Codex Cloud Prompts

Lean prompt set for using Codex Cloud on existing repositories. Put the universal AGENTS.md from this archive at the repository root, then run one task-specific prompt on top of it in Codex Cloud.

Design principle

AGENTS.md holds durable operating rules. Task prompts add only the goal, domain lens, scope boundary, evidence requirements, and expected output for one run. They intentionally avoid repeating the full agent contract and avoid handmade task-sub body schemas that can block Codex Cloud’s native task-sub flow.

Prompt modes

Mode	What it does	Product code
Plan / discovery	Inspect evidence, produce maps, audits, task-subs, synthesis, validation, or roadmap artifacts.	Not changed
Execute	Make one scoped diff or one coherent implementation slice, then report changes and validation.	Changed when justified
Documentation	Generate or refresh docs from repo evidence.	Docs only
Post-execute review	Review the current patch and fix only patch-introduced or patch-exposed issues.	Changed only if patch fix is required

Prompt index

Task planning

Prompt	Use when
Task Shaping.md	Turn a user request into exactly one native task-sub.
Validation & Clarification.md	Verify, rewrite, clarify, or block one task-sub before execution.
Synthesis.md	Merge multiple audit outputs into a deduplicated native task-sub queue.
Evolution & Improvement.md	Build a 1–3 iteration improvement roadmap in evolution_plan.md.

Audits

Prompt	Use when
Check & Fix Plan.md	General behavior audit across domain/API/UI/DB/tests/configs.
Security Audit.md	Reachable security and abuse issues.
API Audit.md	API contract, validation, auth, schema, webhook, and client drift.
SaaS Audit.md	SaaS UX/RBAC/plan/tenant/state/product behavior.
Bot Audit.md	Bot commands, FSM, callbacks, menus, copy, permissions, platform limits.
Test Coverage Audit.md	Critical missing, false, stale, or fragile tests.
Performance Audit.md	Reachable UX/cost/reliability/hot-path performance issues.

Navigation maps

Prompt	Artifact
API Navigation Map.md	docs/api_map.md
SaaS Navigation Map.md	docs/nav_map.md for SaaS routes/surfaces/roles/plans.
Bot Navigation Map.md	docs/nav_map.md for bot states/transitions/commands.

Execute and polish

Prompt	Use when
Bot Polish.md	Polish bot flow/copy/navigation/state/platform behavior.
SaaS Polish.md	Polish SaaS flow/copy/RBAC/plan/tenant states.
Web Polish.md	General web UX/copy/a11y/state polish.
Landing Copy Polish.md	Landing/marketing copy and conversion UX.
Localization Polish.md	i18n/l10n/catalog/locale formatting polish.
Migration & Refactor.md	One coherent migration/refactor slice.
Autodeploy & Infra.md	Minimal CI/CD + Docker Compose + host Nginx setup.

Docs and review

Prompt	Use when
README GEN.md	Generate/update project README from repo evidence.
DESIGN GEN.md	Generate/update DESIGN.md from UI/design evidence.
Patch Checking.md	Review and fix the current patch only.

Recommended flows

• Product task: Task Shaping.md → Validation & Clarification.md → matching execute prompt → Patch Checking.md.
• Unknown UI/product repo: navigation map (Bot or SaaS) → Check & Fix Plan.md + domain audit → Synthesis.md.
• Backend/API: API Navigation Map.md → API Audit.md; add Security Audit.md for sensitive surfaces and merge with Synthesis.md.
• Large improvement path: Evolution & Improvement.md → split roadmap items through Task Shaping.md or Synthesis.md → execute prompts.
• Docs/deploy bootstrap: README GEN.md, DESIGN GEN.md, or Autodeploy & Infra.md.

Task-sub discipline

One task-sub should fit one execute run / one PR. It needs behavior problem/change, exact path:line[-line] evidence, owner/source of truth, fix direction, scope, acceptance criteria, and validation. Prompts should ask Codex Cloud to create task-subs, not force a custom markdown body that replaces the native mechanism.

Priority shorthand:

• P0 — correctness, security, data, tenant, payment, destructive action, or public/internal contract issue with high impact.
• P1 — reachable UX, reliability, architecture, test, or performance issue with meaningful impact.
• P2 — lower-risk evidence-backed issue that is still useful to fix.

AGENTS.md discipline

AGENTS.md is the durable execution contract. Keep high-leverage rules there: instruction priority, repository grounding, external facts/dependencies, execution sequence, checks/honesty, DoD, markdown defaults, and native task-sub policy. Task prompts should add task-specific goal/scope/domain detail instead of repeating the whole contract.

Evidence discipline

path:line[-line] anchors are mandatory for audit findings, task-subs, validation verdicts, navigation maps, evolution plans, and repo-derived docs. Execute prompts use anchors when explaining non-obvious repo facts, skipped adjacent work, or detected behavior; the diff itself remains the primary implementation artifact.

Markdown discipline

Generated docs and plans should use one H1, a clear H2/H3 hierarchy, compact paragraphs, tables only for dense comparisons, and no placeholder-heavy fenced templates. Preserve stable IDs in maps/plans. Omit empty sections instead of filling a template.

Output style

Plan prompts may require strict P0/P1/P2 grouping. Execute/review/doc prompts use compact Russian emoji sections where useful:

• ✅ result / changed scope
• 📁 key files
• 🔎 evidence anchors when relevant
• 🧪 validation
• 👀 manual review or trace
• ⚠️ assumptions / risks
• ⛔ not run / not included

Usage notes

• Run one prompt per Codex Cloud session unless the task explicitly asks for synthesis of multiple outputs.
• Generated maps are discovery aids, not authority; audits must spot-check current repo evidence.
• Execute prompts may change multiple related files/states when needed for a complete coherent result.
• If a plan/doc prompt cannot write an artifact in the current mode, it should print the complete markdown content in chat.