idea-os – Data & Science skill for AI agents

idea-os

An operating system for turning a raw idea into a build-ready plan. Takes a rough problem statement and produces four files: clarifying questions, deep research, a PRD, and a phased execution plan with platform/stack picks, a user-journey diagram, and kill criteria.

Overview

idea-os is a 5-phase sequential pipeline where each phase's output feeds the next — research shapes the PRD, PRD shapes the plan, plan's kill criteria tie back to research insights. Unlike single-command PRD generators, idea-os refuses to write a PRD until research is done, and refuses to write a plan until the PRD is stable. Depth and vocabulary adapt to a two-axis classification (complexity × builder sophistication) so a first-time builder isn't drowning in jargon and a founder gets full rigor.

Source: https://github.com/Slashworks-biz/idea-os — full skill, 11 reference files, 4 asset templates, and a 590-line worked example.

When to Use

• Use when a user shares a raw product idea or problem statement and wants a structured pipeline from clarifying questions through deep research, PRD, and a phased execution plan.
• Use when the user says "I have an idea for…", "help me build X", "validate and plan this concept", or "what should I build?" — and wants files they can take forward, not a one-shot answer.
• Use when a non-technical founder, PM, or hobbyist needs structure to bridge the gap between "idea" and "Monday morning's build queue".
• Do not use for quick sanity-check feedback on a half-formed idea (use idea-refine instead) or for editing an existing PRD (use product-management instead).

The external reference/template/example files are not bundled. Use the inline workflow without claiming those materials were loaded. Save the four linked files in a new dedicated authorized output folder and preserve existing documents; do not overwrite root files blindly. Search counts are a process heuristic, not evidence of research quality. Separate observed demand, community anecdotes and market-size assumptions, use compatible geography/time/customer definitions, and label unvalidated retention/paid-pilot targets as hypotheses.

How It Works

Phase 1 — Triage

Classify the idea on two axes before anything else. Depth of research/PRD/plan and question count scale with complexity; vocabulary scales with sophistication.

• Idea tier (T1/T2/T3) — T1 = weekend utility, T2 = SaaS MVP or AI wrapper, T3 = marketplace / B2B SaaS / regulated.
• Sophistication (S1/S2/S3) — S1 = non-technical, no framework names; S2 = hobbyist, introduce frameworks with definitions; S3 = founder/senior PM, full vocabulary.

State the classification in one line (e.g. "T2 · S2 — moderate SaaS, builder has shipped before") before proceeding.

Phase 2 — Clarify

Write questions.md with 4–18 questions (count scales with complexity), grouped: Who and Pain · Scope and Wedge · Constraints and Goals. Every question must be actionable — the answer has to change what you build. Generic questions are rejected.

After writing, stop and wait for answers. Do not proceed to research until answered or autonomous-mode assumptions are declared.

Phase 3 — Research

Write research.md using WebSearch + WebFetch. Minimum: 5 WebSearches, 2 WebFetches on named competitors, 1 source per TAM number, date on every source. Anything unsourced gets flagged [assumption].

Required sections: problem validation, JTBD, market (TAM/SAM/SOM top-down + bottom-up), competitors (direct/indirect/substitutes + positioning map), SWOT, distribution (first-100-users channel fit), risks, and 3–7 non-obvious insights.

Phase 4 — PRD

Write PRD.md with: falsifiable problem statement, named personas, ranked JTBD, non-goals (mandatory — it's where bad PRDs die), leading and lagging metrics.

Phase 5 — Plan

Write plan.md with: user journey (text + mermaid), platform recommendation tied to research findings, stack in conservative/modern/cutting-edge matrix, phased build (MVP → v1 → target) with kill criteria per phase and first-100-users distribution per phase, metrics per phase, and 3–5 immediate next actions.

Limitations

• Requires user input between phases for best results; if answers are missing, outputs depend on explicit assumptions.
• Produces planning artifacts (questions.md, research.md, PRD.md, plan.md) but does not execute build or deployment work.
• Source quality determines output quality; weak or outdated references can reduce recommendation accuracy.
• Better suited to new-idea validation and early planning than late-stage optimization of an existing shipped product.

Examples

Example 1: Non-technical founder with a consumer-app idea

User: "I want to build a habit tracker for people with ADHD."

idea-os classifies T2 · S1, writes 8 plain-language clarifying questions, runs research with sourced competitor pricing and community signal from ADHD subreddits, produces a PRD with ADHD-specific non-goals (no streaks, no punishment mechanics), and a plan with a single-screen MVP and a kill criterion tied to 14-day retention.

Example 2: Founder with a B2B SaaS idea

User: "I'm thinking about procurement software for mid-market manufacturers."

idea-os classifies T3 · S3, writes 18 questions including procurement-cycle specifics, runs research with Wardley-map option and Porter 5 forces, produces a PRD with tiered personas (buyer/approver/IT), and a plan with a phase-1 kill criterion tied to paid-pilot close rate.

Full source

Full 11-reference skill, 4 asset templates, worked example, and MIT license at https://github.com/Slashworks-biz/idea-os. This antigravity entry is a reference copy — the upstream repo is where ongoing development lives.