DeepSpecGet started
Multi-agent scoring loop is live

Turn rough ideas into
ship-ready specs.

DeepSpec drafts your PRD, TDD, and AGENTS.md, judges them against a quality rubric, and rewrites the weak parts — until everything scores high enough to ship.

Generate your first specSee how it works

Free tier · Sign in with Google · No credit card

Features

Everything you'd hand-build into a spec workflow

DeepSpec ships the boring parts — rubrics, revision loops, and structured outputs — so you can stay focused on the decisions that actually matter.

Multi-agent generation

Specialised drafter, judge, and reviser agents collaborate so each section is written by the agent best suited for it.

Quality scoring loop

A judge model grades every section against a rubric. Low scores trigger automatic revisions until your spec is publish-ready.

Three artefacts in one pass

Get a product PRD, a technical TDD, and an AGENTS.md aligned to each other — not three documents that contradict.

Prompt builder

Craft system prompts the same way: structured brainstorm, generate, ship. Reusable across projects and teammates.

Memory that compounds

DeepSpec learns from your high-scoring specs and avoids repeating critiques that scored low previously.

Bring your own LLM

OpenAI, Anthropic, or a custom OpenAI-compatible endpoint. Your keys, your data, your model of choice.

How it works

From “I have an idea” to “ready for code review”

  1. step 01

    Describe the idea

    A sentence or a paragraph. The drafter agent asks clarifying questions until it has the right shape.

  2. step 02

    Brainstorm into context

    Your answers become a structured context document the rest of the pipeline can reason against.

  3. step 03

    Generate three artefacts

    PRD, TDD, and AGENTS.md are written together so requirements, design, and agent ownership stay aligned.

  4. step 04

    Auto-judge and revise

    The judge scores each section. Anything below the bar triggers a targeted rewrite, up to your plan's revision budget.

  5. step 05

    Ship

    Approve, edit, or grab the raw URLs. Drop AGENTS.md straight into your repo and start building.

Output

Three documents, one source of truth

Each artefact has a deep-link raw URL you can paste directly into another tool, or hand to an AI coding agent.

PRD.xml
<prd>
  <overview>
    Real-time collaborative editor for product specs,
    optimised for engineers as much as PMs.
  </overview>
  <goals>
    <goal id="g1">Eliminate spec drift between PM and eng</goal>
    <goal id="g2">Cut spec authoring time by 70%</goal>
  </goals>
  <user-stories>
    <story id="us-01" priority="must">
      As a PM, I want to draft a PRD by describing
      the idea, so I don't start from a blank page.
    </story>
  </user-stories>
</prd>

Product requirements: goals, user stories, functional & non-functional requirements, risks, and success metrics.

TDD.xml
<tdd>
  <architecture>
    <stack>Next.js 16 RSC + Drizzle + Neon HTTP</stack>
    <pattern>Server actions for mutations</pattern>
  </architecture>
  <data-model>
    <table name="project">
      <field name="status" type="enum">
        BRAINSTORMING | GENERATING | JUDGING | COMPLETE
      </field>
    </table>
  </data-model>
</tdd>

Technical design: architecture, components, data model, integration points, test strategy, and rollout.

AGENTS.md
# AGENTS.md

## Ownership
- `src/app/api/projects/*` &mdash; Eng team A
- `src/components/marketing/*` &mdash; Design

## Conventions
- Server components by default; "use client"
  only when interaction or hooks are required.
- Drizzle queries via `db.query.*` for reads
  with relations; raw `db.select` for narrow ones.

## Workflows
- Always `npm run db:push` after schema edits.

The instruction file for AI coding agents working on your repo: ownership, conventions, and workflows.

FAQ

Questions, answered

Is DeepSpec a wrapper around ChatGPT?
No. DeepSpec orchestrates several specialised agents &mdash; a drafter, a judge, and a reviser &mdash; against a structured rubric. You bring your own model (OpenAI, Anthropic, or any OpenAI-compatible endpoint) and we add the workflow on top.
What does the judge actually evaluate?
Each generated section gets scored 1-10 against criteria like clarity, completeness, testability, and consistency with the brainstorm context. Sections under the threshold get a targeted rewrite request, not a full re-roll.
Can I edit the output?
Yes. Every artefact is editable in-app and exportable to plain markdown / XML. Your manual edits are remembered and influence future judging so the model learns your team's preferences.
What's in the free tier?
10 projects per month, 1 judge run per project, 1 auto-revision per project, and 1 prompt builder slot. The Paid tier raises every limit and unlocks model access controls.
Where is my data stored?
Postgres (we use Neon by default) for project data and an opt-in memory store for high-scoring patterns. We never train models on your data.
Can I self-host?
Yes. The whole app is open and runs on Vercel, Render, or your own Node host. Just point DATABASE_URL at any Postgres and you're good.

Stop writing specs from scratch.
Start shipping them.

Sign in with Google. Generate your first PRD, TDD, and AGENTS.md in under five minutes.

Open DeepSpec