review-recipe
The method for the review stage of a full-lane change: read the machine digest, judge the diff against the frozen spec, and record ordered findings. Use when you are the paqad review stage agent.
- License 许可
- MIT
- 所属仓库
- Eliyce/paqad-ai
- 最近更新
- 2026-11-13
- Stars
- ★ 8
- Forks
- 0
01 前言 Front Matter
- name
- review-recipe
- metadata
- { "cacheable": "false", "model_tier": "reasoning", "paqad-meta": "skill.meta.yaml", "output_format": "markdown" }
- description
- The method for the review stage of a full-lane change: read the machine digest, judge the diff against the frozen spec, and record ordered findings. Use when you are the paqad review stage agent.
02 资源文件 3
- 参考文档agents/openai.yaml166 B
- 脚本scripts/lint-findings.mjs3.6 KB
- 参考文档skill.meta.yaml148 B
03 SKILL.md 正文
What It Does
Gives the review stage one risk-first method and the shape its findings must take.
Use This When
You are the paqad stage agent for this stage. Your injected copy is the truth for this stage; the source skills it copies stay for their other workflows.
Inputs
- The frozen spec text and the review digest text from
paqad-ai stage begin review. - The diff of the change.
Procedure
- Restate the request and its acceptance criteria in two lines at the top of your return.
- Review the change against correctness, regressions, and rollback risk before treating it as complete.
- Confirm or contest each machine finding in the digest. Cite the digest rows in your findings, by file or by rule id.
- Judge the diff in this order: correctness and safety, rollback and migration, contract and compatibility, verification gaps, documentation drift. Mention style only when it raises risk.
- Rank findings from worst to least. A blocker comes first, then major, then minor. Each names its file and what must change.
- Fill the review template under
.paqad/tmp/and runpaqad-ai review record <review-template.json>. It runs thelint-findings.mjscheck itself and refuses a template that fails, printing the errors. It also refuses findings that are all judgment when the digest has a machine row.
Read This Path
- For the full review method and its output template, read
base/skills/adversarial-review/SKILL.mdunder the install directory that.paqad/framework-path.txtnames.
Output Contract
review.jsonfrompaqad-ai review record, with findings in rank order, each citing a file or a digest row.
Escalate / Stop Conditions
- A finding that implies unsafe rollout, data loss or broken customer behaviour: verdict blocked, and say so.
04 同仓库更多 Skills
sample-skill
Sample bundled skill fixture
processing-invoices
Turns a supplier invoice into one ledger line. Use when a supplier invoice arrives and needs booking.
processing-invoices
Turns a supplier invoice into one ledger line. Use when a supplier invoice arrives and needs booking.
understand-idea
Chat and Claude Code: Step 1 of the discover journey: clarifies the goal, who benefits, success and constraints into the idea block. Use when the journey directs it to step 1, not directly.
size-work
Chat and Claude Code: Suggest a size for a ticket from comparable finished work, as a labelled suggestion the team still owns. Use it when someone asks "how big is this?", "roughly how much work is ABC-12?", or wants a backlog item sized. It finds the team's sizing scale at first use, picks two to five of the closest finished tickets with links, and gives a size or range with the drivers behind it and a confidence of low, medium or high, lowered and explained when the area's documents are template text. Every result is labelled a suggested size, never a date or a commitment. It writes nothing into the size field unless the user says "set it", and it refuses to overwrite an existing size unless the user asks to replace it. Produces the sizing block of the record.
refine-ticket
Chat and Claude Code: Improve one or more existing tracker tickets without losing the author's intent. Use it when someone wants a ticket read, diagnosed and tidied up ("this ticket is vague", "fix up ABC-12", "split this, it's too big"). It reads each ticket in full and snapshots it with its last-updated time, says what is wrong with evidence, and proposes a field-by-field before and after change set limited to summary, description, labels, parent, links and new child tickets on a split. It writes nothing until an explicit "apply it" bound to that exact change set, re-checks the ticket for conflicting edits just before writing, never replaces rich descriptions with plain text, never deletes or closes a ticket, and reads each ticket back after applying. Produces the existing and changes blocks of the record.