Prompt Architect

Structured Codex prompt patterns, approval-gated workflows, and safety checklists for reviewable AI-assisted development.

🌐 Languages: | | | | |

English welcome Prompt Architect packages source-backed Codex prompt patterns, approval-gated execution modes, safety checklists, and reviewable AI-assisted development workflows. Start here: Usage modes explains the operating model.

Türkçe karşılama Codex'e iş verirken kapsamı, onayı, güvenliği ve doğrulamayı en başta netleştirmek için kullandığım prompt mimarisi seti. Amaç, işi büyütmeden önce niyeti, sınırı ve kontrol adımlarını yazıya dökmek. Buradan başla: Türkçe anlatım için README.tr.md; modları görmek için kullanım dokümanı.

• Status: v1.1.1 public release
• License: MIT
• Project type: Markdown-based Codex skill package and Prompt Lab knowledge base
• Note: Independent community project. Not affiliated with, endorsed by, or sponsored by OpenAI.

Prompt Architect helps turn vague AI coding requests into scoped, safer, and verifiable Codex prompts. It is built for teams and maintainers who want Codex to inspect first, plan clearly, wait for approval on risky work, and report evidence instead of just saying "done".

Operator Command Center

Signal	Use it when	Start here
Prompt architecture	You need a ready-to-paste Codex prompt with scope, verification and stop rules.	Skill entrypoint
Plan-first control	The repo is broad, risky, security-sensitive or release-adjacent.	PLAN MODE ONLY
Approved execution	A plan has been reviewed and implementation can start.	APPROVED - EXECUTE
Source-backed work	Current docs, MCP behavior, security guidance or repo claims matter.	Source maintenance
Release confidence	You need local checks before commit, push or release.	Validation
Specialist routing	A broad task needs code mapping, docs research, security or release review.	Subagents

Enterprise Evaluator Path

If you need to prove...	Start with	Evidence you get
The repository is public-safe	Public repo checklist	Secret, local-path, private-prompt and generated-artifact guardrails.
The prompt workflow has human control	Usage modes	PLAN MODE ONLY, APPROVED - EXECUTE, STOP / RECOVER and PROMPT_ONLY paths.
The skill can be copied into a Codex workspace	Install guide	File layout, copy path and setup expectations.
The security boundary is explicit	Security model	No-secret, disclosure, prompt privacy and verification rules.
The examples are usable	Examples	Copy-ready prompts with scope, verification and stop conditions.

Operating Guarantees

Signal	Standard
Skill-first routing	Prompts are shaped around reusable skill instructions instead of one-off loose chat.
Approval gates	Risky execution is separated from planning and requires an explicit approval step.
Verification pressure	Generated prompts ask for tests, scans, browser QA or an honest unverified report.
Public-safe knowledge base	Research notes, templates and outputs avoid secrets, private prompts, customer data and local operator paths.

Why This Exists

AI coding agents are useful, but loose prompts create loose outcomes. A good Codex workflow should state the goal, context, constraints, non-goals, verification, output format, and stop conditions before the agent starts changing files.

This repository packages those patterns as:

• a reusable Codex skill,
• prompt templates,
• workflow playbooks,
• security and browser-QA guidance,
• prompt evaluation rubrics,
• source-backed notes for AI coding-agent workflows.

Start Fast

I want to...	Use this
Generate a safe Codex prompt	Skill entrypoint
Keep Codex from editing too early	no-Goal PLAN MODE ONLY
Approve a scoped implementation	APPROVED - EXECUTE
Stop an agent that left scope	STOP / RECOVER
Check public repo safety	Public repo checklist
Review leak-prevention rules	Security model

What You Get

Capability	What it gives you
Goal + Full Prompt	A short objective plus a complete first-message prompt.
no-Goal PLAN MODE ONLY	Strict planning control before any file edits.
APPROVED - EXECUTE	A bounded execution prompt after human review.
STOP / RECOVER	A recovery prompt for premature execution or scope drift.
Browser QA	UI verification instructions for real user flows.
Security constraints	No-secret, approval-gated rules for sensitive work.
Prompt rubric	A way to score prompt quality before use.

Who This Is For

• Developers using Codex for real repository work.
• Prompt engineers building reusable agent instructions.
• Maintainers who want safer public documentation and prompt examples.
• Product, UI, security, and platform teams that need approval-gated AI coding workflows.
• Turkish and English users who want clear Codex workflow examples.

Quick Start

Copy the skill directory into your Codex project:

.codex/skills/prompt-architect/

Then ask Codex:

Use the prompt-architect skill to create a Codex prompt for: [your task]

For strict control, ask for a no-Goal plan-only prompt:

Create a no-Goal PLAN MODE ONLY Codex prompt for: [your task]. Do not allow execution before approval.

For prompt-only output:

Only output the final prompt. No commentary. Create a Codex prompt for: [your task]

Navigation

Goal	Go to
Install or copy the skill	docs/INSTALL.md
Learn the main usage modes	docs/USAGE.md
Copy practical examples	docs/EXAMPLES.md
Understand the skill layout	docs/SKILL_STRUCTURE.md
Run local validation	docs/VALIDATION.md
Maintain source-backed notes	docs/SOURCE_MAINTENANCE.md
Route specialist agents safely	docs/SUBAGENTS.md
Decide skill vs plugin packaging	docs/PLUGIN_READINESS.md
Check public repo safety rules	docs/PUBLIC_REPO_CHECKLIST.md
Prepare GitHub metadata	docs/GITHUB_SETTINGS.md
Improve discoverability	docs/SEO.md
Review prompt/security boundaries	docs/SECURITY_MODEL.md
See planned improvements	docs/ROADMAP.md

Trust Signals

Area	Standard
Public safety	No secrets, private prompts, customer data, local paths, or private URLs.
Human control	Risky work starts with PLAN MODE ONLY and waits for explicit approval.
Verification	Prompts require tests, QA, scans, or clear "unable to verify" reporting.
CI gate	The dependency-free Prompt Lab validator runs in GitHub Actions.
Source discipline	Current or unstable claims use source cards with confidence and outdated-risk notes.
Documentation	README stays concise; deeper guidance lives in docs/ and knowledge/.
Maintenance	Changelog, release notes, security policy, contribution guide, issue templates.

What This Is Not

This is not:

• an official OpenAI project,
• a Codex replacement,
• an app, API, hosted service, or automation platform,
• a package that calls external APIs by itself,
• a place to store secrets, private prompts, customer data, or credentials,
• a replacement for code review, security review, or engineering judgment.

It is a practical instruction system for designing, running, and reviewing Codex workflows with less ambiguity.

Core Workflow Modes

Goal + Full Prompt

Use when a persistent high-level objective helps Codex keep the definition of done in view.

no-Goal PLAN MODE ONLY

Use when Codex must inspect the repository and produce a plan before any edits. This is the safer default for broad, risky, security-sensitive, production-adjacent, or multi-file work.

APPROVED - EXECUTE

Use only after reviewing a plan. The execution prompt should tell Codex to follow the approved plan, preserve unrelated work, implement in small phases, verify, and stop if new facts expand scope.

STOP / RECOVER

Use when Codex starts editing too early or leaves scope. The recovery prompt should stop work, list changed files and commands run, avoid automatic revert, and return to plan-only mode.

Repository Structure

.
|-- .codex/skills/prompt-architect/
|   |-- SKILL.md
|   |-- commands.md
|   |-- response-modes.md
|   |-- codex-patterns.md
|   |-- examples.md
|   `-- knowledge/
|-- .github/
|   |-- ISSUE_TEMPLATE/
|   |-- workflows/
|   `-- pull_request_template.md
|-- docs/
|   |-- INSTALL.md
|   |-- USAGE.md
|   |-- EXAMPLES.md
|   |-- FAQ.md
|   |-- SKILL_STRUCTURE.md
|   |-- VALIDATION.md
|   |-- SOURCE_MAINTENANCE.md
|   |-- SUBAGENTS.md
|   |-- PLUGIN_READINESS.md
|   |-- PUBLIC_REPO_CHECKLIST.md
|   |-- SEO.md
|   |-- GITHUB_SETTINGS.md
|   |-- SECURITY_MODEL.md
|   `-- ROADMAP.md
|-- knowledge/
|   |-- distilled/
|   |-- templates/
|   |-- sources/
|   |-- outputs/
|   `-- logs/
|-- README.md
|-- README.tr.md
|-- package.json
|-- scripts/
|   |-- validate-prompt-lab.mjs
|   `-- install-prompt-architect.ps1
|-- SECURITY.md
|-- CONTRIBUTING.md
|-- CHANGELOG.md
`-- RELEASE_NOTES.md

Prompt Ledger

Reusable generated prompts are logged in:

knowledge/outputs/generated-prompts.md

Small one-off PROMPT_ONLY outputs are logged only when reusable, important, or explicitly requested.

Public Safety Rules

This repository is designed for public, reusable prompt and workflow patterns. It must not contain:

• API keys, tokens, credentials, cookies, private keys, or private URLs,
• customer data or internal company information,
• private system prompts,
• proprietary implementation details,
• unredacted logs, screenshots, local paths, or personal notes.

Before publishing changes, use docs/PUBLIC_REPO_CHECKLIST.md and docs/SECURITY_MODEL.md.

For local release checks, run:

node scripts/validate-prompt-lab.mjs

Contributing

Contributions are welcome when they improve Codex prompt clarity, safety, examples, documentation, or source-backed workflow guidance. See CONTRIBUTING.md.

Security

Do not open public issues for vulnerabilities, leaked credentials, private prompts, or accidental disclosure. See SECURITY.md.

License

MIT. See LICENSE.