playbook-first-name-cleaning

Run each row's raw first name, last name, and company through the locked gpt-4o-mini prompt, which returns a greeting-ready first_name_clean value in JSON. Then apply the six deterministic guards to abstain on placeholders and flag ambiguous rows for review before sending.

The playbook was benchmarked on gpt-4o-mini with JSON response format and max_completion_tokens of 200, scoring 96/100 on an adversarial 100-row benchmark at about $0.21 per 1,000 rows. A nano-class reasoning model needs a 2000-token cap or it returns empty content on every row.

No. A measured head-to-head showed a 435-token short prompt saved only $0.0145 per 1,000 rows but produced 3 broken greetings in 30 rows that none of the guards caught. Cutting the example block also drops the prompt under the 1,024-token cache floor, raising cost.

That happens when the guard's normalize function, which strips everything outside [0-9a-z], is used as an emptiness test — it reduces non-Latin names to empty strings. The fix is a Unicode-aware has_letters check alongside normalize, which affected 4 of 100 benchmark rows.

Unusable rows return an empty first_name_clean string and are excluded from the campaign and routed to review. The playbook never falls back to generic greetings like 'there' or 'friend', and never outputs 'N/A', which would render into a live email.

Yes, clay-table.md and clay-workflow.md specify a Clay AI column plus a JavaScript formula version of the six guards. However, only the script path is verified at 96/100; the Clay column and JavaScript guards are marked as unexecuted specifications requiring a smoke test.