lovable-reverse-engineer-spec

A portable agent skill that turns a Lovable.dev project into a structured, reimplementable specification — without locking you into React, Supabase, or shadcn.

Works with:

• Claude Code — via SKILL.md
• Codex / Cursor / Aider / Cline — via AGENTS.md
• GitHub Copilot — via .github/copilot-instructions.md

What problem does this solve?

You prototyped on Lovable. It works. Now you want to:

• Rebuild it on a production stack (Next.js + Prisma, .NET + Postgres, whatever).
• Onboard teammates who weren't in the chat.
• Audit what the AI actually shipped vs. what you asked for.
• Feed a spec into a spec-driven framework (GSD, Spec-Kit, BMAD, Kiro).

Reading the raw chat transcript doesn't scale. The repo alone misses the intent, the failed attempts, and the business rules expressed in prose. This skill reads both and emits a versioned spec plus a GSD-compatible planning tree.

Inputs

Flag	Required	Purpose
--repo <path>	✅	Local clone of the Lovable GitHub repo
--chat <path>	✅	Export from loveable-chat-history-capture
--teams-chat <path>	optional	Teams/Slack/email transcript (e.g. CHAT.md) where stakeholders discussed scope, strategy, external context. Pulls in signal that's absent from Lovable chat and the repo.
--out <path>	optional	Output dir (default: <repo>/.reverse-engineered/)

Outputs

<out>/
├── spec/                          stack-agnostic, portable
│   ├── 00-overview.md
│   ├── 01-domain-model.md
│   ├── 02-features.md
│   ├── 03-user-flows.md
│   ├── 04-ui-design.md
│   ├── 05-data-model.md
│   ├── 06-integrations.md
│   ├── 07-business-logic.md
│   ├── 08-non-functional.md       ← populated via interactive Q&A
│   ├── 09-decisions.md            ← ADRs mined from chat
│   ├── 10-evolution.md            ← pivots + abandoned attempts
│   ├── 11-production-adjustments  ← POC shortcuts flagged for rework
│   └── implementation-hints/      ← Lovable-specific notes (optional)
├── .planning/                     GSD-ready
│   ├── PROJECT.md
│   ├── REQUIREMENTS.md
│   ├── ROADMAP.md
│   ├── BACKLOG.md                 ← items deferred past v1
│   └── phases/NN-slug/PLAN.md     only when project warrants breakdown
└── REVERSE-ENGINEER-REPORT.md     coverage matrix, confidence, TBDs

Install

Claude Code

# option A: clone into the skills dir
git clone https://github.com/brendangooden/lovable-reverse-engineer-spec \
  ~/.claude/skills/lovable-reverse-engineer-spec

# option B: symlink from anywhere
ln -s /path/to/lovable-reverse-engineer-spec ~/.claude/skills/lovable-reverse-engineer-spec

Then invoke in any Claude Code session:

"Reverse-engineer the Lovable project at C:\repos\my-lovable-app using the chat export at C:\chat-history\my-lovable-app."

Codex / Cursor / Aider / Cline

Copy AGENTS.md and prompts/ into the root of your working project (or any parent your agent loads AGENTS.md from).

GitHub Copilot

Copy this repo's .github/copilot-instructions.md + AGENTS.md + prompts/ into your working project.

Usage

1. Clone the Lovable repo locally.
2. Export chat history with loveable-chat-history-capture (bun run export).
3. Ask your agent to run this skill with the two local paths.
4. Answer the NFR elicitation questions (audience size, compliance, SLA, etc.).
5. Answer the POC-vs-production review (which shortcuts to carry forward vs adjust vs drop — the skill also auto-surfaces concerning patterns it found).
6. Review REVERSE-ENGINEER-REPORT.md for coverage and gaps.
7. Feed .planning/ROADMAP.md into GSD (/gsd-new-project --from-roadmap) or hand /spec/ to any agent to rebuild.

Design principles

• Evidence-linked — every claim cites a chat id or path:line. No hallucinated features.
• Stack-free spec — /spec/ never names React, Supabase, shadcn. Those move to implementation-hints/.
• Chat as primary intent source — code shows what was built; chat shows what was meant. Conflicts are surfaced, not papered over.
• Preserve failed attempts — abandoned paths are ADR context, not noise.
• POC ≠ production — a Lovable prototype is ground truth for what was built, not what should ship. The skill asks up-front what to keep vs adjust vs drop and auto-surfaces concerning patterns (CORS wildcards, PII in logs, missing FKs, etc.) for per-item decisions.
• Three signal sources — (1) the repo shows what was built, (2) Lovable chat shows what was meant at dev level, (3) Teams/Slack chat (optional) shows what was meant at stakeholder level. Merging all three catches the gaps a two-source review would miss.
• Never modify inputs — the repo and chat export are read-only.

What qualifies as a "Lovable project"?

At least two of:

• lovable-tagger in devDependencies
• componentTagger() called in vite.config.ts
• .lovable/ directory at repo root
• Auto-generated header in src/integrations/supabase/client.ts

If fewer than two match, the skill aborts. It's for Lovable specifically, not "any React app".

Related

• loveable-chat-history-capture — required upstream tool
• GSD (Get Stuff Done) — spec-driven framework this tool's .planning/ output targets
• GitHub Spec-Kit — alternative destination for /spec/ content
• addyosmani/agent-skills — skill-format inspiration

License

MIT — see LICENSE.