From "we should do something with AI" to a ranked shortlist — markdown skills, one folder.
Version: see VERSION · September 2026 · https://github.com/ditomax/idea
For users, START.md is enough — three steps, no installation. Conversation and output files follow the user's language; headings in the templates stay in English.
If you are reading this from a link and have no local copy yet:
- Get one —
git clone https://github.com/ditomax/ideaor unzip the latest release from https://github.com/ditomax/idea/releases — into a folder the user chooses (default:./ideanext to their other work). That download is the only network access; the "no network" rule inAGENTS.mdapplies from then on. - Work from inside that folder and read its
AGENTS.md. Do not act as the Director from this README — the rules live inRULES.mdandskills/. - Layout: standalone (this folder as is) unless the user has a project with a
planning/folder — then the suite belongs atplanning/suite/idea/andplanning/AGENTS.mdis the entry point, not this file. Thatplanning/form is produced by skill-suite-setup (see The suite), never assembled by hand. - A clone gives updates via
git pull; the user's work inideas/is ignored by git, so nothing of it is ever committed to a clone.hooks/andguard.pyare for skillset developers — ignore them.
idea is the first of three skillsets — idea → maquette → build — that take an organisation from a vague wish to a shipped product without anyone re-explaining what an earlier step already established. This one finds and evaluates ideas:
| Stage | "Team member" | Who sits at the table | Result | approx. |
|---|---|---|---|---|
| 1 collect | interviewer | one employee (solo), or one collector who interviews colleagues, or an import of existing ideas | cards/<ID>-card.md — one idea card per idea, evaluation-ready but unscored |
45 min per session, 3–5 cards |
| (merge) | consolidator | the collector between sessions, or the committee inside a round — optional | new cards from existing ones: consolidate (n → 1), umbrella (n → 1, the sources stay as variants), split (1 → n); the originals stay untouched and traceable | 15 min per operation |
| 2 evaluate | committee moderator | a committee, one writer at the keyboard | evaluations/<ID>-eval.md per card and 10-shortlist.md — ranked, with reasons |
30 min per card, 30 min for the shortlist |
Finished results look like the files in examples/ — one card, its evaluation and a two-entry shortlist for a fictitious company. A Director skill (idea) reads the folder, tells the user where they stand, and calls the stage. The user types start, then next, redo or stop.
Collect never judges; evaluate never generates; merge only regroups what people decided belongs together — it never scores and never drops a line. That separation is the method: the interview surfaces pains before technology, the committee assesses concept maturity — never the person, never the idea — with ten fields, K.o. gates and an index anyone can recompute by hand. Where a K.o. field is unknown, the committee estimates or waives; ideas are never sent back and never silently dropped.
| Repo | What it does |
|---|---|
| idea — this repo | vague wish → ranked shortlist (10-shortlist.md, contract H1) |
| maquette | one shortlist entry → clickable model and brief (60-brief.md, contract H2) |
| build | brief → product, with concept documents as the source of truth |
| skill-suite-setup | puts the three into one project folder (planning/) with a profile and a single entry point that knows which skillset is up; builds customer-specific versions |
Each skillset works on its own. Anyone who wants more than one of them, or a customer-specific version, gets the planning/ form from skill-suite-setup instead of standalone folders side by side.
- Out — H1.
10-shortlist.md(contractH1/2) is read by maquette. The committee names under Maquette order which entry is built next and, for an umbrella entry, which variant the demo shows; maquette ≥ 0.6.0 proposes exactly that entry, the user confirms or picks another — always exactly one. maquette 0.4–0.5 reads the file asH1/1(lists the entries, ignores the additions). Sparring prefills from the chosen entry instead of asking again. Every entry is self-contained, an umbrella entry including its Variants table. - In — foreign cards. Ideas that already exist in a spreadsheet or another tool enter through collect's import step, mapped onto cards with the gaps marked
[unknown]; the mapping is remembered per format.
Both seams are optional. maquette also starts from an idea in prose; idea also works without maquette.
Compatibility. Out: 10-shortlist.md is contract H1/2 (additive to H1/1), fully read by maquette ≥ 0.6.0, readable as H1/1 by maquette ≥ 0.4.0. In: none (foreign cards enter through collect's import). Version triples tested together are listed in skill-suite-setup/compat.md. Changes: CHANGELOG.md.
idea/
START.md three steps for the human
AGENTS.md entry point for Codex — turns the agent into the Director
CLAUDE.md the same for Claude
VERSION
README.md this file
RULES.md shared rules for all stages — frontmatter, write rules, conversation rules, git, contracts
QUESTIONS.md every question the skillset asks, with stable IDs — the tailoring surface for profiles
CHANGELOG.md what changed per version
hooks/ pre-commit guard for development clones (see Release)
ATTRIBUTION.md where the method comes from
profile/ optional customer-specific constraints (empty = core defaults)
examples/ fictitious finished results — a card, an evaluation, a shortlist (Example GmbH)
ideas/ the users' work, one subfolder per organisation (not in the repo)
templates/ one template per result file (binding content definition)
00-idea.md control file
card.md idea card
evaluation.md ten-field evaluation
10-shortlist.md handover to maquette (contract H1/2)
skills/
idea/ Director
idea-collect/
idea-merge/ card operations: consolidate, umbrella, split
idea-evaluate/
The workspace above is the standalone form. Inside a project folder the suite is copied to planning/suite/idea/ (read-only), the work lives in planning/idea/, the profile in planning/profile/. maquette and build sit next to it under the same planning/. The Director recognises the layout by the planning/ folder — nothing to configure.
This folder is the workspace — repo and ZIP have the same structure. Users download the ZIP of a release, unzip it, open the folder in their AI app and type "start" (see START.md). AGENTS.md (Codex) and CLAUDE.md (Claude) are read automatically and make the agent the Director — nothing to install, no symlinks, no global skill folders.
Developers who want the skills globally can additionally:
ln -s "$PWD/skills/"idea* ~/.codex/skills/ # or ~/.claude/skills/
No dependencies, no network access, no telemetry. ideas/ is excluded from the repo via .gitignore.
Chat tools without folder access (plain ChatGPT, Le Chat, Perplexity): the collect stage of idea can run there as a single prompt file. Ask us for one, or render it yourself with skill-suite-setup: python3 setup.py new <name>, then python3 setup.py prompt <name> idea-collect — the file lands in dist/; a profile is optional, its scope and tailored questions go into the prompt. The other stages need a folder.
Customer-specific variants (restricted topics, IT constraints, standards, corporate design, the customer's own review process, questions skipped or added) do not fork this repo. They live in a profile/ folder the Director reads at start; a profile may restrict, never loosen. The format is specified in skill-suite-setup/PROFILE.md; a minimal example is in profile/README.md. QUESTIONS.md lists every question the skillset asks, with stable IDs — read it before a session, and use the IDs in a profile to skip or add questions.
All skill text, template headings, frontmatter keys and status values are English. The conversation follows the user's language (German → informal "du"). The language of the result files is resolved by the Director: English by default → profile → the language the user writes in → an explicit statement by the user ("I write German, the cards shall be English").
Optional. Skills never create a repository; if one exists, a commit marks a frozen state (a card done, an evaluation done, the shortlist done) and nothing else — the policy is in RULES.md §6.
Development clones activate the customer-data guard once: git config core.hooksPath hooks (the hook calls guard.py from the sibling skill-suite-setup repo and blocks commits that carry customer markers). Release ZIPs are built with skill-suite-setup/release.py, which ships only git-tracked, allowlisted, guard-clean files.
Repository: https://github.com/ditomax/idea — releases at https://github.com/ditomax/idea/releases.
New version: bump VERSION, tag vX.Y.Z, GitHub release with the folder attached as idea-vX.Y.Z.zip. Users update by downloading the new folder and copying their ideas/ (and profile/) across.
collect merges our ki-ideenfindung (interview method for mid-sized companies, idea card) and idea-work (mode discipline, elicitation toolkit, pre-mortem). evaluate is our konzeptskizze skill and the Kriterienkatalog Konzeptbewertung v0.3 — ten fields, K.o. gates, index and confidence unchanged — moved from a one-on-one to a committee setting. See ATTRIBUTION.md.
MIT — see LICENSE.