AI said “done.” Was it?
Described transcript
On screen: A black screen with a soft amber glow in the middle. The word “Done.” slams onto the screen in huge white letters with an amber halo, and an amber check mark sits under it. Then the check mark shatters into a question mark: thin white cracks spread from it while the word shakes.
Narration: AI said “done.” Was it?
On screen: A dim field of tall glass monoliths in haze; the camera pushes in toward one of them, which has a glowing amber core. A large amber “39” slams in, with “agents” in white below it and then “under 12 hours” in amber.
Narration: Thirty-nine agents rebuilt a site in under twelve hours.
On screen: Close on the amber core of a monolith, the camera still pushing in. A huge amber “11” punches in, with “false claims” in white beneath it and, smaller, “in one review”. Then “wrong twice, too” appears in amber at the bottom.
Narration: Reviewers found eleven false claims. They were wrong twice, too.
On screen: A dimmed terminal window; “replayed output” is written small in the top-right corner. Over it, in two lines: “You have probably been burned by “done””. The terminal types npx bmad-plus review gate site; echo "exit code: $?" and prints “site: incomplete — coverage 0/1 completed (0 waived), 0 open finding(s), 0 refuted”, three lines of reasons (no coverage.json, no findings.json, one selected file never accounted for: site/contact/index.md) and “exit code: 2”. The word “incomplete” is outlined in orange and the view zooms in on it. Then a dark band appears with “5 steps · 5 commands” in large amber letters.
Narration: You have probably been burned by “done” before. Here is how to check: five steps, five commands.
On screen: A dark, glassy stack of slabs lit by amber light, the camera pushing in slowly. At the top left, a large amber “1” and the title “Write the rules first”. Three cards follow, about a second apart, each with a thin amber bar on its left: “From scratch”, “Facts from the code” and “Simulated? Labelled.” At the bottom, five short bars form a progress line; the first one is amber.
Narration: One. How do you stop invented facts? Write the rules first: from scratch, facts from the code, simulated means labelled.
On screen: A terminal with “replayed output” in the top-right corner. The command npx bmad-plus install --yes --lang en --packs core,design,memory --tools claude-code is typed. The installer prints “BMAD+ Installer v0.23.0”, “Installing to: /home/dev/app”, “Selected packs: Core, Memory, Design”, “Detected IDEs: Claude Code”, “Global brain created: /home/dev/.bmad-plus/brain”, “6 agents, 9 skills, 5 files copied”, “1 IDE(s) configured” and “Installation complete — Getting started”. A dim line says “getting-started box, 40 lines” in place of the box, and the last line, “BMAD+ is ready! Talk to Atlas to get started”, is outlined in orange. A band at the top reads “One install gives your assistant its roles”. The progress line still shows one amber bar.
On screen: A terminal. The command npx bmad-plus intake record is typed with the Design pack’s intake.yaml, an answers file and three answers: surface_kind=landing, audience=“Urban professionals who buy coffee most days” and job=“Get a first-time visitor to open a loyalty card”. The output reads “Recorded surface_kind, audience, job in _bmad-output/design/intake.json.”, then “design-read (design): ready — 3 answered, 0 defaulted, 0 unknown, 10 not asked yet.” and “Ask next (at most 6 per message):” followed by the first questions (brand_assets, design_system, aesthetic_family and more). Orange rings circle “3 answered”, then “10 not asked yet”, then “at most 6 per message”. At the bottom left, a large amber “2” and “Ask first”; the progress line has two amber bars.
Narration: Two. What should it ask before it starts? The intake lists what is known and what is missing.
On screen: A single glass slab splits into many identical slabs that fan out in an arc, each edge lit with a thin amber line, while the camera orbits around them. At the top left, an amber “3” and “Fan out”. A large amber counter climbs from 1 to 39, with “agents” in white below it. Three bars of the progress line are amber.
Narration: Three. How do agents stay out of each other’s way? One brief per page, one copy of the repo each.
On screen: A terminal with “replayed output” in the top-right corner. The command npx bmad-plus swarm init --definition swarm/page-fanout.json --units swarm/units.json is typed; the output reads “site-pages-20261019-g3ip: fanout of 3 units, budget 26, approval not-needed.”, “Units listed, no seed.” and “Next: bmad-plus swarm briefs site-pages-20261019-g3ip”. Then npx bmad-plus swarm briefs site-pages-20261019-g3ip prints “3 of 3 units need a sub-agent, in waves of 10.”, with “in waves of 10” outlined in orange, a first wave listing the briefs of the pages about, pricing and contact, and a line starting with “Disclosure:”: “Each page was written by a separate sub-agent from the same brief; read them together before publishing.” At the end, a dark band at the bottom shows “39 agents”, “133 commits” and “244 pages” in amber, one after the other.
Narration: BMAD+ deals the briefs and caps each wave. Your assistant launches the agents.
On screen: A dark field crossed by orange rays of light. At the top left, an amber “4” and “Make someone else try to break it” on two lines. Five thin amber lines draw one after another, each with a word above it: “facts”, “reference”, “examples”, “languages” and “accessibility”. Four bars of the progress line are amber.
Narration: Four. Who checks the work? Five readers that wrote none of it, replaying commands on a clean server.
On screen: A closed door of dark glass and brushed metal, with an amber beam across its face; the camera pushes in slowly and the door stays shut. In amber monospaced letters: “incomplete · exit 2”, and below it, in grey, “coverage 0/1 completed”.
Narration: And the gate says no until every file is accounted for.
On screen: Thin slabs of glass drift in the dark. Three cards appear one after another. “Your files are kept”, in quotation marks, is struck through by a red line and replaced by “backed up, then replaced” in amber. “/animated” is struck through and replaced by “does not exist”. “26 frameworks” is struck through and replaced by a large amber “27”.
Narration: A promise that your files are kept. A command that never existed. A count off by one.
On screen: A dark grid cube with fine amber lines. At the top left, an amber “5” and “Fix, then prove”. Three rows count up, each ending with a green check mark: “19 fix commits”, “1,879 tests” and “0 accessibility violations”. The five bars of the progress line are all amber.
Narration: Five. Fix, then prove: nineteen fix commits, 1,879 tests, zero accessibility violations.
On screen: A dim dark sky. In large white letters: “A person kept the last call.” Then, in grey, “Not yet measured:” and, in amber, “do readers find it clearer?”
Narration: A person kept the last call. Nobody has measured yet whether readers find it clearer.
On screen: Dark glass monoliths with a warm edge light. A list builds line by line, each with a large amber numeral and a white label: “5 commands”, “3 swarm files”, “1 review gate” and “7 mistakes to avoid”.
Narration: You get five commands, three swarm files, the review gate and seven mistakes to avoid.
On screen: Dark blue haze. In amber monospaced letters, “npx bmad-plus install”, with an amber underline drawing itself under it; then “bmad-plus.rochetta.fr” in white and, smaller and grey, “docs · guides · case study”.
Narration: Start with step one.
On screen: The glass monoliths again, the camera pulling back. “Done” is a claim. appears in white, with “Done” in quotation marks, then fades. A large amber question mark appears with white cracks spreading from it, as in the first shot, and the picture fades to dark blue.
Narration: “Done” is a claim. Replay it.