Copy-paste prompts. Give the goal and how you'll know it's done; let the router pick the playbook. Invocation per host: Use the … skill (OpenCode), /… (Claude Code), $… (Codex).
Use the poteto-mode skill. sumOfSquares in src/math.ts returns the wrong result
for negative numbers. Repro first, then fix and verify. Keep the diff minimal.
What happened: matched the Bug fix playbook, reproduced before touching code (negatives failed, positives passed), delegated the fix to one scoped poteto-agent, then verified the real module directly — 4 cases green — and named the principles that shaped each decision (fix-root-causes, prove-it-works, laziness-protocol, guard-the-context-window). Final diff: one reducer body.
From upstream's examples (mined from Lauren Tan's sessions):
Use the how skill. How do we cancel runs? Do we have an n+1 when we look up
every run to cancel?
What to expect: the Investigation playbook — read-only, no edits. Explorers fan out over the subsystem (2–4 angles), a synthesizer merges findings, and the answer cites files, not impressions. Use it whenever you're tempted to ask "are we sure?".
Use the interrogate skill. Review this PR and try to break it.
What to expect: independent reviewers attack the diff (assumptions, edges, races, failure modes) and the lead aggregates to one severity-ranked verdict with file references. On hosts without per-role models you get independent passes instead of model diversity — the skill says so.