A Claude Code skill that takes a messy, half-formed idea and walks it through eight fixed stages until it becomes something you can actually hand to an agent: a battle-tested prompt, plus a packaged artifact (a new skill, slash command, or spec) when the idea warrants one.
The output of a run is a single handoff document. At the top sits the final optimized prompt, ready to paste verbatim into a fresh agent session. Below it: the packaged artifact and its usage notes, then the decision log.
Every run executes all eight stages, in order, no skipping. Two of them stop and interview you; the rest flow on their own.
| # | Stage | What happens | Method |
|---|---|---|---|
| 1 | Clarify | One-question-at-a-time interview: purpose, constraints, success criteria | inline in SKILL.md |
| 2 | Draft prompt | First optimized prompt, targeting Claude Code by default | stages/2-draft.md |
| 3 | Grill | Relentless interrogation of the brief and draft; weak assumptions die here | inline in SKILL.md |
| 4 | Shape check | Decides the right Claude Code form: skill, slash command, subagent, hook, MCP, or plain prompt | stages/4-shape.md |
| 5 | Optimize | Diagnose, fill gaps, compress into full + quick versions | stages/5-optimize.md |
| 6 | De-AI | Strips AI-writing tells from every user-facing artifact | stages/6-deai.md |
| 7 | Package | Builds the artifact — a tested skill if the idea is reusable, polished prompt + spec files if it is one-off | stages/7-package.md |
| 8 | Handoff | Assembles the final document and prints it in full | inline in SKILL.md |
Stage outputs land in .idea-forge/<slug>/NN-<stage>.md inside the current project, so an interrupted run resumes from the last completed stage instead of starting over.
idea-forge is fully self-contained. Each stage's method is native to this repository: the small ones live inline in SKILL.md, the larger ones in stages/ and load only when their stage runs. Nothing is invoked from outside; there is nothing else to install.
The methods are distilled from a set of excellent MIT-licensed projects — superpowers, mattpocock/skills, prompt-master, ECC, humanizer, claude-howto. THIRD-PARTY-NOTICES.md maps each stage to its source. The originals go deeper than the slices this pipeline needs; they are worth installing on their own merits, just not required here.
Swapping any stage's method is an edit to one file in stages/.
git clone https://github.com/yourlcfr/idea-forge.git ~/.claude/idea-forge-repo
ln -s ~/.claude/idea-forge-repo ~/.claude/skills/idea-forgeRestart Claude Code (or reload skills) so the slash command registers.
/idea-forge <dump your messy idea here>
Answer the interview questions at stages 1 and 3 — or wave them through — and collect the handoff at the end.
The contract in SKILL.md reflects how I run it: interaction in Indonesian, no AI attribution on any artifact, target tool defaulting to Claude Code. Edit the Contract section to taste; nothing else depends on it.
Run scripts/check.sh before committing any change (or wire it as your pre-commit hook: printf '#!/usr/bin/env bash\nexec scripts/check.sh\n' > .git/hooks/pre-commit && chmod +x .git/hooks/pre-commit). It lints the invariants that hold the pipeline together — cross-references resolve, one canonical handoff filename, dead wording stays dead, no dashes inside templates, exactly eight stages, only stages 1 and 3 interview. Three rounds of adversarial training showed that every serious defect in this skill was text colliding with text; the linter catches that class before it ships.
MIT