A six-stage multi-agent methodology, run in a fixed order: feasibility and prompt structuring → a read against the intent → the plan → a pre-mortem against the plan → the revision → the build. Each stage has one job, and no stage gets skipped because the work looks small.
01 · Voice Capture & Vault Ingestion
Ideas get spoken into the Unli Rice iOS app, transcribed to markdown, and synced to the local vault — the same ground-truth memory every connected MCP agent reads from, so no tool starts a session cold.
02 · Feasibility & Prompt Structuring
Gemini Spark is the first point of entry. It reads the raw vault text, checks technical feasibility, benchmarks what already exists, and structures the problem — context, root causes, and what to look at — into the handoff prompt Claude works from. Its output is raw material, not a plan.
03 · The Intent Gets Attacked First
Before anything is planned, GPT-6 Astra tries to break the intent doc: is the stated problem the real one, and is the “explicitly not this” list actually complete? That list is what a swarm's questions get checked against later, so a gap in it surfaces as improvised code three stages downstream. Astra runs the standing AI-failure lens list in the same pass. It returns objections and does not rewrite the intent — if the read changes what's wanted, that's a new intent doc superseding the old one, never an edit that quietly makes the original look right.
This stage exists because the chain used to have three checks and none of them looked at the intent. The plan got attacked, the model's behaviour got attacked, and the business case got attacked six months in — by which point a wrong premise had been built.
04 · The Plan — and Claude Doesn't Build It
Claude turns that prompt into the architecture: a file-by-file task list, the edge cases, the gotchas. It writes the plan into the project's own docs behind a short handoff pointer. If Claude finds itself starting a long mechanical pass, that's the signal to stop and write the plan instead — building is somebody else's stage.
05 · Pre-Mortem, Then Revision
Codex tries to break the plan on paper before any code exists: where the design fails, what the tests miss, which invariants are unguarded, what the plan assumed about the repo without checking. It returns objections, not a rewrite. Claude then takes a second pass — revising what the objections change, and saying plainly which it declines and why. The plan is settled when that pass ends, not before.
06 · The Swarm Builds It — Over an MCP Bridge
The settled plan is dispatched to the Antigravity swarm through a local MCP bridge rather than by hand. The bridge writes the plan into the target project folder and launches the headless agy CLI there with an explicit directory — necessary, because the swarm resolves its workspace from its own project state, not the shell's, and an unaddressed run silently attaches to whatever project was open last. It has done exactly that, into the wrong repo.
Nothing the swarm says about itself counts as evidence: a run whose tools were denied still exits clean and reports success. Every build is checked against git diff and a real test run. Anything the plan didn't anticipate goes back to Claude as a plan change, not an improvised fix. Near release Claude runs the security audit — secrets, server-side authorization, per-record ownership, receipt validation — and the app ships through the relay to the App Store. Afterward, Unli Disk clears simulators and dead-end repos off my own machine.
Visual map of the pipeline: from voice capture through the intent read, the plan, the pre-mortem, and the swarm that builds it — with the MCP bridge that dispatches the build, and the return loops where one tool checks another’s output before work continues.