ideationv0.26.1

Plan real work before the code gets written.

ideation interviews your idea until every claim carries evidence, sets four adversarial critics on the plan, and produces a contract you approve before anything is built. Then your agent builds it — one reviewed commit per phase, on its own branch. Six of its seven stages are allowed to stop the run.

Claude Code
$/plugin marketplace add nicknisi/ideation$/plugin install ideation@ideation
pi
$pi install npm:pi-subagents$pi install npm:@quintinshaw/pi-dynamic-workflows$pi install npm:@juicesharp/rpiv-ask-user-question$pi install git:github.com/nicknisi/ideation
5evidence gates4critics, read-only0files before you approve1commit per phase

The contract

Every run produces one document to decide on: the evidence behind each gate, scope tiers nested so cutting one does not unravel the rest, an interactive phase graph, copyable run commands, and a decision log that keeps rejected approaches rejected.

docs/ideation/<feature>/contract.html
BOOKMARK GARDENexpress recommended
Gates 5/5
problemthe visit log that already exists
goalssave → revisit → grow, one loop
criteria4 of 6 carry check commands
scopeseasons cut, recorded as deferred
consistencyschema.ts · bookmarks.visits
Scope
Mustsave → seed · visit → growth · wilt query · garden view
Shoulda configurable wilt window
Couldseasons — deferred
Phases · 3 waves
01020304node verify.mjs — 6 checks
Fig. 1 — generated at the end of the interview, before any execution.

The seven stages

Each stage has a defined output and defined grounds for stopping. Stopping is not error handling — it is the mechanism. Scroll and the thing grows.

01Brainstorm

A position on whether the idea is worth building at all. Conversation only — it writes no files, enforced by its tool list.

Stops whenit concludes no — and writes nothing either way.

02Interview

Five gates, each backed by an artifact rather than a confident sentence: the problem, the goals, the success criteria, the scope, and consistency with the code you already have.

Stops whena gate cannot be evidenced. It keeps asking; generation starts at five of five.

03Critique

Four read-only critics — scope creep, over-engineering, hidden dependencies, unfalsifiable criteria — file findings at three severities.

Stops whena blocker is filed. The plan goes back, and the interview can reopen on that gate.

04Contract

One document to decide on, plus the specs that will be executed against it and a routing verdict: walk away, watch it run, or run it yourself.

Stops whenalways. Nothing proceeds until you approve, and the tier you pick bounds the specs.

05Execute

Phases build in waves on an isolation branch, each reviewed before it commits, each commit naming its spec so a resumed run skips it.

Stops whenfive ways — including a reviewer that returns no verdict, or a commit that cannot name its files.

06Harvest

At most three candidate lessons from the run, deduped against what is already recorded. Accept, edit or dismiss each one.

Stops whenyou dismiss them. A learning store nobody curates is only a log.

07Next run

The following interview opens by naming the lesson it is applying, out loud, as its first question. learnings.md is committed with the repo, so it is the team's memory rather than one machine's private state.

Stops whennever. The only stage that cannot stop you — it just points back at 02.

What it refuses

Three constraints that are structural rather than advisory — they are enforced by tool lists and commit rules, not by good intentions.

No blind commits

There is no git add -A anywhere in it. Every commit names the files it wrote; one that cannot refuses instead.

No question cap

The interview runs as long as the evidence takes. A cap would mean shipping an unevidenced gate, so there isn’t one — in practice it is short, because every question arrives with a recommended answer.

No silent edits

Critics are read-only by construction. They file findings; every fold-in lands in the decision log where you can see the reasoning and reverse it.

Commands

/ideation

The one door. Interview, critics, contract, then the fork — you choose express or full review once the evidence is on the table.

/ideation:chart

Too big for the door — you cannot even see the route? It maps the open decisions, one ticket per session, until the way is clear, then hands the route to the interview. Plans only; it builds nothing.

/ideation:brainstorm

Whether to build it at all. Pure conversation — no files, enforced by its tool list.

/ideation:express

The same door with routing pre-answered: one confirmation, an isolation branch, straight into execution.

Execution commands — autopilot, execute-spec, get-goal-prompt — are documented in the field guide.

What it writes

Everything lands in your repository as reviewable files, committed with the code — not held in a service.

docs/ideation/
├── learnings.md              the team's memory, curated by you
└── bookmark-garden/
    ├── contract.html         gates, tiers, phase graph, decisions, open questions
    ├── interview.md          every question and answer, with evidence
    ├── critique.md           findings at three severities
    ├── verify.mjs            the contract's own success checks
    └── specs/
        ├── 01-soil-schema.md
        ├── 02-growth-stages.md
        └── 03-garden-view.md

Install

Installs as a plugin in your coding agent. Nothing to configure, no template to fill in, and no service to sign into — everything it produces is a file in your repository.

Claude Code
$/plugin marketplace add nicknisi/ideation$/plugin install ideation@ideation
pi
$pi install npm:pi-subagents$pi install npm:@quintinshaw/pi-dynamic-workflows$pi install npm:@juicesharp/rpiv-ask-user-question$pi install git:github.com/nicknisi/ideation

Two lines, paste them together — the first registers the marketplace, the second installs the plugin from it.

Then run /ideation and describe the messy version of your idea. A worked example — one invented feature through all seven stages — is on the walkthrough page.