Blog

6 min readfeature

The builder round: one prompt, a whole app, filmed end to end

One prompt into an empty directory; 47 tool calls later a full-stack app with a green 39-test suite and a live-verified API. The whole build, on film.

  • #landing
  • #media

Every lane in the capture catalog compresses a story into seconds — a bugfix, a rename, a wiki. The lane nobody had filmed was the whole point of the product: give the agent one prompt in an empty project and let it build the entire app. The builder round filmed exactly that, start to finish, and the cut is now the landing page's capability dex № 023 — "The builder round."

The fixture

The world the camera boots into is an empty directory holding exactly one file: PRODUCT-BRIEF.md. The brief describes Glyph Quests — a small full-stack quest tracker with an architecture contract, not a TODO list:

  • a JSON file store with validation, seeding, atomic persist and corrupt-file recovery (store.ts);
  • a REST API layer over it (api.ts) and a Bun HTTP server (server.ts) on a fixed port;
  • a dependency-free vanilla frontend (public/index.html, styles.css, app.js) that must also work standalone — opened from disk, with no server, on localStorage;
  • a test suite (quest.test.ts) covering store CRUD, validation, atomic persist, reload and end-to-end API tests over a real Bun.serve on an ephemeral port;
  • and a mission section: run the tests until they all pass, verify every endpoint with curl, fix what is broken, and end with a one-paragraph summary.

The only keystroke a human sends is the mission prompt itself. Everything after it — every read, write, edit, test run, server boot and curl — is the agent's own work, at 30 characters per second, the speed a person types.

The session

The filmed run (against the live stack: bridge + engine, agnes-3.0-flash through the Agnes AI upstream) is a 12-minute session the receipt records event by event: 97 events, 47 tool calls, 13:16:11 to 13:28:12. The agent read the brief, planned four todos — backend, frontend, tests, verify — and then worked the list:

  • wrote store.ts, api.ts, server.ts and the three frontend files, plus README.md and a .gitignore that keeps data/ out of git — eleven files by the closing inventory, +2211 −47 lines on the session footer;
  • wrote quest.test.ts (39 tests) and ran the suite — green on the first completed run, then kept hardening it: cleaned a stale field and an unused import out of store.ts, rewrote the server.ts fetch handler to pass the raw body string into the API layer so JSON validation errors surface as 400s, converted flaky it.each blocks to plain it blocks, and made the updatedAt assertion millisecond-safe — each fix followed by another suite run, five and then seven times in a row to prove stability;
  • booted the server on :8787 with a clean data dir and ran the live curl round-trip: seeded list, stats, a 201 create → PATCH → delete round-trip with the id confirmed gone afterward, a 404 on an unknown id, a 400 on a validation error, and the statics (/, styles.css, app.js) checked for content-type — unknown paths falling back to the index;
  • then killed the server and verified the kill — a two-second curl probe for "STILL UP" before calling it dead — and closed with the one-paragraph summary the brief asks for, all four todos checked.

The suite's first green run, mid-fix-iteration — the todo panel still shows test work in progress

The verification beat: the curl round-trip script and its 404/400 JSON answers on screen, the fourth todo in progress

The closing summary — 39 tests green, five endpoints verified live, 4/4 todos

An earlier session the same morning (run 3 of the round) built the same brief with a different plan — eight todos, 68 tool calls — and also finished green with the same curl verification. Two independent sessions, two different plans, one outcome; the shipped cut is the second one.

What the camera saw — and what it took to film it

The lane is honest, and so is this section: filming a 12-minute build tested the capture harness harder than anything before it. The four cuts it took are now lessons in the workflow contract (docs/tui-capture-workflow.md):

  • Cut one anchored stills on bare filenames and fired on the model's plan text minutes before any work happened. Cut two anchored on the todo panel header and died — the header scrolls off the 36-row fold while the panel's rows stay. Cut three made every wait soft (a miss logs and captures the pane as-is — a long real session is worth more than a perfect anchor) but its green-suite anchor excluded " fail"… and cut four shipped that exclude: bun renders a 0 fail row on a green piped run, so the wait excluded its own match for ten full minutes. The stills for "verifying the server" and "app built" came out byte-identical — both the same idle summary, captured twice after the session had already ended.
  • The fix for the shipped cut was not an image edit: still 06 was re-curated from the frame record — frame 568 of 1514, timestamped mid-round-trip by the session receipt and verified by the vision lane (the curl script, its 404/400 answers, the live spinner, the todo panel). The substitution and its evidence are documented in the capture root's CURATION.md; the superseded still is preserved, not deleted.
  • The assembler learned the same lesson: the frames after the session ends are harness time, not story, and the video now trims that trailing idle to one final dwell (rule 2b in assemble.py). The shipped cut is 651 frames at 8fps — 81 seconds paced from 1514 captured frames.
  • And the launch itself: this sandbox reaps the launching shell's process tree, which killed two silent nohup/setsid attempts eight seconds in. The round's driver runs detached inside tmux, which survives.

Every shipped still passed the three-layer audit (pane truth, pixel verification, vision check) and the video passed frame-sampled vision audit 5/5 — the same gates as every other lane; the curated still carries its receipt-timeline evidence alongside.

What ships

  • The lane: public/landing/tui/build-session/ — seven audited stills plus the mp4/webm pair, catalogued in the tui README.
  • The dex entry: capability № 023, "The builder round," on the landing page — the reel plays the whole build, muted, from the writing-app poster.
  • The harness fixes: the fifth-cut anchors (the green wait anchors on the green-unique 0 fail row with no exclude; the verify wait reads the session receipt — the agent's own record — instead of screen text), the assembler's trailing-idle rule, and the workflow contract's new failure playbook rows. The next builder round films on the first cut that matters.

Receipts

  • The lane and its audits: public/landing/tui/build-session/ + public/landing/tui/README.md (capture workflow: docs/tui-capture-workflow.md).
  • The scenario spec (fifth cut): scripts/tui-capture/scenarios/index.js.
  • The session receipt: 97 events, 47 tool calls — the curation note quotes the timeline (CURATION.md in the capture root, outside the repo).
  • The changelog: the round ships in the 0.0.33 release entry (folded with the live-preview round).

What ships next

The catalog's next lane candidates are the ones the dex still marks pending — each gets the same treatment: one real session, one honest cut.

Spotted something wrong or worth adding? The record is the repo — open an issue.

All posts