A disposable dev machine builds the rest
For decades, building software has meant a developer typing the application code line by line. This story is a different job: one developer authors precise specifications, and an Amplifier-driven "dev machine" implements against them.
The reframe only earns your attention if the machine can actually ship at scale — so here's the proof.
word4 is a word processor built by the autonomous dev-machine pattern. Its committed pattern report documents the run — no cherry-picked demo.
Real leverage, verified. But why can't you just point a model at the whole thing and let it code?
A model can't reach real scale in one pass. And a session run long enough to try just degrades: context rots, and the output stops being trustworthy.
The pattern sidesteps both problems at once — by refusing to run long sessions at all.
No session runs long enough to degrade. Memory lives in files, not in the context window — so quality never rots between batches.
Disposable keeps quality from decaying. But how do you know each batch is actually correct?
The bundle ships 8 templated recipes that gate the build loop. The health check loops until the codebase is clean — autonomous output only ships once the gates pass.
Disposable plus gated makes autonomous output safe to ship. Now watch it run on a real app.
Chris Park is the single developer behind Ridecast. The same developer, product, and spec-driven workflow are independently corroborated by the committed Ridecast Phase 4 story he submitted.
Now the headline number has a home — here's what one day of the machine produced.
Chris Park — single developer behind Ridecast / Ridecast2
Ridecast Phase 4 story (.docx), committed 2026-03-14
Phase 4 story: 17 sessions, 44 features across earlier Ridecast phases
Per the committed deck, one day on Ridecast2 shipped 22 production features across 8 dev sessions — with zero lines of app code written by a human.
The developer wrote the spec; the gated, disposable machine shipped the code.
It's reusable tooling, not a one-off: a hard-gated three-mode progression and templated recipes. The generated machine lives in your project repo with zero runtime dependency on the bundle.
Write the spec. Let the machine ship the code.
Feature status: Foundry (amplifier-bundle-dev-machine) live/complete, last commit 2026-06-18. word4 shipped. Ridecast2 reported Active, verified 2026-03-06 in the committed deck.
Research performed:
cd amplifier-bundle-dev-machine && git remote -v → git@github.com:ramparte/amplifier-bundle-dev-machine.gitgit rev-list --count HEAD = 164; git log --format='%ai' → 2026-02-25 to 2026-06-18git log --format='%an' | sort | uniq -c | sort -rn → Sam Schillace 80, Marc Goodner 76, Samuel Lee 4, ramparte 4sed -n '1,60p' bundle.md; ls modes/*.md (3); ls agents/ (6)ls templates/recipes/*.yaml | wc -l = 8cd word4 && sed -n '540,560p' docs/autonomous-development-machine-pattern.md (278 features, 106 sessions, 5.5 days, 445 commits, 3,714 tests, ~89K LOC); git rev-list --count HEAD = 5,246 at current HEADcd amplifier-stories && git log --diff-filter=A --all -- '*dev-machine-22-features-one-day.html' → c05d143, cpark4x, 2026-03-07git log -- 'stories/ridecast-phase4/*' → 91c9f84, Chris Park, 2026-03-14Gaps: The Ridecast2 source repository is NOT present under /home/ramparte/dev/ANext. All Ridecast2-specific numbers (22 features, 8 sessions, tests 58 → 151, +93, zero human-written app code) come only from the committed deck ("verified 2026-03-06" from that repo's git log + STATE.yaml) and could not be independently re-verified from git here.
Primary contributors: Sam Schillace (80 commits) and Marc Goodner (76) on the foundry; Chris Park, the single developer behind Ridecast/Ridecast2.