Skip to content

agent.md

Latest commit

 

History

History
93 lines (64 loc) · 4.13 KB

agent.md

File metadata and controls

93 lines (64 loc) · 4.13 KB
Raw

Pre-Deployment Checklist — AI Agent Version

Machine-facing companion to the checklists in checklists/. The base checklist (checklists/base.md) is the source of truth for what to verify; stack-specific files (e.g. checklists/go.md) add extra items. This file tells an AI coding agent (Claude Code, Cursor, Copilot, etc.) how to run them before a deploy.

Point your agent at the base checklist plus any stack file that matches the project. The checklists provide the items; this file provides the rules.

1. Copy-paste prompt

Paste this into your agent, in the repo you are about to deploy:

Fetch and read the BASE deployment checklist:
https://raw.githubusercontent.com/PinoyFreeCoder/deployment-checklist/main/checklists/base.md

Then, if a checklist exists for this project's main language/stack, ALSO fetch it
(e.g. .../checklists/go.md, .../checklists/java.md) and treat its items as
additional checks on top of the base. If none exists, use the base alone.
(Or, if the checklists were copied into this repo, read those local copies instead.)

Audit THIS repository against every checklist item that applies to it:
- Web-only repo: skip the Mobile-Specific section.
- Mobile-only repo: skip the Web-Specific section.
- Full-stack / both: check everything.
- Include every item from the matching stack checklist, if any.

For EACH item:
- Verify against the actual code, config, and dependencies. Do NOT assume.
- Cite the evidence: a file path + line, a config value, or command output.
- Mark: PASS / FAIL / N/A. N/A must include a one-line reason.

Then output, in this order:
1. Summary counts: PASS / FAIL / N/A totals.
2. A FAIL table: item, section, severity ([BLOCKER]/[SHOULD]/[NICE]), evidence, exact fix.
3. Verdict: GO or NO-GO. NO-GO if ANY [BLOCKER] is FAIL.

Rules:
- Report only. Do NOT deploy. Do NOT run destructive commands.
- Do NOT fix anything yet — list the fixes and wait for my approval.

2. Output contract

The agent's report must contain these three parts, in order.

2.1 Summary counts

PASS: <n>   FAIL: <n>   N/A: <n>   (total <n> items checked)

2.2 FAIL table

Item Section Severity Evidence Fix
(short item text) (e.g. Web / Security 4.2) [BLOCKER] (file:line or config or command output) (exact change)

Only FAIL rows appear here. If there are zero FAILs, write No failures.

2.3 Verdict

VERDICT: GO

or

VERDICT: NO-GO — <n> blocker(s) unresolved

Rule: NO-GO if any [BLOCKER] item is FAIL. A single unresolved blocker blocks the whole deploy, regardless of how many other items pass.

3. Verify-not-assume rule

This is the rule that makes the audit trustworthy.

  • Evidence or it's a FAIL. An item is PASS only with proof: a file path + line, a concrete config value, or the output of a command the agent ran. "Looks fine", "probably set", or "the framework handles this" is not evidence — mark it FAIL or investigate further.
  • Check the real artifact, not the intent. For "no secrets in repo", the agent greps the history, not just the current tree. For "TLS only", it inspects the actual config, not a comment claiming it.
  • N/A must be justified. Example: PCI DSS — N/A: no card data handled, payments delegated to Stripe Checkout (see app/checkout/route.ts).
  • Report first, fix second. The agent reports findings and proposed fixes. A human approves before any fix is applied and before any deploy. The agent must never deploy or run destructive commands on its own.

4. Notes

  • Keep this file and the checklists in checklists/ in sync. When you add a checklist item (base or stack), the agent picks it up automatically — no change needed here unless the output format changes.
  • Severity labels ([BLOCKER] / [SHOULD] / [NICE]) are defined in the base checklist's legend and reused by every stack file.
  • Adding a stack checklist? See CONTRIBUTING.md.
  • The human Sign-off block still applies: a person records the final GO/NO-GO, even when an agent produced the audit.