One sentence in. The right rules out.

a dance-studio SaaS · routine ops — no skill needed

$ “can you pull from remote? the client put a doc there”

Recognized as mechanical git work and handled directly — no skill fired, no ceremony. The classifier knows when to stay out of the way.

an agency CRM · an explain question

$ “explain this ID-constants file properly — this is where the drift happens”

The lighter skill, not a full blueprint — a single-file convention question answered from the catalog, drift documented at its source.

a café site · a fix-it ask

$ “read how we do interfaces in the catalog — the site navigation config still isn't set up right”

Found three duplicate type families, a circular import, and four dead auth links — all fixed against the catalog shape, build- and render-verified in the same loop.

a mailing-metrics tool · a planning ask

$ “help me plan what our AI agents are going to be — that could be a blueprint on its own”

A tracked, three-phase build — schema, roster, runtime — plus an admin fleet surface, every phase spec'd from the fetched patterns before a line was written.

an agency CRM · a build request

$ “replace the calendar's create/edit with a multi-step wizard, like the account one”

The wizard shipped by mirroring the existing pattern — and a follow-up blueprint reused the same meeting-token API instead of building new invite infrastructure.

a mailing-metrics tool · a build request

$ “launch the MCP server build — the blueprint queued it as its own task”

One blueprint's task queue dispatched a second named skill — the MCP server and its connections UI shipped in the same run. Skills compose.

a dance-studio SaaS · a security catch, at write time

$ “(no prompt — a write tripped the guard mid-build)”

The guard flagged a required database-role grant as a leaked credential — a false positive. The agent refused to bypass it, surfaced the tradeoff, and the owner approved a scoped fix that kept the real leak rule live everywhere. Protection, not just automation.

Composure's own repos (dogfooding) · a routing question

$ “why did the generic review skill run instead of ours?”

The classifier's own routing gap became a blueprint, the docs were corrected, and the fixed pipeline reviewed the fix — the system catching and correcting itself, live.

a nutrition-coaching app · a product-planning ask

$ “auto-generate the captions — and build our own engagement dashboard”

The planner mapped per-tier screen depth first; its output directly scoped an expanded blueprint that superseded the narrower one — plugins composing across the suite.

a nutrition-coaching app · an enforcement misfire, handled with judgment

$ “(no prompt — an enforcement hook fired on the wrong trigger)”

A sub-agent's brief tripped the testing hook on ordinary prose. The orchestrator recognized the misfire, refused to obey the injected directive blindly, verified nothing was lost, and documented the gotcha — enforcement AND skepticism, both real.

an operations dashboard · a goal to pursue

$ “run a campaign to bring the assistant up to par — pick from the backlog”

Audit baseline → phased blueprints → review against the blueprint's own boundaries → tests on every changed file — then the campaign resumed itself and opened Phase B unprompted.

a child-led AI learning platform · a foundation build

$ “set this project up for success — screens should import cleanly from the entity registry”

Two entity tiers generated from the registry, tenant routing named — then a loop drove the entire mobile build through self-scheduled check-ins, each reporting real shipped milestones.

a dance-studio SaaS · a goal to pursue

$ “let's build the entire back office now — for the studios, for the parents”

The very first message of the session — and ten phases later the whole multi-tenant backend existed: schema and tenancy, people, calendar, enrollment, billing, comms, role-switched dashboards, parent portal, automations. The owner's own tally: 52 agents, seven hours.

a dance-studio SaaS · a goal to pursue

$ “kick off P0 — but make it a campaign, resumable across sessions”

Six phases — data truth, bug floor, CRUD, entity registry, product core with billing, mobile parity — each phase its own blueprint, each verified by a live e2e stream, ending committed, pushed, and released.

a practitioner's site + a research SaaS · a full build

$ “start both products — the practice site and the research SaaS”

One genesis call birthed both repos with a pre-authorized overnight loop — green by morning. Then the compose chain locked the brand, voice-locked the copy to a proof ledger, and built seven pages on the locked system: a live site.