A real brief failed ignition. Here's what we fixed, same day.

We test Insomniac Agents on real, unglamorous ideas — not cherry-picked demos. One test brief: a weekend mobile bicycle tune-up service for commuters (internally we called it "tuneatdoor"). One sentence, one service, no frills. It's exactly the kind of brief a solo founder would actually type in.

It didn't finish.

Here's what we found, and fixed, in one sitting:

1. A single real service tripped the offerings schema. Our site-copy schema requires 2–6 offering items per business — a page with just one "service" reads like an unfinished template. But "tuneatdoor" only had one real service. The model dutifully generated exactly what the brief described (one offering) and the run failed schema validation. The fix: when a brief describes a single service or product, the agent now splits that ONE offering into sensible tiers or variants a real business in that trade would actually sell (standard vs priority turnaround, one-off vs subscription) — honestly derived from what the brief already says, never inventing a capability it doesn't support. If that still produces fewer than 2 offerings, the stage retries once with the exact defect named, instead of failing outright.

2. The fallback slug was ugly. With no founder-given business name, onboarding needs some slug before the agent ever runs — and it was falling back to raw brief text, producing something like a-weekend-mobile-bicycle-tune-up-service-for-commuters-in-we. Once the agent generates a real brand name later in the pipeline, we now use that instead and quietly correct the slug.

3. Our own critique reviewer flagged two things that weren't wrong. The self-review stage that checks site copy for honesty and quality was rejecting the tenant's own config-injected contact email as a "fake placeholder," and flagging the renderer's own **bold** emphasis syntax as broken markdown. Both are platform conventions the reviewer just didn't know about — so we now tell it explicitly which facts are correct by construction, instead of asking it to guess.

4. A single-price honesty-gate conflict. A late-pipeline check meant to catch invented prices was tripping on legitimate single-price businesses. Resolved the same day.

We re-ran the exact same "tuneatdoor" brief afterward: all 11 pipeline stages completed, no manual intervention. 197 tests green.

None of this is a victory lap — it's the actual state of building an AI agent platform: real briefs surface real bugs, and the job is finding them before a paying customer does. We'll keep posting the failures here, not just the launches.