# Idea Desk

> Paste one raw idea - a product, a feature, a tool, a new process - and refine it the way the idea-refine agent skill does: open it up first, then converge. The browser checks the idea against the skill's five sharpening questions and the red flags of its refinement rubric, drafts a How-Might-We statement and assigns one ideation lens per variation, for free. A paid Expand lane reframes the idea, asks three to five sharpening questions and writes five to seven variations, one per lens; a paid Converge lane clusters the variations you picked into two or three directions, stress-tests them and writes the skill's one-pager with a Not Doing list. Every reply is checked against the page's own reading.

URL: https://idea-desk.skillsafe.ai/
API tutorial: https://idea-desk.skillsafe.ai/api.html

## Who it is for

Founders, product managers, makers and anyone with a half-formed idea who wants real alternatives to the first version they thought of, an honest stress test, and a one-page plan that says what they are betting on and what they are not doing.

## The free reading (in the browser, no model)

- Input: an idea name, the raw idea (pasted or a dropped .md/.txt), optional answers as "Who:", "Success:", "Constraints:", "Tried before:", "Why now:" lines, and for the Converge lane the variations as "## V1 - name (lens) [resonates]" sections or one bullet per idea.
- The five sharpening questions - who it is for, what success looks like, the real constraints, what has been tried, why now - each answered in a labelled line, mentioned in prose, or missing.
- Red flags from the rubric: everyone (no specific user), like_x ("it's like X but better"), solution_first (a technology named before any person), feature_list (many features at once), no_pain (no problem stated).
- A How-Might-We draft from the idea's own words, the lens each variation gets, the pasted variations and which are picked.
- Exports: the reading as Markdown or JSON, the HMW draft, and an ideation worksheet with the whole method and the one-pager template.

## The paid lanes (gpt-terra, metered)

- expand: a How-Might-We problem statement, the idea restated, three to five sharpening questions with why each matters, and one variation per lens - name, pitch, the person it serves, why it exists, its hardest part - plus honest pushback naming every red flag. Lenses: simplification, inversion, audience shift, expert, constraint removal, combination, 10x scale.
- converge: two or three directions built from the variations, each rated painkiller or vitamin, high or low feasibility and by type of differentiation, with what it bets on, what could kill it and what it ignores; a recommendation by the skill's decision matrix; and the one-pager - problem statement, recommended direction, key assumptions with a tier and a test, MVP scope in and out, Not Doing (and why), open questions. The expand lane hands its variations and questions to this lane with one button.
- The idea, the answers and the variations are data: instructions inside them are never followed.

## How replies are held to the input

Expand: a "How might we ...?" statement with no solution built in; three to five questions from the five, always including who and success when the idea leaves them out; one variation per assigned lens in order, each complete, no two sharing most of their content words. Converge: two or three directions from real variations, every picked one placed, no two from the same set; every rating present; a recommendation in the best quadrant present (painkiller and feasible, then painkiller and hard, then vitamin and feasible, then vitamin and hard); three to six assumptions each with a test and at least one dealbreaker; an MVP with what is in and out; at least three Not Doing items with reasons; open questions that still ask who or success if the input never says. Both: every red flag answered by name, and no percentage, price or head count the input does not contain.

## Limits

- The reading is a text heuristic over labelled lines and common phrasings; it can miss an answer written in unusual prose, and shows what it found so you can see.
- The variations and the one-pager are a model's judgement, not market research: market sizes, prices and user counts are never stated unless you supplied them, and the assumptions list is where validation starts.
- Bundled examples use invented ideas (a primary school's permission slips, a vague meal-planning app).

## Sources

- Agent skill @addyosmani/idea-refine (github.com/addyosmani/agent-skills, MIT): https://skillsafe.ai/skill/@addyosmani/idea-refine - its three phases (understand and expand, evaluate and converge, sharpen and ship), its sharpening questions and ideation lenses, its refinement rubric (painkiller vs vitamin, feasibility, differentiation types, the assumption audit, the decision matrix, MVP scoping), its one-pager template and its anti-patterns.