skeptic-review
Evidence-based production-readiness audit in 13 sections — every claim needs a file:line, N/A needs a reason, the verdict comes from the table, not the impression. Use only when explicitly asked — /skeptic-review, "production readiness review", "is this ready for production", "audit this checklist", "skeptiku audit", "kas see süsteem on production-valmis". Do NOT auto-trigger on ordinary feature or bugfix work. Pairs with production-skeleton — the skeleton builds the floor, this audit proves it holds.
hallikkristjan-code · 已有 AI 审核记录
安装与使用
运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。
npx skills add hallikkristjan-code/production-skeleton --skill skeptic-review技能说明
name: skeptic-review description: Evidence-based production-readiness audit in 13 sections — every claim needs a file:line, N/A needs a reason, the verdict comes from the table, not the impression. Use only when explicitly asked — /skeptic-review, "production readiness review", "is this ready for production", "audit this checklist", "skeptiku audit", "kas see süsteem on production-valmis". Do NOT auto-trigger on ordinary feature or bugfix work. Pairs with production-skeleton — the skeleton builds the floor, this audit proves it holds.
Skeptic Review — evidence-based production-readiness audit
Contents: Overview · Process (6 steps) · Report shape (+ "I can do this now" rules) · Checklist (13 sections) · Red flags · Typical mistakes
Overview
A 13-section skeptic's checklist, graded ONLY on evidence found in the code.
Core rule: a claim without a file:line is not a status, it is a guess.
The default posture is skeptical: a section is ❌ until evidence shows otherwise.
This audit covers breadth: all 13 areas in one pass. For per-path depth in the
security section, hand off to a dedicated security review (in Claude Code: /cso).
Reply in the language the user wrote in. Section names and statuses stay as written here so reports are comparable.
Process
- Map the project (type, stack, users, data). Decide for every section: applies or N/A. N/A needs a one-sentence reason in the report (e.g. "no payments → PCI DSS scope is zero"). Translate domain items: a "dropshipping partner" is every external integration (API, feed, webhook, agent).
- Collect evidence with 2–3 parallel read-only sub-agents on a cheap
model, split for example as security+auth+secrets / automation+monitoring+
tests / data+backup+privacy. Every fact as
file:line+ one sentence. Agents report MISSING findings too. Orchestration rule: each evidence agent writes its report to a FILE and the synthesizer reads the files — a background agent's return value does not always reach the synthesizer's context, and a synthesis that waits for it stalls. - Verify by running, not by reading the README: tests, type-check,
npm audit(or the ecosystem's equivalent), and the production-skeleton probe:node <skeleton-skill>/scripts/probe.mjs . --json(<skeleton-skill>is the production-skeleton skill directory; add--url <deployment>when a deployment is reachable). Its rows arefile:lineevidence for sections 1, 2, 3, 5, 6, 9, 10, 11 and 12 — cite them. Read ops documents (runbook, backup doc), but check every claim in them against code or a workflow. - Grade item by item: every checklist item gets ✅ / ⚠️ / ❌ / ❓ / N/A plus an evidence reference. An item without evidence cannot be ✅. ❓ means not verified: the evidence lives where you cannot look (a vendor dashboard, MFA settings, a live deploy you cannot reach) or a check could not run. Say what you searched and where; ❓ is not ❌ (absence was not shown) and never ✅. A probe PASS is evidence that a pattern exists, not that the section is ✅ — follow it to where it is wired in or runs.
- Report shape — see below, exactly in that order.
- Offer to save the report as
docs/ops/skeptic-review-YYYY-MM-DD.md(today's date). Ask; the audit itself is read-only.
Report shape (exactly this order)
- Top-3 risks — concrete, with evidence, impact described in the user's language, not in audit jargon.
- Section table — all 13 sections, N/A included (status + 1–2 sentences with evidence).
- Skeleton scorecard — the probe's score line and its failing rows, verbatim.
- Quick wins — fixes under about an hour, as a separate list.
- Not traced — 2–4 lines: the route families, tools or jobs you did not
open, and the checks you made by search only. End with one line: a per-path
security review (in Claude Code:
/cso) is the next pass for these. - Verdict. "Ready" is allowed only when no applicable section is ❌ or ❓ AND no required skeleton gate is FAIL, STALE or UNKNOWN. Otherwise the verdict names the blocking sections and gates; for ❓ and UNKNOWN it names what a human has to confirm.
- I can do this now: — 1–3 numbered lines (rules below). Last block.
"I can do this now" rules (last block of the report)
- 1–3 lines, only actions the agent can take ITSELF in this environment (write a test, add an alert, patch a dependency, write a runbook, fix a config) — not "talk to the host" or "decide the RPO".
- Pick from the top-3 risks and quick wins the ones that are DOABLE — the block is not a repeat of quick wins, it is their executable subset in the form verb + object + time estimate.
- Ordered by risk, not by ease — the first line addresses the worst finding the agent can touch. A ❌ finding goes before a convenience fix.
- Do not run them — the audit is read-only; wait for a number or "do 1 and 3".
- If every applicable finding needs a human decision (price, legal, access): "I can do this now: nothing — the next step is your decision on X." Padding the list with trivial edits is a mistake.
Checklist (13 sections)
- Architecture — stack? where does it run? where is the database? backups? disaster recovery? RPO and RTO AS NUMBERS (backup frequency → RPO; a tested restore procedure → RTO)?
- Identity and access — admin MFA? API-key management? where are secrets
(env / Vault / Secrets Manager; any leaks in the repo)? least privilege
(RLS, roles, scopes)? Then trace the seams:
- Every path: for each documented rule on data (a
draftstatus "hidden from users", a revoked token, a role, a time window, an OAuth scope), list every read and write path that touches that data — routes, MCP/API tools, service worker, exports — and cite where each enforces the rule. One path without it is ❌. - Side paths: where a guard exists (
requireAdmin, a ver
- Every path: for each documented rule on data (a