CtrlK
BlogDocsLog inGet started
Tessl Logo
SkillPublic

aet-req-review

Gate pipeline for requirement deliverables — enforces automated quality gates and interactive user validation before handover. Trigger when: (1) structured PRD/spec/SDD reviews are required, (2) executing a Stage demanding review-then-revision cycles, (3) workflows necessitate HCritic automated checks coupled with user-in-the-loop revision.

Quality

79%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide
SecuritybySnyk

Passed

No findings from the security scan

Fix and improve this skill with Tessl

tessl review fix ./jiuwenclaw/jiuwenswarm/agents/harness/code/rails/sdd/design_rail/skills/aet-req-review/SKILL.md
SKILL.md
Quality
Evals
Security

Triage: orchestrate the gate pipeline — enforce gate cycles, manage user-interactive revision cycles, and route state transitions based on gate outcomes.

  • The pipeline comprises: a SubAgent gate loop → a user-interactive revision step → a conditional re-gate loop.
  • Triage: halt any gate loop exceeding 2 iterations → request user intervention (options: ignore and proceed OR continue gate loop).
  • Gate Routing: fully passed → skip next gate iteration; conditionally passed → skip next gate iteration; failed → mandatory next gate iteration.
  • Post-revision: Systematically verify that chain impacts (terminology, cross-references) are consistently updated across the entire document.

[R1] SubAgent Gate

  • Validate input completeness by confirming the deliverable path exists. The methodology frame above carries the resolved absolute path and an Exists: flag — trust it and hand that absolute path to the Subagent. Do NOT shell out to test -f / [[ -f ]] to check it: those are not portable to Windows PowerShell, and the failing probe aborts the gate.
  • If the frame reports Exists: no, stop and ask the user for the correct path — do not start the gate against a missing deliverable.
  • DO NOT read the content yourself.
  • Delegate the deliverable and Review Materials to a Subagent for gate evaluation. Use the subagent call the system prompt prescribes: call subagent_spawn with a subagent_type from the system's "Available subagent types" list, then subagent_wait in the same turn. Use the synchronous task_tool only when the system prompt says the registered subagent tool is task_tool — task_tool accepts only the synchronous subagent types, and explore_agent is not one of them, so calling task_tool with it is rejected before the gate starts. (Reference the prompt template in ).
  • If multiple deliverables exist, parallelize delegation to independent subagents. Final routing decisions MUST follow the worst-case result principle.

[R2] Deliverable Revision (via Subagent)

  • CRITICAL: Do NOT read the deliverable content yourself — delegate the revision to a Subagent to avoid context explosion from large file reads.
  • Delegate to a Subagent (use task_tool with an available subagent type), providing:
    • The gate results from R1 (issue list with severity and location)
    • The deliverable file absolute path
    • Instructions: read the deliverable, address each gate issue sequentially, apply fixes via write_file (in-place overwrite), and return a brief revision summary.
  • Diagnostic (pass to Subagent): never apply superficial text patches — trace back to the deliverable's generation methodology, identify impacted steps, and re-execute from scratch if necessary. Fix the source material, not just the document surface.
  • The Subagent must ensure document-wide consistency for all chain impacts (terminology, cross-references) and NEVER introduce new non-conformities.
  • After the Subagent returns, use its revision summary for R3 — do NOT read the deliverable yourself.

[R3] User Gate & Revision

  • CRITICAL: Before calling ask_user, you MUST output the complete review results as visible content (not just in reasoning). The output MUST include:

    • Gate grade (fully passed / conditionally passed / failed)
    • List of issues found (with severity and location)
    • Revisions applied (what was fixed)
    • Final score if applicable
  • After outputting the review results as content, call ask_user with a structured question — the confirmation must be a choice, never a free-text box. A free-text confirmation lets the user "confirm" with an empty input, which the transport cannot carry, and the gate then re-asks the same question forever.

    {
      "query": "文档审查完成,是否需要修改?",
      "questions": [{
        "question": "是否需要修改?如有其他意见请说明,无需修改请直接确认。",
        "header": "审查结论确认",
        "options": [
          {"label": "无需修改,进入下一阶段"},
          {"label": "需要修改(请补充具体意见)"},
          {"label": "Other"}
        ]
      }]
    }
  • The Other option is the frontend's free-text slot (rendered as 「其他(自定义输入)」): picking it opens an input box for the concrete modification request. Keep the literal label Other.

  • NEVER merge the review summary into the question text: the summary is visible content, the question carries only the confirmation. Keep the option labels free of the words 返工 / 拒绝 / rework — the rail treats them as a rework request.

  • If the user requests modifications:

    • Apply the requested revisions to the deliverable in-place.
    • Ensure document-wide consistency for all chain impacts.
  • If the user approves (no modifications needed):

    • Proceed directly to [R4].

Empty Submission / Skip / Cancel Semantics

These three user actions all mean "no modification requested". Treat them identically and go straight to [R4]:

User actionAnswer text you receiveAction
Selects "无需修改,进入下一阶段"the option labelProceed to [R4]
Selects "Other" and types the modification requestthe typed textApply the revision, then follow the User Gate below
Clicks 「跳过」 (skip)用户已跳过此题,未选择任何选项。Proceed to [R4]
Clicks 「取消」 (cancel)用户已取消本次问答,未作答。Proceed to [R4]
  • NEVER call ask_user again after an empty submission, a skip or a cancel: re-asking after an empty submission stalls the flow — the user has already signalled "no modifications".
  • Treat the answer as modification feedback ONLY when the user picked "需要修改(请补充具体意见)" or typed a request into the "Other" box. If they picked the modification option without any detail, ask once for the details; if that follow-up comes back empty, skipped or cancelled, treat it as no modification and go to [R4].

[R4] End & Advance

  • Output a brief review summary: grade, issue count, fix count.
  • Output a clear transition message telling the user what stage is next:

    "文档审查完成。审查结果:[grade],发现问题 [X] 个,已修复 [Y] 个。现在进入下一阶段:[下一阶段名称]。"

    • The next stage name is specified in the methodology frame above — use it in the message.
  • CRITICAL: After outputting the transition message, call the sdd_advance tool to proceed to the next stage. Do NOT guess or skip.
  • Halt all further gate-pipeline actions after calling sdd_advance.

SubAgent Delegation Prompt Template (Used in [R1])

You are an authoritative Quality Assurance Reviewer. Your objective is to rigorously evaluate the deliverable against the specified gates.

## Input
- Review Materials: <path_to_review_materials>
- Deliverable: <path_to_deliverable>

## Task
1. Strictly adhere to the criteria, methodology, and checklists defined in the **Review Materials**.
2. Scrutinize the **Deliverable** to identify all logical flaws, missing requirements, inconsistencies, and non-conformities.
3. Output a definitive gate grade. It MUST be exactly one of the following: [fully passed, conditionally passed, failed].
4. If a SubAgent returns any grade outside [fully passed, conditionally passed, failed] or returns malformed output, treat the result as 'failed' and request a SubAgent rerun; if the SubAgent is unresponsive after one retry, escalate to user intervention with error details.
5. Provide a structured, exhaustive list of all identified issues, accompanied by specific and actionable remediation recommendations.
  • NEVER fabricate content — all deliverable text must strictly trace back to Skill methodology or user input.
  • Never re-ask: an empty, skipped or cancelled answer IS an answer — it means "no modifications". Re-asking the same question deadlocks the gate and the flow never reaches the next stage.
  • Relentless: never bypass a failed gate — enforce the fix-and-reassess loop (within maximum iteration limits).
  • Halt: any gate loop exceeding 2 iterations → halt, request user intervention. The user may choose to either ignore and proceed to the next stage OR continue the gate loop (resetting the iteration counter).
  • ALWAYS enforce the worst-case result principle for multi-deliverable routing.
  • ERROR RECOVERY: If a Skill or SubAgent crashes after processing begins, capture partial results, set state='error', notify the user with the error log, and offer choices: retry (maximum one automatic retry), skip-to-end, or abort. Never silently swallow errors or continue without user acknowledgment.
  • Self-contained: Do NOT attempt to load any external skills via skill toolkit — the methodology is inlined. Use the built-in ask_user tool for all user interactions.
  • Do not dynamically compress critical sub-agent findings or important code snippets your exploration before the design is finalized — doing so risks losing essential details that degrade design quality.
  • Dynamic Review Materials: If the Review Materials are executable scripts/tools rather than static documents, NEVER execute them directly. Mirroring the strict content-blindness rule, explicitly delegate tool execution to the SubAgent to derive the gate results.
  1. Deliverable Path to be Reviewed

    • If the path is missing -> abort with fatal error. The gate pipeline cannot proceed without a deliverable.
  2. Review Materials Path

    • Encompasses reviewer role and checklist (may be unified into a single document).
    • If the path is missing -> prompt the user: "No Review Materials found. Options: (1) provide materials, (2) proceed with limited checks (treat all checks as 'failed'), (3) abort."
  • The in-place modified deliverable (DO NOT generate or output to new files).

Initial Gate Loop

  • IF R1 gate PASS (fully OR conditionally) THEN execute R2 revisions -> proceed directly to R3.
  • IF R1 gate FAIL AND R1 iteration count < 2 THEN execute R2 revisions -> return to R1 for re-gate.
  • IF R1 gate FAIL AND R1 iteration count >= 2 THEN halt -> request user intervention (options: ignore and proceed to R3 OR continue gate loop with reset iteration counter).

User Gate

  • IF R3 user makes ZERO modifications — including an empty confirmation, a 「跳过」 (skip) or a 「取消」 (cancel) — THEN proceed directly to R4. NEVER re-ask the same question.
  • IF R3 user MAKES modifications THEN optionally re-gate (ask user: "是否需要复审?" with options ["需要复审", "跳过复审"]).
    • IF "需要复审" THEN return to R1 for re-gate.
    • IF "跳过复审" THEN proceed to R4.
Repository

XiaoLuoLYG/GOD

Path

jiuwenclaw/jiuwenswarm/agents/harness/code/rails/sdd/design_rail/skills/aet-req-review/SKILL.md

Last updated
First committed

Is this your skill?

If you maintain this skill, you can claim it as your own. Once claimed, you can manage eval scenarios, bundle related skills, attach documentation or rules, and ensure cross-agent compatibility.