Skills / pick-gtm-agent-pattern

/pick-gtm-agent-pattern

Most agents are over-built — an orchestrator-worker mesh doing a job a plain pipeline would do better.

pick-gtm-agent-pattern runs the gate sequence that picks an agent's build pattern and forces the "patterns NOT used" list — so over-engineering dies in planning, before a line of code is written.

MIT Claude Code skill planning-time anti-over-engineering

What it does

Point it at a new agent idea, a workflow you want to turn into an agent, or an existing agent you suspect is over-built. It fills an intake block from your request and the codebase, then works a fixed gate sequence 0→5, stopping at the first gate that holds — no skipping, no escalating past a gate that already answered.

The output is a one-page verdict: the dial tier, a NEURAL / SYMBOLIC / HUMAN flowchart, the patterns explicitly NOT used (the list is the point), a v0 scope, and the top-3 failure modes mapped back to gates.

How it works

  1. Read the canonical gates doc in full. It evolves; memory of it is stale. The intake block and gates live inside it.
  2. Fill the intake block from your request + the codebase — including shared plumbing to reuse and competitor/parity evidence. Missing fields get estimated and marked [assumed].
  3. Work Gates 0→5 in order, stopping at the first that holds. Never skip a gate; never escalate past one that already holds.
  4. Produce the 5-part verdict — tier, flowchart, patterns NOT used, v0, failure modes → gates. Under a page.
  5. Feed it into the build — fold the verdict into the agent's spec, and record the calibration back into the doc so the gates keep learning.

The two frameworks it encodes

reference

The 21 control-flow patterns

Chains, routers, ReAct, orchestrator-worker… and the one choosing rule: can you draw the flowchart before runtime?

Open diagram → reference

The 10-80-10 dial

Three build tiers — neurosymbolic, guarded judgment, deterministic pipeline — plus the two-question ambiguity test and the pyramid discipline.

Open diagram →

Red flags it kills

Install & use

# copy the skill into your Claude Code skills directory
cp -r skills/pick-gtm-agent-pattern ~/.claude/skills/

Then invoke it directly, or just describe the agent:

/pick-gtm-agent-pattern

# or, in plain language —
# "what pattern should this agent use?"
# "make this workflow an agent"
# "is this over-engineered?"