10xjobs-careercoach — working rules

Read docs/PROCESS.md before building or changing any skill — the ritual (design gate → issue → prior-art check → build → independent review → live run → harness → closing review) is not optional and each step carries its receipt.

Precedence for every design judgment: PRINCIPLES.md → docs/design-cowork-coaching-goals.md → docs/design-cowork-coaching.md → docs/skill-shape.md. Conflicts mean the chain is wrong — fix it, don't pick a winner.

Always-on interrupts (full versions + receipts in PROCESS.md):

• Patch, grep-verify, and commit in ONE command block — a commit that claims an unverified change is a lie.
• Verify the file, not narration: grep before believing any "written/fixed".
• Never test or solo-review a skill you authored; spawn an independent reviewer against the precedence chain.
• New rules bind to their trigger moment, name the failure they prevent, and get measured (multi-trial). A measured miss earns a moment rule in the loop, a hint in patterns, or a script — never a step.
• A skill's SKILL.md is goal + loop shape (standard · budget · the measured moment rules · exits); the how lives in references/patterns.md. tests/test_invariants.py fails the moment a skill drifts from the shape.
• Skill prose is checked like everything else — check_files.py FAILs a relative link that doesn't resolve.
• Tests: python3 tests/run.py (pytest is not assumed). Conduct harnesses: tests/always-on/README.md has the run/judge commands; they run against fresh temp workspaces only, never a real one.
• A real candidate workspace is personal data: never plant test data in one; harness runs use fresh temp workspaces and lock any real workspace for the run.
• After skill changes, refresh the deployed copy: cp -r skills/* ~/.claude/skills/ — and know its trap: cp never deletes, so a file REMOVED from the repo must be removed there by hand.
• Session transcripts and candidate data never enter the repo.