cook — independently scanned and version-tracked by SaferSkills.
SaferSkills independently audited cook (Agent Skill) and scored it 100/100 (green). The audit ran 55 deterministic rules across Security, Supply Chain, Maintenance, Transparency, and Community; it found 0 high-severity and 0 lower-severity findings. The full rule-by-rule trace and per-finding evidence are below. Free, methodology-open.
Findings & checks · 0 flagged
Every scanned point with the score it earned and what moved between them.
First recorded scan — no prior version to compare against.
The primary manifest — the file an agent reads to learn what this artifact does.
ultrathink.
Cook is an incremental loop: break work into tasks, implement one, verify it works, move to the next. The key discipline is that nothing is "done" until there's evidence it works — but how you verify adapts to the situation.
If the request is clear, start. If it's ambiguous or multi-faceted, ask clarifying questions. If the work is large or multi-path, plan first (/give-plan). Otherwise, just start.
For a small single-concern request, just implement, verify, and continue.
For multi-step work or collaboration, use .claude/scripts/tasks.py to track concrete outcomes. Task files persist in .tasks/, so a fresh context can pick up where the last one left off.
Every task requires three fields: title, desc, expected. Read the authoring rubric before writing tasks — bad tasks are worse than no tasks:
.claude/scripts/tasks.py --help for the short-form rubric and examples.claude/scripts/tasks-authoring.md for the full guide.claude/scripts/tasks.py --task-file <slug> add "<title>" "<desc>" "<expected>"
.claude/scripts/tasks.py --task-file <slug> list
.claude/scripts/tasks.py --task-file <slug> verify <id> "<evidence>"
.claude/scripts/tasks.py --task-file <slug> done <id>Pick the next unblocked task, make the change, then hand off to a tester and (for risky work) a reviewer — isolated teammates that judge the change independently. See .claude/skills/test/teammate.md and .claude/skills/review-code/teammate.md for how to spawn them.
/cook owns implementation. The tester owns verification. Give the tester:
You can still add durable tests, fixtures, or stable selectors when they belong to the product change itself. Do not stuff temporary verification tactics into /cook just to get through one run.
Before invoking verification, confirm the edits actually landed on disk. If the change you expect is missing from the diff, fix that first; a passing check against unchanged code is worthless evidence.
Do not mark a task done on confidence alone. The tester proves the behavior. The reviewer checks that the implementation is correct, scoped, and aligned with the repo. For risky or non-trivial work, spawn a reviewer before marking the task complete.
Update task status as you go so the execution trail stays trustworthy.
After the task list is complete, review the combined diff before declaring success. Cross-task issues often appear only in the final aggregate: mismatched assumptions, naming drift, incomplete ripple updates, or verification that was too narrow. Spawn a reviewer for this final pass.
When done, summarize:
<request>$ARGUMENTS</request>
~30 seconds. Free. No account. Every finding cites a rule and a line of evidence.