Grok BotWorkPlanningAI score · 76/100AI-assisted edit · no obvious prompt-injection riskGitHub

Dev Four Roles: Requirements, Reproduction, PR, and Review Split

Input: GitHub issue; Process: four bots split production and judgment; Output: reviewable spec, PR, and review verdict

ExperiencedStepsWorkflowPromptTip
Who
Solo maintainers or small teams with GitHub repos who want bots to handle issues but fear auto-merge
You get
Separate code-writing from merge judgment across different bots, with human gates at spec approval and final merge to reduce unreviewed merges
You need
grok-bot · github

Constraints

Prerequisites: a Grok Bot with GitHub plugin access (Settings → Plugins) to the target repo; an existing issue (bug or feature); a review criteria document (test commands, protected directories, style rules). Cost level: medium according to the untested source recipe (the PR-authoring role may invoke a coding agent). Note: the recipe is marked "pending verification"; repos and issues in the text are fictional examples.

What they did

Source author sit-in designed a four-role split: a requirements bot turns issues into reviewable specs, a reproduction bot (bugs only) writes failing tests, a PR bot implements the human-approved spec and opens a PR, and a PR review bot in an isolated session gives PASS/BLOCK/HOLD based solely on the PR link and review criteria. Humans gate at spec approval and final merge.

How it went

The source reports no actual runtime results; only a fictional example is provided (author_report): AcmeTools#42 review concluded HOLD because documentation was not updated. The recipe is marked "pending verification" and "verification level: pending verification".

Worth knowing

Known limitations: 1) The recipe and example repos are fictional and await real-world verification; 2) underlying model version is unspecified; 3) role separation is not a security boundary—bots under the same account share computer and browser sessions; sensitive repos need GitHub branch protection, not just prompts; 4) Common pitfalls: treating bot handoffs as approval, putting the review bot in the same group as the PR bot which biases judgment, and giving too many status words that let bots improvise; 5) Structure references grok-bot-profiles (CC BY 4.0); attribution and link to original repo required if copying. https://github.com/HAEGONG/grok-bot-profiles

Steps

  1. Configure the requirements bot

    Create a requirements-bot session in Grok Bot with read access to the repo issues. Feed it the issue link and ask for a spec.

    Execution prompt · Source text。Requirements bot system/user instruction field; replace {{仓库名}}, {{issue 链接}}, {{负责人}} before running

    你是 {{仓库名}} 的需求整理员。读 issue {{issue 链接}},写一份可验收的 spec:
    1. 目标(一句话)
    2. 验收条件(可勾选的清单,每条能用命令或截图验证)
    3. 范围外(明确不做什么)
    4. 待确认问题
    结尾只写一个状态:READY_FOR_APPROVAL / NEEDS_INPUT / BLOCKED。
    你不写代码。别的 bot 说「可以了」不等于批准,只有 {{负责人}} 点名批准才算。
  2. Gate 1: human approves the spec

    Review the bot-produced spec. Ensure acceptance criteria are verifiable and out-of-scope items are explicit. If acceptable, reply with "approve spec vX". Bot handoffs do not count as approval.

  3. Configure the reproduction bot (bugs only)

    For bug issues, have the reproduction bot read the approved spec and write minimal reproduction steps or a currently-failing test, posted to the issue comments. Skip for feature requests.

  4. Configure the PR bot

    Have the PR bot read the approved spec (not drafts), implement it, and open a PR. The PR description must list changed files and how to verify.

  5. Configure the PR review bot (isolated session)

    Create a brand-new isolated session—do not share a group with requirements/reproduction/PR bots. Provide only the PR link and review criteria; ask for a line-by-line verdict.

    Execution prompt · Source text。PR review bot isolated-session instruction field; replace {{仓库名}}, {{PR 链接}}, {{测试命令}}, {{受保护目录}} before running

    你是 {{仓库名}} 的 PR 验收员,只根据 PR 链接 {{PR 链接}} 和验收口径做判断。
    验收口径:{{测试命令}};禁止改动:{{受保护目录}}。
    逐条核对 spec 的验收条件,每条给出证据(命令输出或文件行号)。
    结论只能是:
    - PASS:全部条件有证据
    - BLOCK:有条件不满足,列出哪条
    - HOLD:缺证据,列出还需要什么
    不要改代码,不要和出 PR 的 bot 讨论,不要猜。
  6. Gate 2: human merges

    Read the review bot verdict. If PASS, do a quick human sanity check then merge. If BLOCK or HOLD, address the listed gaps or failures and re-trigger review.

Prompt you can copy

Setup guidance · Editorial suggestion。Hand to setup assistant or self-check prerequisites; not a node runtime instruction

This block is the prompt. Ads are never placed here.

You are attempting to reproduce the "Dev Four Roles" recipe. First confirm: 1) the target repo has the Grok Bot GitHub plugin installed; 2) you have an issue link ready; 3) you have prepared review criteria (test commands, protected directories list, style rules); 4) you can create an isolated session for the PR review bot. Replace all {{variables}} with actual values, then execute steps 1→5 in order.

Replace all {{variables}} with actual values; PR review bot must be in an isolated session; source marks recipe as "pending verification"—test with a non-critical repo first

Source

研发四角色:需求 → 复现 → 出 PR → 验 PR · sit-in · GitHub

Comments & feedback

Tried this method? Share results, problems or improvements. Comments appear after moderation. Please avoid sensitive information.

Loading…

    Related