Principles and recipes

Use principles when they change a decision

The router's principle index lists all 24 leaves; the skill catalog describes each one. Read the relevant leaf instead of reciting every principle for each task.

Principle	Pushes toward
separate-before-serializing-shared-state	Owned outputs
fix-root-causes	Tracing the actual mechanism
prove-it-works	Real evidence

WARNING

No principle expands scope, overrides a read-only request, or authorizes publication.

Write a useful prompt

State the goal, a pass/fail done check, the proof you want to inspect, known facts such as repro steps or logs, and real constraints such as read-only investigation or a review checkpoint. Leave room for the agent to choose the implementation unless the method itself is a requirement. For a noisy report, ask for a plain-language restatement before editing; share a theory of the cause after that first reading to avoid anchoring the search.

Save z-mode for work that needs rigor. Use verified native model and effort settings when requesting a smaller budget, and bound review panels to useful independent coverage under the native host contract. Installed skills do not create a separate model-routing configuration.

For help rewriting an ask, explicitly invoke sharpen. It adds expert methods, original sources, and an independent check while leaving your choices open. It returns the prompt without doing the task; review and send the rewrite separately.

Recipes

Claude Code

/how trace argument parsing; read-only
/in-your-own-words <long or messy ask>
/sharpen <the prompt you are about to send>
/z-mode read this report; restate the underlying issue and investigate the evidence; don't change any code yet
/interrogate this diff; findings only
/swarm check these three packages; one owned report per package
/z-mode fix duplicate output; reproduce first; keep changes local
/z-mode pause; record the exact resume action

Codex

$how trace argument parsing; read-only
$in-your-own-words <long or messy ask>
$sharpen <the prompt you are about to send>
$z-mode read this report; restate the underlying issue and investigate the evidence; don't change any code yet
$interrogate this diff; findings only
$swarm check these three packages; one owned report per package
$z-mode fix duplicate output; reproduce first; keep changes local
$z-mode pause; record the exact resume action

Pitfalls

• Overlapping writers.
• Stale head receipts.
• Unscoped history mining.
• Assuming a started agent completed.
• Leading with a theory before the issue is understood.
• Accepting a design while experimentally answerable questions remain.
• Polishing an abstract plan instead of testing its uncertain assumptions.
• Repeating unchecked work before the workflow earns trust.
• Reporting a speedup without vetting the measurement.
• Correcting the same mistake in chat instead of preventing it in the repository.

Scripts use absolute paths from the real installed skill; Git helpers inspect the target project. Retired automation sources are documented in provenance.