Version 1.0.0 — see CHANGELOG.md. Your installed version is the version: line in SKILL.md.
A Claude Code skill that builds one new feature with several AI models at once. It writes a spec with acceptance criteria you can test, agrees the interfaces between the parts before anyone codes, splits the work by file evenly across the models, has each builder write the failing test first, has a different model on a different subscription review every part against the spec, integrates, runs the full suite, and makes one commit.
The one idea: parallel building fails at the joins. Each builder makes a private assumption about the seam between its part and the next, and the assumptions disagree. So the seams — signatures, data shapes, routes — are written first, and every brief carries them word for word. Builders fill in behind a contract; they do not invent one.
Sister skills: multi-fix fixes a list of defects the same way; sliced-review finds them. All three share one rule: the model that wrote something never checks it, and the check is never on the same bill.
git clone https://github.com/Mariomarquezt/multi-build ~/.claude/skills/multi-buildRecommended alongside it: sliced-review, for the backend wiring notes (references/runners.md) this skill points at.
The number is how many models take part.
| you type | who builds | who reviews | use it when |
|---|---|---|---|
/multi-build 2 |
Sol | Grok | a small feature, one or two files |
/multi-build or /multi-build 3 |
Sol, Grok, Opus — equal shares | each reviews the next, A→B→C→A | the normal case |
/multi-build 5 |
those three + GLM Flash + Composer — equal shares | fixed ring Sol → Opus → Grok → GLM → Composer → Sol | a large feature with many separable parts |
/multi-build fast |
Composer | Sol | cheap and quick |
/multi-build dry |
nobody | nobody | write the spec, seams and split, then stop and show them |
Suffixes: report builds a Markdown + HTML build report without asking; no wrap skips the closing wrap-up. They combine: 2 report, 5 dry.
The model roster is the author's measured stack. Swap in whatever backends you can reach; the procedure does not change.
| Path | What it is |
|---|---|
SKILL.md |
The skill: modes, then ten steps from spec to one commit |
templates/spec.md |
The feature spec: goal, acceptance criteria, out of scope, constraints, seams, the split, what shipped |
templates/build-brief.md |
What every builder is told: the whole spec, every seam, its files, the no-touch list, test-first procedure, output shape |
templates/review-brief.md |
What every reviewer is told: spec, seams, diff, the builder's red/green output, one of three verdicts |
CHANGELOG.md |
What changed in each version |
- Never dispatch a builder against an open question or a missing seam. Ask first.
- Never report a criterion as met without naming the test that proves it, and having run it.
- Never let a model review its own part — and "own" means the model family.
- Never let a part and its review sit on the same subscription.
- Never change a seam silently after consumers built against it.
- Never commit per builder. The tree is only known good after the full-tree pass.
MIT — see LICENSE.