## Problem L6 chain step 10 (`tests/dogfood/l5-rows/10-consultant/prompt.txt`) currently reads: ``` @bro let cto weigh in on monolith vs microservices for auth. Don't ask questions. ``` The prompt names `cto` explicitly. Per Human review (2026-05-17 on !2904): real users describe *intent*, not which agent/skill to use. Same anti-pattern as the step-14 prompt that was rewritten in #2904 Phase B. The scenario under test is bro's **consultant-spawn judgment** (which consultant — architect/cto/pm/ceo — is the right read for this question). By naming `cto` in the prompt, the test bypasses that judgment and checks only the mechanical spawn. The hook + spawn would fire even if bro's classification logic is broken. ## Evidence (verified) - `tests/dogfood/l5-rows/10-consultant/prompt.txt` — the literal prompt above - Same anti-pattern as step-14 (#2904 Phase B commit 967d448) — explicit-naming shortcut breaks the scenario's actual judgment test - Step 14's fix pattern: rewrite to a natural ask; trust `tmb_agent-creator` skill's description-match heuristic ## Plan 1. **Rewrite the prompt** to a natural user ask that does NOT name a consultant. The ask should clearly land in CTO territory (tech strategy / scaling / tech-stack trade-offs) so bro's consultant-classification logic picks `cto` correctly. Example: ``` @bro I need a tech-strategy read on monolith vs microservices for our auth layer — what scales better at our size? Don't ask questions. ``` This shape matches the `cto` template description ("Technical strategy + tech-stack trade-offs. Scaling, dependency posture, build/CI direction") and lets bro's `tmb_agent-creator` skill auto-classify the right consultant. 2. **Update the README.md** in the row dir to reflect the new prompt + design ("no explicit consultant naming; bro classifies"). 3. **Verify the scorers still pass with the new prompt**: - `tools-required.json` should require `Task` (spawn) tool - `outcome.sql` should assert ≥1 row in `agent_runs` with `agent_type='cto'` - If existing scorers require something the old explicit-naming guaranteed, update them to match the new realistic flow ## Acceptance criteria - New prompt does not match `\b(architect|cto|ceo|pm)\b` (no role naming) - New prompt naturally fires `tmb_agent-creator` and bro classifies as `cto` consultant - L6 chain step 10 passes (the row was passing before via explicit naming; should still pass with realistic flow) - README clarifies the realistic-prompt design (cite #2904 Phase B step-14 as precedent) ## Out of scope - Reviewing every other L6 prompt for similar anti-patterns (other 13 are clean per audit on !2904 Phase B notes) - Changing the consultant-classification logic in `tmb_agent-creator` skill ## Coordination - Same fix shape as #2904 Phase B step-14 commit 967d448 - Doctrine: real users describe intent, not agent/skill names ## Note on source Discovered during the L6 prompt audit (post-Phase-B review of !2904) when scanning all 14 chain prompts for the same anti-pattern as step-14. Step 10 is the only other prompt that names a TMB agent/skill explicitly.