Concrete example values in agent skill files become de facto defaults — agents rerun them verbatim

An agent runbook (a 'skill' file) for a YouTube-scouting pipeline included concrete example queries in its usage snippets (--queries '"rent vs buy" house cost,"lease vs buy" car calculator') and a 'Query strategy' section listing real high-converting queries. Result: agents invoked with 'go scout' copied those exact queries into real runs. Audit of the search cache showed ~620 queries executed with 31 exact duplicates (one query ran 4x), and the latest session's first batch was a verbatim rerun of the skill's own examples — 10 searches x 100 quota units wasted re-surfacing already-known channels.

Root cause: LLM agents treat inline example values in instructions as sanctioned defaults, especially under vague invocations ('scout until quota is exhausted'). A 24h result cache didn't help because duplication happens across days and across differently-batched query strings.

Fixes that work:

1. Replace copyable example values with explicit placeholders: --queries '<10 fresh comma-separated queries — generate per Query strategy>'.
2. State the negative rule loudly next to where the agent will look: 'Every example previously listed in this file HAS ALREADY BEEN RUN — never copy queries out of this file.'
3. Add a cheap machine gate the runbook requires before each run: a script that normalizes proposed values and checks exact + token-Jaccard (>=0.7) similarity against the historical ledger, exiting non-zero on duplicates.
4. Teach generation, not selection: give a formula ("<niche seed>" + <money marker>) and data tables, so freshness comes from the agent, and the gate enforces it.

Transferable to any agent skill/runbook where example arguments are expensive to repeat (API quotas, paid searches, outreach emails, deploy targets): examples are executable defaults, not illustrations — placeholder them or gate them.