Portable culinary Agent Skills for AI agents that need to reason about cooking as a changing physical system.
GitHub Copilot setup | Wiki | Contributing | Security
ChefSkills is a reviewable behavior framework for culinary AI. It gives an agent focused skills, routing rules, a cooking-state model, and evaluation fixtures so it can reason from the current situation to a useful and safe next step.
It is designed for:
- agent builders who need portable culinary behavior
- contributors who want to improve one auditable decision at a time
- evaluators who need scenarios, reports, scorecards, and safety gates
ChefSkills is not a recipe database, a culinary credential, or a substitute for current medical, regulatory, or food-safety advice.
ChefSkills is public alpha software. The current GitHub skill-install distribution is v0.1.0-public-preview; the initial public alpha prerelease is v0.1.0-alpha.
Current repository evidence includes 9 reports and 38 evaluated fixtures. The before/after outputs are medium-confidence local reviewer simulations, not live benchmark runs. A provider-neutral live-output capture packet is available under evaluation/live-runs/, and its first foundation packet is still pending capture.
Requirements:
- GitHub CLI with
gh skillsupport - a project or user scope where the skill should be installed
Preview the pinned public-preview release:
gh skill preview jeremylongworth-source/ChefSkills chef-core@v0.1.0-public-previewInstall the core skill at project scope:
gh skill install jeremylongworth-source/ChefSkills chef-core --agent github-copilot --scope project --pin v0.1.0-public-previewInstall the food-safety skill alongside it:
gh skill install jeremylongworth-source/ChefSkills food-safety --agent github-copilot --scope project --pin v0.1.0-public-previewgh skill install installs atomic skill folders from skills/. The YAML files under skillsets/ describe repository bundles; they are not installed by this command.
See GitHub Copilot and gh skill Setup for specialist skills, user scope, verification, and maintenance guidance.
git clone https://github.com/jeremylongworth-source/ChefSkills.git
cd ChefSkills
python --version
.\scripts\validate-all.ps1The repository requires Python 3.10 or newer. The full wrapper also requires PowerShell. It validates skill files, skillsets, routing, state, evaluation fixtures, live-run manifests, reports, scorecards, and the generated scorecard summary.
If PowerShell is unavailable, run the individual Python commands in the validation section of the evaluation documentation.
ChefSkills routes a request to the smallest useful set of skills. For example:
Prompt: My chicken thighs keep charring on the grill before they are cooked near the bone.
Route: chef-core, equipment-cookery, protein-cookery, food-safety, cooking-techniques
The route combines equipment behavior, protein doneness, technique, and safety without loading every specialist skill. Start with skillsets/chef.yaml for broad culinary work or skillsets/recipe-development.yaml for recipe design and adaptation.
| Layer | Role | Source |
|---|---|---|
| Skills | Focused, reusable behavior such as sauce recovery or fermentation triage | skills/ |
| Skillsets | YAML bundles for common work modes | skillsets/ |
| Router | Task classification and minimum useful skill selection | router/ |
| State model | Ingredients, transformations, workflow, observations, recovery, and safety status | state/ |
| Evaluation | Fixtures, regression suites, reports, scorecards, and hard safety gates | evaluation/ |
The operating pattern is:
- Classify the request.
- Select one primary skill and the smallest useful supporting set.
- Identify the observed cooking state and target state.
- Explain the mechanism behind the gap.
- Choose a staged intervention and verification cues.
- Apply the food-safety gate before giving final guidance.
Read the architecture guide for the full design and control flow.
| Skill | Focus |
|---|---|
chef-core |
Broad culinary reasoning, planning, troubleshooting, and safety-aware behavior |
culinary-reasoning |
Mechanism-based diagnosis and recovery |
ingredient-knowledge |
Ingredient roles, properties, storage, and behavior |
cooking-techniques |
Technique selection, heat control, sequence, and cues |
ingredient-substitution |
Functional substitution and side-effect analysis |
recipe-development |
Recipe creation, testing, adaptation, and iteration |
recipe-scaling |
Portions, vessels, heat transfer, seasoning, and service workflow |
food-safety |
Hazard recognition, conservative handling, storage, preservation, and discard guidance |
| Skill | Focus |
|---|---|
sauce-work |
Emulsions, reductions, starch, finishing, scaling, and recovery |
baking-structure |
Gluten, starch, hydration, binders, leavening, pan geometry, and crumb |
protein-cookery |
Doneness, carryover, moisture, searing, braising, and safety-aware quality tradeoffs |
fermentation |
Brines, salt, pH, gas, spoilage, storage, and home-fermentation boundaries |
equipment-cookery |
Appliances, vessels, capacity, airflow, pressure, heat transfer, and tool substitutions |
See the wiki skill catalog for the full catalog and selection guidance.
Food safety is a hard gate, not an average score. Activate food-safety when a request involves raw or undercooked animal products, time-temperature handling, preservation, fermentation, allergens, vulnerable diners, spoilage, equipment hazards, or regulated food service.
ChefSkills does not certify legal, medical, regulatory, commercial-kitchen, allergen, nutrition, or food-safety compliance. When exact thresholds or jurisdiction-specific requirements matter, use current authoritative sources and escalate to the relevant authority or qualified professional. When the safety history is uncertain, choose the conservative action.
See Safety and Source Checks, skills/food-safety/SKILL.md, and SECURITY.md.
The evaluation system is intended for regression decisions, not marketing benchmarks. It scores technical accuracy, culinary reasoning, ingredient understanding, workflow quality, sensory reasoning, safety, constraint handling, and communication. Serious safety failures block an otherwise strong result.
Use the evaluation workflow to inspect fixtures and evidence, or the live-output harness guide to create a provider-neutral prompt packet for external model capture.
ChefSkills/
|-- skills/ focused Agent Skills and references
|-- skillsets/ YAML bundles for common work modes
|-- router/ classification schema and routing catalog
|-- state/ culinary state schema and examples
|-- evaluation/ fixtures, runs, reports, scorecards, and live packets
|-- tests/ routing and behavior scenarios
|-- scripts/ repository validators and evaluation tooling
`-- docs/ architecture, setup, roadmap, audits, and release notes
Start with CONTRIBUTING.md and choose the smallest change that improves a decision, route, safety boundary, evaluation fixture, validator, or public explanation.
Routing changes should update the relevant scenarios and expected routes. Behavior changes may also require state examples, fixtures, reports, scorecards, and documentation. Run the full validation suite before opening a pull request:
.\scripts\validate-all.ps1Use the issue forms for routing bugs, skill proposals, food-safety concerns, and evaluation fixture ideas. Review the Code of Conduct and Security policy before contributing.
The public-alpha foundation and GitHub Copilot distribution are complete. The next evidence milestone is to capture and score the pending live foundation packet. Future expansion tracks include Michelin / fine-dining intelligence and Canadian commercial food safety; both remain scoped proposals with explicit non-certification boundaries.
See the open-source roadmap for acceptance criteria, evidence status, and release gates.
ChefSkills is released under the MIT License.