Launch Control

One idea in. A whole week of on-brand launch content out, written, made, and graded by a swarm of Claude agents.

Built for Claude Build Day. Give it a goal, a call to action, and a nonprofit's website. A swarm of Opus 4.8 agents researches the brand, plans a 7-day content arc that crescendos to the event (e.g. a Saturday beach cleanup), writes the copy for X / LinkedIn / Instagram, renders the images and UGC/launch videos, and grades its own work against a rubric, regenerating anything that fails before a word goes out. Connect your socials through Zernio and approve each post or auto-post the whole week. Agents watch the comments and reply in your voice.

See docs/brief.md and docs/rubric.md.

What's new vs. off-the-shelf

• The agents grade themselves against a fixed rubric and regenerate failures. "Done" is verifiable by the model, no human: every slot green + the URL responds.
• It doesn't stop at content. It ships and works the comments.
• It reruns on any campaign by changing three inputs.

Orchestration in one breath

one planner → one critic loop (grade → fix → re-grade, every slot in parallel)
            → one rubric (docs/rubric.md) → one done-check

Two zero-dependency scripts make the orchestration simple, repeatable, and verifiable by the model with no human:

# rerun the engine on ANY campaign — swap the three inputs
node scripts/launch.mjs --goal "..." --cta "..." --website "https://..."

# "done" = one command, one exit code: rubric tests + URL 200 + scorecard green
node scripts/verify.mjs --url <deployed-url> --run --goal "..." --cta "..." --website "..."

docs/rubric.md is the single contract — the critic enforces it, the offline test suites (lib/*.test.ts, no API key) prove it, and verify.mjs (+ CI on every push) checks it. Rerun on a new problem tomorrow by changing three strings.

Run it

cp .env.example .env.local   # fill in ANTHROPIC_API_KEY, FAL_KEY, ZERNIO_API_KEY
npm install
npm run dev                  # http://localhost:3000

Enter a goal, a CTA, and a nonprofit website, then Generate the week. The plan returns with a green scorecard. Click "Make image / video" on any media slot to render it (cached + spend-capped). Connect socials and publish.

Keys (never committed)

Env	What
ANTHROPIC_API_KEY	Claude (Opus 4.8): research, planning, copy, the critic
FAL_KEY	fal.ai: images, UGC video, motion/launch video
ZERNIO_API_KEY	Zernio: connect + publish to X / LinkedIn / Instagram

All secrets live in .env.local (gitignored). Only .env.example (names, no values) is in the repo.

Cost

Video is the expensive part. Renders are cached per prompt, and a process-wide ceiling (MAX_VIDEO_SPEND_USD, default $20) hard-stops runaway spend. Claude runs on Build Day credits.

Stack

Next.js 15 (App Router, TypeScript) · Anthropic SDK (Opus 4.8) · fal.ai · Zernio.

Layout

app/          UI + API routes (generate-week, generate-media, connect, publish, comments)
lib/          anthropic (brains + critic), fal (media), zernio (social), cache, types
docs/         brief.md, rubric.md, architecture.md

Multi-agent orchestration — how we built it this fast

We didn't build this with one agent working top to bottom. We ran a crew of Claude Code sessions in parallel, orchestrated by Claude Classroom — a coordination skill we built earlier, at home (not during Build Day) and used here. Instead of serial work, the sessions self-organized:

• Ownership + file claims. Each session declared the area it operated, then claimed files before editing. One session owned the entire UI (landing, console, sidebar, liquid glass, the scroll story); others owned backend/data (auth + quota, the Supabase store, channels, the comment watcher). Two agents never touched the same code.
• A shared backlog. Work was posted to a shared board and pulled by best fit, so nothing was duplicated and the critical path stayed unblocked.
• A pre-commit guard. The skill blocks committing a file another live session holds, so "merge hell" became a non-event and every change landed clean on main.
• Continuous sync. Every session broadcast what it shipped, so the crew kept a live picture and could rebase instantly.

The payoff: the marketing landing, the full console + Asset Bay + Calendar + Channels, the self-grading media pipeline, auth/quota, and every integration (Claude, fal.ai, Zernio, Supabase, Open-Meteo, Luma) came together in parallel, in a single Build Day — not one feature after another.

Vendored at .claude/skills/claude-classroom/; see USED_HERE.md for provenance.