<yêu cầu>" # một phát rồi thoát
1. The standard entry point when you don't know which path to pick
disco --creator -p "/skill:distill-ml-knowledge identify <source or task anchor>; scope, ground, construct, and verify the operating skill graph."
distill-ml-knowledge determines the anchor then drives through 4 phases: scope → ground → construct → verify. If the source is a repository it usually reuses create-repo-skill; if the source is a paper it takes the paper path (see 06). Only when there is evidence of a recurring capability gap does it switch to design-meta-skill.
2. Create a repo skill
disco --creator -p "/skill:create-repo-skill Create and verify a repository skill for /absolute/path/to/repo."
Short form (Creator picks the entry itself):
disco --creator -p "Create a repo skill for /path/to/repo."
disco --creator -p "Create a skill for /path/to/repo using Python /path/to/env/bin/python"
Also delegate the choice of extraction scope and import into the library after verification passes — only add this when you really want to skip those two approval gates:
disco --creator -p "Create a repo skill for /path/to/repo with auto decide and auto import."
The workflow does: analyze the repo structure → prepare/check the Python inspection environment (via prepare-repo-skill-env) → write runtime guidance → record provenance → hand off to verify-repo-skill.
3. verify-repo-skill to guard what
- Generate usability cases with assertions;
- Self-refinement at the content level;
- Run native examples/tests safely if the repo has them;
- Static quality gate;
- Record coverage + review artifacts.
Only after passing all of these is it considered "ready". The review artifact is not located in the skill's runtime tree.
4. Refresh — when upstream has changed
disco --creator -p "/skill:refresh-repo-skill Refresh /absolute/path/to/existing-skill against the current repository at /absolute/path/to/repo."
Short form: disco --creator -p "Refresh the skill at /path/to/repo/skills/example-skill against the current /path/to/repo code."
Refresh keeps guidance that is still correct, and only compares the outdated parts against the current source baseline.
5. Extend — when the skill is still correct but missing areas
disco --creator -p "/skill:extend-repo-skill Add streaming inference coverage to /absolute/path/to/existing-skill using /absolute/path/to/repo as evidence."
To distinguish clearly: refresh = content has drifted from upstream; extend = correct content but does not yet cover the new workflow.
6. After verifying: where to deploy
Creator proposes one scope before importing after you approve:
- Tied to a specific task/project/dataset/evaluator/machine, or reuse is uncertain →
<project-dir>/.agents/skills/(only loaded when the project is trusted). - Self-contained, has provenance, cross-project verified → then propose
~/.disco/agent/skills/. - One graph does not split in half between two scopes.
The skill repo goes through the specialized import path ~/.disco/agent/skills/repositories/repo-skills/ along with rebuilding the sibling router — do not push routing metadata through the general graph importer.
Finally, the Creator writes a handoff for the new Researcher session.
7. Catalog of 15 meta skills (to know what runs automatically, what must be invoked)
Manual invocation (entry point): distill-ml-knowledge, create-repo-skill, refresh-repo-skill, extend-repo-skill, import-repo-skills-to-agent, create-paper-skills, design-meta-skill.
Loaded automatically by the workflow — do not invoke manually per checklist: prepare-repo-skill-env, verify-repo-skill, paper-skills-distiller, plan-paper-skill-modules, create-paper-module-skill, prepare-paper-recovery-env, recover-paper-result, analyze-paper-recovery.
Next step
06-creator-paper-to-skills.md — the paper → skill path requires a TOML config file.