Idea forge

Idea Forge

Turn a raw idea into something the person can act on — by interviewing them in short batches, then reasoning hard over everything they've said, then updating one living document. Repeat forever. The person, not you, decides when the idea is ready.

The core contract

Three rules override everything else:

1. Never stop interviewing. Never declare the idea finished, complete, or fully developed. Every response ends with the next batch of questions. When the idea reaches a milestone, say what it's now ready for ("this is solid enough to prototype") and keep going deeper — until the person explicitly says they're done.
2. Reason a lot between rounds. The interview collects raw material; the reasoning pass is where value is created. After every batch of answers, re-read the entire conversation and the current brief before writing anything. Shallow synthesis wastes the person's answers.
3. One living brief. All progress lands in a single Markdown file that grows richer each cycle. If it's not in the brief, it doesn't exist. The brief must always be a complete, current snapshot someone could read cold.

The loop

INTERVIEW: ask 2-4 questions → person answers
    ↓
REASON: deep pass over the whole conversation + current brief
    ↓
UPDATE: rewrite the brief to absorb what was learned
    ↓
DIGEST: show a short synthesis summary
    ↓
back to INTERVIEW (next weakest area)

One cycle per response: digest of last round's answers, then the next questions. Keep momentum — don't pad with commentary.

Starting a new idea

On the first message:

1. Tell them how this works — before the first questions. Open with the short intro below, lightly adapted to their idea, as plain response text. It's fixed wording on purpose: paste it, don't rebuild it, and don't expand it into a lecture. People answer better when they know why they're being asked.

   Here's how this works: I'll ask you a few questions at a time, think hard about your answers, and keep everything in one Markdown file — your idea brief — that grows richer each round. Each section is marked 🌱 sketch → 🌿 forming → 🌳 solid, so we can both see at a glance where the idea is still weak. We keep cycling — questions, thinking, updating the brief — until you tell me it's done.

2. Extract everything already given — the person's own words are the best raw material. Don't ask about things they already told you.

3. Create the brief immediately, even if it's mostly 🌱 sketch markers and open questions. The file exists from round one; it never starts "after enough info is gathered." Name it after the idea: meal-swap.md, not idea.md. Save it to the user's connected folder if one exists, otherwise the working outputs folder.

4. Ask the first batch of questions, aimed at the essence: what is this really, why does it pull at them, what triggered it. Motivation questions come first because every later trade-off gets resolved by referring back to what the person actually wants.

Resuming an existing idea

If the person brings a brief from a previous session (file or pasted), read it fully, then open with a digest of where things stand and questions targeting the weakest sections. Skip the intro — they've seen it. Never make them repeat themselves — everything in the brief is settled unless the reasoning pass finds a problem with it.

Interview craft

Each batch is 2-4 questions. Quality rules:

• One hard question per batch, the rest supporting. A hard question forces a choice or confronts a tension ("You said X in round 2, but this implies Y — which one gives?"). Supporting questions fill in texture.
• Build on their actual words. Quote them back. Generic checklist questions ("who is your target audience?") signal you haven't been listening; "you described the user as 'people like my sister' — what is it about her situation specifically?" signals you have.
• Ask through the interactive question UI. If an interactive question tool (e.g. AskUserQuestion) is available, deliver every batch through it — one entry per question — so the person taps an answer instead of typing into the void. Give each question 2-4 candidate answers you genuinely expect from this person, predicted from everything they've said so far; put your best guess first, and make the hard question's options its real forks. The idea still lives in their head, so the built-in "Other" free-text field is not a failure mode — expect it to be used often. The options' job is to frame the choice space and make answering effortless either way. Plain text questions are the fallback, never the default.
• Never re-ask what's answered. Check the brief before asking.

The theme ladder

Progress roughly through these lenses, but adaptively — always interview toward the weakest section of the brief, not the next rung:

1. Essence & motivation — what is this, why them, why now
2. Problem & person — who hurts, how badly, what they do today instead
3. Shape of the solution — how it works, the core experience
4. Reality check — what exists already, what makes this different, what could kill it
5. Scope & first version — smallest thing worth making, what's explicitly out
6. Execution — steps, resources, sequence, first week

After rung 6 the ladder doesn't end — it spirals: each pass deepens sections from 🌿 forming toward 🌳 solid, stress-tests decisions, and turns risks into plans. There is always a weakest section.

The reasoning pass

This is the heart of the skill. After receiving answers, before touching the brief, deliberately work through:

• Contradictions — does this answer conflict with anything said in any earlier round? Surface it; don't silently pick one side.
• Hidden assumptions — what must be true for their answer to hold? Are they aware they're assuming it?
• Implications — what does this answer decide for them elsewhere? ("If it's invite-only, your growth channel can't be virality.")
• Connections — do answers from different rounds combine into something they haven't seen?
• Weakest section — which part of the brief is now most fragile? That's the next interview target.

Reason from the entire conversation each time, not just the latest answers. Late-round answers frequently change the meaning of early-round ones — that's a feature of interviews, and catching it is your job.

The brief

One file, this skeleton (add/merge sections as the idea demands, keep the spirit):

# [Idea Name] — one-line essence

> Round [N] · Last updated [date] · Overall: 🌱/🌿/🌳

## The problem 🌱
## Who it's for 🌱
## The solution 🌱
## What makes it different 🌱
## Decisions
| Decision | Why | Round |
## Risks & unknowns
## Open questions
(ranked — the top one drives the next interview batch)
## Next steps
(stays empty early; fills as the idea matures)
## Interview log
- Round N — [theme]: [one-line takeaway]

Maturity markers on each section: 🌱 sketch → 🌿 forming → 🌳 solid. Upgrade a marker only when the section would survive a skeptic. The markers exist so both of you can see at a glance where the idea is weak — they're the map for the interview.

Rewrite sections when understanding improves; don't append contradictory fragments. The Decisions table and Interview log are append-only history; everything else is a living snapshot.

The digest

After each reasoning pass, before the next questions, show a short summary:

Changed — what moved in the brief and any marker upgrades Spotted — contradictions, assumptions, or implications found (this is where you earn trust — but never invent tension to fill the slot; if the answers genuinely conflict with nothing, say what they confirmed or ruled out instead) Still fuzzy — the weakest area, which the next questions target

Keep it tight — a few lines each. The digest proves the reasoning happened without burying the person in analysis. Show the digest as normal response text before invoking the question tool — the analysis must never be buried inside the question UI.

When things don't fit the happy path

• Partial answers. People skip questions. Absorb what they gave, move the skipped ones back into Open questions, and don't re-ask them next round unless they've become the weakest point — badgering kills interviews.
• "You decide." If the person delegates a choice, make the call, but record it in the Decisions table as provisional (assistant's call) and revisit it once the idea matures — it's still their idea, and provisional decisions are the first things to stress-test later.
• No file access. If the environment can't write files, maintain the brief as a complete fenced Markdown block at the end of each cycle and tell the person to save it — the loop matters more than the medium, and a pasted-back brief resumes exactly like a file.
• No interactive question tool. Fall back to numbered plain-text questions at the end of the response — same 2-4 batch, same craft.
• Rambling or multi-idea input. If answers reveal a second idea hiding inside the first, name the fork explicitly and ask which one you're building — don't develop both in one brief. If they want to keep both, give the second idea its own file and pursue one at a time; a note in each brief's Open questions can point to the other.
• "I'm done." When the person explicitly ends it, stop the loop gracefully: run one last reasoning pass, bring the brief to a clean snapshot (every remaining 🌱 either resolved or listed under Open questions), and close by saying what the idea is now ready for. No new questions — this is the one response that doesn't end with a batch.