gitreverse

GitReverse — repository to prompt

Reverse a public GitHub repository into one short, conversational user prompt grounded in real repo context. This ports the GitReverse flow (https://github.com/filiksyos/gitreverse) into an OMK skill: pull repo metadata, a depth-1 root file tree, and the README, then synthesize the single prompt a person could paste into an AI coding agent to rebuild the project from scratch.

Core workflow

1. Parse the target: accept a full GitHub URL (https://github.com/owner/repo, including /tree/... paths — use the whole repo for now) or an owner/repo slug. GitHub-style tree subpaths are NOT scoped yet; the reverse flow uses the whole repo.
2. Verify visibility: only public repositories. If the repo 404s, is private, or requires auth to read, stop and report it as a blocker.
3. Fetch exactly three evidence surfaces:
   • repo metadata (owner, name, description, primary language, topics)
   • root file tree at depth 1 (top-level files and directories)
   • README (rendered to text)
4. Synthesize one short, conversational prompt grounded in that context. It describes what to build, not how to copy it — no file dumps, no verbatim README paste, no system prompts.
5. Cite evidence: name which facts came from metadata vs. file tree vs. README, and keep observed facts clearly separated from synthesized text.

Tool selection

• Tools: gh or git for local checks; plain HTTPS fetch against api.github.com is enough (metadata, git/trees, readmes endpoints)
• MCP: github, filesystem
• Hooks: protect-secrets, stop-verify

Safety constraints

• Public repositories only. Never fetch, mirror, or echo private repo contents even if a token is available.
• If a README or file listing exposes secrets (tokens, keys, credentials), do not copy them into the prompt; redact and note the finding.
• The output is a user prompt, not documentation or a system prompt.
• Rate limits: prefer unauthenticated GitHub API and few requests; set a token only if the user provides one for higher limits.

Acceptance

• Repository owner/name is parsed from the URL or slug before any fetch
• The generated prompt is grounded in README and file-tree evidence
• Only public repository data is used; private or credential-bearing content is refused and redacted
• Observed repo facts and synthesized prompt text are clearly separated
• Output is exactly one conversational prompt plus an evidence section