Idea-to-build

This release came out of a real post-mortem: a project built without idea-to-build that died by optimizing the parameters of a thesis that was never true. idea-to-build was front-loaded (critique once, before the build) and had no presence in the execution loop where that failure happens. v0.5.0 fixes that.

/recheck — re-critique a built project with real results. Come back with your execution data and the critic re-runs in re-critique mode under one question: is the thesis failing, or only the tuning? It rules out build-drift and proxy metrics, gives the thesis a fair bounded test, diagnoses by signal presence, and returns 🟢 keep / 🟡 pivot the thesis / 🔴 kill — pointing back to the alternatives you rejected so you don't tunnel a dead approach.

The critique separates the thesis from the parameters and diagnoses which is failing instead of defaulting either way — so it avoids both tuning a dead premise forever and killing a good thesis that was merely under-tuned.

The planner tests the premise first (cheap, fair, bounded) and prefers parallel experiments when they're cheap and reversible, rather than serially tuning one.

The scaffolded CLAUDE.md carries the discipline into the build so the building agent thinks wide instead of tunneling: the core thesis, the rejected alternatives as fallbacks, the #1 premise-risk, and explicit guardrails (measure the real outcome not a proxy; separate thesis from parameters when results disappoint; check build fidelity; parallel-test cheap reversible options; come back and /recheck).

Generation gains a reversibility / parallel-testability lens (two-way vs one-way doors).

Install

• Claude CLI: /plugin marketplace add winchxyz/idea-to-build, then /plugin install idea-to-build@idea-to-build
• Cowork / Claude Code: add that marketplace, or Create plugin → Upload plugin with idea-to-build-plugin.zip (attached).

Full notes: CHANGELOG.md