| name | final-answer-completeness |
| description | Post-draft check that every explicit question and sub-request in the task is answered in the response, in the form the task asked for — nothing dropped, deflected, or answered in the wrong shape. |
| license | MIT |
final-answer-completeness
Trigger (observable): The task contains more than one question or directive (multiple question marks, numbered questions, embedded format instructions), and a draft response exists and is about to be sent.
When NOT to activate: Single-question tasks whose draft visibly answers it; mid-drafting (run at send time, not while composing); the user rescoped the questions mid-session (latest instruction wins).
Procedure
- Extract every interrogative and directive from the task verbatim — including question two inside a paragraph, 'also tell me...', and shape directives ('as a table', 'in 3 lines', 'yes or no first').
- Map each to the exact location in the draft that answers it.
- Unmapped items: answer them now, or state explicitly that they are unanswered and why — a response that silently answers 2 of 3 questions reads as evasion.
- Check shape compliance: a 'yes or no?' answered with an essay, or '3 bullets' answered with 5 paragraphs, fails even when the content is right.
- Exact-format rule: when the task specifies output lines "exactly" / "of the form" or says they are parsed by a script or machine, emit those lines VERBATIM at column 0 in plain text — no bold, italics, indentation, bullets, or code-fence prefix on the line itself — and place NOTHING after a line the task designates as last.
- Check the first two sentences carry the primary answer, not preamble.
Required output
The corrected draft, plus (for 3+ question tasks) a one-line-per-question mapping showing where each is answered.
Verification
- Every question extractable from the task maps to a specific location in the sent response, or carries an explicit unanswered-because statement.
- Each shape directive (count, format, yes/no-first) is honored or explicitly declined with a reason.
- When the task mandates exact or machine-parsed output lines, those lines appear verbatim at column 0 in plain text — no bold, italics, indentation, bullet, or code-fence prefix — and nothing follows a task-designated last line.
Known risk: Mapping tables appended to short answers, bloating a two-question reply. Mitigation: below 3 questions the check is silent — fix the draft, skip the table.
Max intended cost: ≤150 added output tokens beyond the answers themselves; one re-read of the task.
Evidence status: EXPERIMENTALLY_TESTED — this repair content lifted a weak model on procedural tasks in controlled runs; the effect is task-concentrated, and automatic router selection of micro-skills is a separate, unproven layer.
Lineage: Grounded in an evidence-backed completion-integrity principle (partition the request into its parts and close each one) and in a documented reasoning-failure mode — confidence in a response outrunning the evidence that it actually covered the request. Sibling of an artifact-completeness check; this one covers the response text itself.
