fires when
Use whenever the plan is to spawn, fan out, or parallelize agents \u2014 ANY multi-agent work, not just when models are mentioned. Triggers: \"spawn/kick off subagents\", \"a bunch of agents in parallel\", fleets that audit/comb/sweep/hunt across files, verify or skeptic stages, reviewer fan-outs, loop-until-done runs, authoring or editing a Workflow script (agent()/parallel()/pipeline()), or writing agent-definition frontmatter. Every spawned agent must get an EXPLICIT model tier assigned BEFORE it launches: reserve expensive head models for a small bounded set of final judges/synthesizers; mid-tier for judgment (review, verify, hunt); cheap tier for mechanical sweeps (grep, list, count, rename). Activate even when the user says nothing about models or cost \u2014 an omitted model silently inherits the pricey session model, and fan-out multiplies that waste by the agent count.
best for
Anyone about to spawn subagents, write a Workflow script with agent()/parallel()/pipeline(), or run an eval/optimization loop, who wants to avoid silently inheriting the expensive session model across every spawned instance.
measured · trigger
best score
14/18
description-triggering measured 2026-09-04
measured · quality
Not measured yet. A quality number needs its own eval: 15 or more tasks, each run with the skill and without it.
security scan · 8/8
✓prompt-injectionProse is all reader-facing methodology (tier table, rules); no directives aimed at overriding the reading agent's instructions.
✓data-exfiltrationNo network sends of repo/user data; only local progress-beacon file writes described.
✓secret-detectionNo keys/tokens/credentials present anywhere in the file.
–dangerous-commandsSkill contains no scripts/ directory and runs no shell commands.
✓obfuscationPlain markdown tables and prose; no encoded blobs or eval-of-decoded-data.
✓external-fetchesNo URLs or download/fetch instructions anywhere in the skill.
✓credential-accessNo reference to ~/.ssh, ~/.aws, .env, keychain, or any credential store.
✓privilege-escalationNo sudo/setuid/cron/profile persistence; only a benign ~/.claude/progress/current.json status file mentioned.
scanned 2026-09-08 · cerberus:security-scan
side effects & deps
read-only
needsnothing
what’s inside
SKILL.md7.2 KB
metadata.yaml782 B
evals/result.json6.8 KB
evals/rubric.json3.2 KB