Skip to content

AGENTS.md

Latest commit

 

History

History
92 lines (71 loc) · 3.29 KB

AGENTS.md

File metadata and controls

92 lines (71 loc) · 3.29 KB

AGENTS.md

Purpose

Build and maintain this ChatGPT app with the lowest practical friction and high reliability.

Deployment Intent

  • Intent: PRIVATE_SELF_USE
  • Profile: Private app for personal or internal use. Not intended for OpenAI submission.

Operating Rules

  • Start with the requested result.
  • Do not change Deployment Intent without Ramon approval.
  • Missing Deployment Intent, auth, payment, production, external write, persistent data, or destructive scope is a question for Ramon, not a technical failure.
  • Ask one question when a missing fact chooses between next actions.
  • Continue with an assumption only when it affects local uncommitted edits and failure will be visible in tests, logs, build output, or local manual checks.
  • Write the quality plan before implementation.
  • Keep app-value.md current before adding or changing tools.
  • Keep the SDK conformance matrix current while editing.
  • Record verification evidence as checks pass.
  • Ask one question only when ambiguity changes behavior, data, auth, payments, or production.
  • Read current repo instructions, scripts, tests, docs, and git state before editing.
  • Before editing, record scope, files to change, success output, and verification step.
  • Make the smallest change that solves the real problem.
  • Do not add speculative features.
  • Do not port a whole product UI; expose focused capabilities ChatGPT can orchestrate.
  • Do not change unrelated files.
  • Validate all external input with schemas.
  • Never commit secrets.
  • Ask before auth, payments, production, data deletion, or irreversible changes.
  • Keep business state, UI state, and cross-session state separated.
  • Keep presentation, domain, persistence, integration, workflow, and external service code in separate modules.
  • Treat shared contracts, auth, stored data, payments, production, migrations, and external integrations as protected surfaces.

Process By Stage

  1. Start: identify outcome, allowed changes, missing facts, and whether one question is required.
  2. Read: inspect instructions, file structure, scripts, tests, docs, logs, and git state.
  3. Plan: write scope, risks, contracts, and verification commands before editing.
  4. Edit: make the smallest reversible change tied to the requested result.
  5. UX: verify every touched visible state by use, test, story, screenshot, or local reproduction.
  6. Verify: run checks matching changed behavior and record evidence.
  7. Close: report scope, files changed, evidence, assumptions, untested items, risks, rollback, and next action.

Required Verification

Before calling work complete, run:

npm run verify

For UI changes, also inspect the visible widget states:

  • Initial.
  • Loading.
  • Empty.
  • Success.
  • Error.
  • Recovery.

Final Report

Include:

  • Scope.
  • Files changed.
  • Checks run and result.
  • Untested items.
  • Known risks.
  • Rollback path when applicable.

Process Gate

Work is not complete until these files are updated:

  • DEPLOYMENT_INTENT.md
  • QUALITY_PLAN.md
  • SDK_CONFORMANCE.md
  • PROCESS_EVIDENCE.md
  • quality-contract.md
  • sdk-fit.md
  • tool-contracts.md
  • risk-matrix.md
  • failure-model.md
  • verification-report.md

If any required check is skipped, mark it as Not verified with the reason.

If a skipped check depends on Ramon approval or a missing decision, ask for that decision before final review.