Three grills, one job: kill bad plans before you build
The grill family interrogates a plan until it has no holes. grill-me is the bare stress-test. grill-with-docs grills the plan against the project's own docs and writes the decisions down. grill-decide-report is Chris Castillo's autonomous version: instead of asking you 28 questions one at a time, it recovers everything you already said into a Feedback Catalog, then decides every open question itself and hands back a Decision Report.
- grill-me: quick one-at-a-time stress test of a plan.
- grill-with-docs: grills against your docs, writes a glossary + ADRs.
- grill-decide-report: decides the open questions for you, writes the report.
use grill-with-docs on this plan:
[YOUR PLAN]
Walk every decision branch against the project docs, one question at a time, with your recommended answer for each.use grill-decide-report to finish the remaining decisions autonomously.
Recover what I already told you into a Feedback Catalog, decide every open question, and write the Decision Report. Only kick back the ones only I can decide.GSD did the planning. It could not go multi-agent.
The stack went GSD, then superpowers, then gbrain. GSD ('Get Shit Done') was good at one thing: detailed phased planning. But it was sequential by design, one atomic task per fresh sub-agent, never parallel. That is not multi-agent, so GSD is out of the stack. The grills now do GSD's old detailed-planning job, except they run multi-agent, multi-LLM, and multi-environment.
- Evolution: GSD then superpowers then gbrain.
- GSD's value was detailed planning. Its limit was sequential-only.
- Grills take over the planning job, now multi-agent.
Plan layer first, then the multi-agent build muscle
Grill is the PLAN and DECIDE layer. It produces a locked Decision Report, then hands that to the BUILD layer: parallel Ralph, Claude Code Teams, or the Workflow orchestrator. That build layer is where the multi-agent muscle lives. The planning itself is also multi-agent: run several grills at once, route deep decisions to Opus and lighter enumeration to Sonnet, and run the whole thing headless anywhere on the fleet.
- Multi-agent: several grills run in parallel.
- Multi-LLM: deep decisions on Opus 4.8, enumeration on Sonnet 4.6.
- Multi-environment: headless across Hetzner, Mac, and the VPS fleet.
claude-as team --model claude-opus-4-8 -p "use grill-decide-report ..." --dangerously-skip-permissions
claude-as team --model claude-sonnet-4-6 -p "use grill-with-docs ..." --dangerously-skip-permissionsTake the Decision Report and dispatch a parallel build:
/ralph [task] --parallel N
(or spawn a Team when the agents need to coordinate live.)We tested it on a real project. It found a real bug.
We ran the grills against a real project on the server, a2a-foundation on Hetzner, two agents in parallel on two different models. grill-decide-report on Opus 4.8 produced 28 decisions across 13 areas, honored all 5 project invariants, and correctly refused to auto-decide the 2 things only Mike can decide. grill-with-docs on Sonnet 4.6 produced 30 grilling questions, flagged 8 code-vs-doc contradictions, and caught a genuine bug in the code. Both runs exited clean.
- Opus 4.8: 28 decisions, all 5 invariants honored.
- Refused the 2 owner-only blockers and kicked them back.
- Sonnet 4.6: 30 questions, 8 contradictions, 1 real bug found.
- Synced fleet-wide: Windows, Hetzner, Mac, vps1/2/3.