Decomposer

Overview

Decomposer is an AI-powered planning tool that takes your messy, multi-part goals and transforms them into clear, actionable plans. Whether you're starting a business, learning a new skill, or tackling a home project, Decomposer breaks down your goal into 8–14 concrete steps with honest labels that tell you what AI can handle, what requires human judgment, and when you need expert help.

The tool extracts hidden assumptions, evaluates quality across five dimensions, surfaces cross-cutting risks, and gives you a structured plan you can actually follow—complete with copy-ready sub-prompts for each task. Unlike generic chat interfaces that produce paragraphs that fall apart when you try to act on them, Decomposer creates persistent, shareable plans with measured quality gates.

Why we built it

People bring sprawling, multi-part problems to AI and get responses that feel smart but fall apart during execution. Plans miss steps, repeat themselves, or assume magic helpers exist. Most people don't think in no-overlap/no-gaps trees, nor should they need to learn prompt-engineering to get a usable plan. Decomposer is the planning layer that makes problems legible, routes work honestly, and gives you portable sub-prompts to execute step by step—it's the front door to AI that ensures you start with structure, not soup.

How it works

1. Describe your goal – Paste your objective (with optional file attachments). The system forms an "observed brief" and proposes assumptions and constraints you can edit, making the edges of your plan explicit.

2. AI generates a structured plan – Using GPT-5, Decomposer creates 8–14 task cards with titles, specific bullet points, honest effort labels, dependencies, and paste-ready sub-prompts for each step.

3. Quality evaluation – A built-in evaluator scores your plan across five dimensions: coverage, redundancy, actionability, label routing, and synthesis fit. Low scores trigger targeted regeneration.

4. Granularity coaching – The system flags tasks that are too broad (should be split) or too small (should be merged) to keep your progress real.

5. Follow through – Edit cards inline, track status (Not started → In progress → Done), copy sub-prompts, export to Markdown, and share persistent links with others.

Feature highlights

• Template-Free Decomposition – Works across any domain without rigid frameworks
• Honest Routing – Clear labels showing what requires AI vs. Human judgment vs. Expert help (🤖/👤/🧰)
• Quality Evaluation – Five-dimension scoring (coverage, redundancy, actionability, label routing, synthesis fit)
• Assumptions Extraction – Surfaces hidden constraints and makes plan edges explicit
• Plan Synthesis – Coherent summary with risk identification
• Granularity Coaching – Flags tasks that are too broad or too small
• Plan Sharing – Compressed storage with 4-5x compression ratio and persistent shareable links
• Portable Sub-Prompts – Copy-ready prompts for each task to use with any AI tool
• Multi-Model Support – GPT-5, GPT-4, GPT-3.5 (default: gpt-5-mini)
• Live Editing – Edit cards inline with auto-save and localStorage persistence
• Export Options – Copy as Markdown or plain text

Notes

• Current Status: Phase V Complete (v0.4) – Full web UI with editing, sharing, and export capabilities
• Tech Stack: Built with Next.js 15, TypeScript, React, and Tailwind CSS
• Deployment: Hosted on Render with Redis persistence
• Model-Agnostic: Thin adapter architecture allows swapping AI models as quality and cost evolve
• No Lock-In: Everything exports cleanly with copy-ready formats