GitHire

Humans frame. AI executes. Architects judge.

English · 中文

Website · Case 01 · Blog · Skill · Open an issue

---

GitHire is an AI-native engineering methodology, written as a reusable six-step workflow and validated against real production incidents.

AI coding speed already outruns human review speed. Reviewing code after it's written is too late, so the work shifts:

• Humans frame — write the issue, set constraints, declare non-goals, define success.
• AI executes — generates the code, runs tests, opens the PR.
• Architects judge — make the final tradeoff before merge.

The six steps below place each role at the right point in the pipeline:

1. Issue — frame the problem with the six-section Prompt Spec (Goal / Constraints / Non-goals / Verification / Architecture notes / Existing context).
2. Sandbox — a long-running dev environment with real dependencies and production-shape data.
3. Execute — Claude Code or Codex implements in the sandbox.
4. AI Review — a second agent with different priors (performance, security, anti-patterns).
5. Architect ★ — the most expensive 30 seconds in the workflow. Human reads the diff with system-side context (QPS curves, prior incidents).
6. Production — merge, ship, write the decision trail back to the Issue.

If you only adopt one step, adopt step 5.

Real cases

The site is an operating log, not a tutorial. Every case is a real event:

• Case 01 · A 22-line SCAN took down production Redis — Codex wrote a r.scan(match='model_detail::*') + pipelined HGETALL in 5 minutes; the architect step was skipped; production was on fire 13 hours later; 25 hours of fix-chain commits until a Redis SET replaced the SCAN. Walked along all six steps to show where each one was abandoned.

More cases are queued; if you've shipped a real incident worth walking, open an issue and we'll co-write the post-mortem.

Install as Skill

GitHire is published as an installable Skill so any AI agent can apply the same method. The skill lives at realRoc/skills alongside the matching Prompt Spec skill.

# Install both skills
npx skills add realRoc/skills

# Or just one
npx skills add realRoc/skills --skill githire
npx skills add realRoc/skills --skill prompt-spec

The Skill points agents at the live site for current method, plus the bundled references/method.md for offline fallback. Detailed install page: https://realroc.github.io/git-hired/skill.html.

Note on path migration: skills used to live under realRoc/git-hired/skills/githire. They now live at realRoc/skills/skills/githire. The local skills/ directory in this repo is a dev workspace artifact (symlinked to .agents/) and is gitignored — clones will see it empty, which is expected.

What belongs here

Open an issue if you want to:

• contribute a real-incident case for the operating log;
• propose a refinement to the six-step workflow;
• challenge a claim in the FAQ;
• improve the Prompt Spec template;
• discuss how AI should support — not replace — architect judgment.

A polished proposal is not required. A clear question is enough.

Links

• Website: https://realroc.github.io/git-hired/
• Case 01: https://realroc.github.io/git-hired/case-redis-scan.html
• Blog: https://realroc.github.io/git-hired/blog.html
• Skill landing: https://realroc.github.io/git-hired/skill.html
• Skill source (canonical): https://github.com/realRoc/skills
• Issues: https://github.com/realRoc/git-hired/issues