Pick gtm agent pattern

Pick GTM Agent Pattern

Run the gate sequence that picks an agent's build pattern and forces the "patterns NOT used" list, so over-engineering dies in planning.

The gates live in ONE place — read them, never reproduce from memory: pick_agent_pattern.md (in this skill's folder)

The framework the gates encode is explained in three reference files:

• reference/agent-build-patterns.html — the 21 control-flow patterns (chains, routers, ReAct, orchestrator-worker…) and the choosing rule: can you draw the flowchart before runtime?
• reference/the-dial.html — the dial: the build tiers (neurosymbolic 10-80-10 · guarded judgment ~5-90-5 · language-as-deliverable ≈5-60-35 · deterministic pipeline 0-100-0/0-100-5), the two-question ambiguity test, and the pyramid discipline (most agents are pipelines; only a handful reason).
• reference/gtm-fleet-switchboard.md — the 46-agent Model Switchboard fleet, each agent's trigger, deliverable, and tier. Use it to find the closest analogue to the agent you're planning and start from its tier. Source of truth: https://labs.bowtiefunnel.com/agent-models/

Steps

1. Read the canonical gates doc above in full (it evolves; memory of it is stale). The intake block and gates are inside it.
2. Fill the doc's intake block from the user's request and the codebase — including the two reuse lines (shared plumbing to reuse; competitor/parity evidence, i.e. a feature-by-feature table against the product being rebuilt). Fields the user didn't supply: estimate them and mark [assumed] — only stop to ask if the description is too vague to pass Gate 2 (the doc's 3-question rule). Include the filled intake in your output. If a parity pass hasn't been done for a competitor rebuild, recommend doing it BEFORE the gates so scope is settled.
3. Work Gates 1→5 IN ORDER, stopping at the first that holds. Do not skip a gate; do not escalate past a gate that holds. At Gate 3, look up the closest fleet analogue in reference/gtm-fleet-switchboard.md and start from its tier.
4. Produce the doc's 5-part output format (verdict line with the dial tier, NEURAL/SYMBOLIC/HUMAN flowchart, patterns NOT used, v0, top-3 failure modes → gates). Under a page.
5. If this feeds a build: fold the verdict into the agent's spec, and record the calibration result in the doc's Edge Cases section so the gates keep learning.

Red flags

• Reciting gates from memory instead of reading the doc → stale gates, missed rules.
• Producing a verdict without the "patterns NOT used" list → the list is the point.
• Orchestrator–worker chosen because "the job is big" → re-check Gate 2; big ≠ unknowable.
• Most verdicts coming out neurosymbolic → gates escalated too easily; a healthy fleet is a pyramid (see reference/the-dial.html).