Prompt Optimizer

Prompt Optimizer (oh-my-codex Edition)

Analyze a draft prompt, critique it, match it to oh-my-codex installed skills, and output a complete optimized prompt the user can paste and run in Codex.

When to Use

• User says "optimize this prompt", "improve my prompt", "rewrite this prompt"
• User says "help me write a better prompt for..."
• User says "what's the best way to ask Codex to..."
• User says "优化prompt", "改进prompt", "怎么写prompt", "帮我优化这个指令"
• User pastes a draft prompt and asks for feedback or enhancement
• User says "I don't know how to prompt for this"
• User explicitly invokes the prompt-optimize skill by name

Do Not Use When

• User wants the task done directly (just execute it)
• User says "优化代码", "优化性能", "optimize this code", "optimize performance"
• User says "just do it" or "直接做"

How It Works

Advisory only — do not execute the user's task.

Do NOT write code, create files, run commands, or take any implementation action. Your ONLY output is an analysis plus an optimized prompt.

Run this pipeline sequentially. Present results using the Output Format below.

Phase 0: Project Detection

Before analyzing the prompt, detect the current project context:

1. Check if AGENTS.md or CLAUDE.md exists in the working directory — read it for conventions
2. Detect tech stack from project files:
   • package.json → Node.js / TypeScript / React / Next.js
   • go.mod → Go
   • pyproject.toml / requirements.txt → Python
   • Cargo.toml → Rust
   • build.gradle / pom.xml → Java / Kotlin / Spring Boot
   • Package.swift → Swift / iOS
   • Gemfile → Ruby
   • composer.json → PHP
   • *.csproj / *.sln → .NET
   • Makefile / CMakeLists.txt → C / C++
3. Note detected tech stack for Phase 3 and Phase 4

If no project files found, flag "tech stack unknown" in Phase 4.

Phase 0.5: Lightweight Skill Catalog Inventory

Before recommending any skill, build a lightweight view of the current project skill catalog and resolve likely aliases. This phase is a routing index, not a full skill-body review.

Token budget rule:

1. Prefer the runtime skill index cache at .omcodex/cache/skill-index.json. If missing or stale, suggest running npm run catalog:skill-index or let tooling fall back to catalog/frontmatter lookup.
2. Use catalog/manifest data such as src/catalog/manifest.json, src/catalog/generated/public-catalog.json, or templates/catalog-manifest.json to confirm source, intent, layer, and conflicts.
3. If catalog files are missing or stale, inspect only each SKILL.md frontmatter block (name, description, intent, layer, origin, version). Do not read every skill body.
4. Prefer .agent/skills/local/**/SKILL.md over upstream copies when duplicate names exist.
5. Include upstream sources under .agent/skills/upstream/**/SKILL.md only as candidates, using metadata/frontmatter first.
6. Group available skills by intent/layer instead of dumping a long flat list: planning, execution, debugging, testing, review, verification, research, documentation, frontend, backend, security, orchestration, skill-management.
7. Shortlist at most 2-5 candidate skills or chains, then read full skill bodies only for the shortlisted skills if the prompt needs precise workflow details.
8. For short, clear prompts, skip full body reads entirely and use metadata plus the best-practice chains below.
9. Recommend only skills that appear in the current project/runtime catalog. If a well-known capability is absent, name the nearest installed substitute and state the substitution.

This keeps prompt optimization fast: catalog/frontmatter lookup is cheap, while full skill loading is reserved for ambiguous or high-risk choices.

Common aliases and substitutes:

Do not recommend phantom slash commands or unavailable skills. In optimized prompts, refer to skills in natural language: "Use the conductor skill..."

Phase 1: Intent Detection

Classify the user's task into one or more categories:

If a prompt contains multiple intents, split them instead of flattening them into one instruction. Preserve explicit order signals such as "先 X 再 Y".

Phase 2: Scope Assessment

For new requirements, prefer conductor when the work is HIGH/EPIC or when the prompt asks for durable context, a track, spec/plan/review artifacts, or multi-session continuity. Do not force conductor for TRIVIAL/LOW tasks; that adds ceremony without improving execution.

When using grill-me or any other workflow, check whether it changes the project's durable intent layer:

• Requirements, scope, acceptance criteria, architecture, workflow rules, or long-lived delivery plan changed -> include conductor refresh/new-track or reconcile steps in the optimized prompt.
• Pure implementation technique, debugging method, review pass, or verification gate with no intent change -> use the workflow directly and do not add conductor solely for formality.

Phase 3: oh-my-codex Skill Matching

Map intent + scope + tech stack to installed oh-my-codex skills.

By Intent Type

Best-Practice Skill Chains

Use ordered chains in the optimized prompt. Do not output an unprioritized pile of skills.

If a selected workflow changes durable intent, insert conductor refresh or reconcile before implementation and after review.

By Tech Stack

By Scope Escalation

Phase 4: Missing Context Detection

Scan the prompt for missing critical information:

• Tech stack — Detected in Phase 0, or must user specify?
• Target scope — Files, directories, or modules mentioned?
• Acceptance criteria — How to know the task is done?
• Error handling — Edge cases and failure modes addressed?
• Security requirements — Auth, input validation, secrets?
• Testing expectations — Unit, integration, E2E?
• Performance constraints — Load, latency, resource limits?
• UI/UX requirements — Design specs, responsive, a11y? (if frontend)
• Database changes — Schema, migrations, indexes? (if data layer)
• Existing patterns — Reference files or conventions to follow?
• Scope boundaries — What NOT to do?

If 3+ critical items are missing, ask up to 3 clarification questions before generating the optimized prompt. Incorporate answers into the output.

Phase 5: Workflow & Model Recommendation

Lifecycle position:

Research → Plan → Implement (TDD) → Review → Verify → Commit

For MEDIUM tasks, start with plan unless the task explicitly needs durable artifacts. For HIGH/EPIC tasks, start with conductor; use ralplan as a blueprint helper only when conductor is unavailable or the user explicitly requests it.

Model recommendation:

Multi-prompt splitting (for HIGH/EPIC scope):

• Prompt 1: Research + Plan (research/deepsearch skill, then conductor for HIGH/EPIC or plan for MEDIUM)
• Prompt 2–N: Implement one phase per prompt (each ends with verify skill)
• Final Prompt: Integration test + code-review across all phases
• Use project-session-manager skill to preserve context between sessions

Output Format

Present analysis in this exact structure. Respond in the same language as the user's input.

Section 1: Prompt Diagnosis

Strengths: What the original prompt does well.

Issues:

Needs Clarification: Numbered list of questions (omit if Phase 0 auto-detected the answer).

Section 2: Recommended oh-my-codex Skills

Start with a compact catalog summary based on catalog/frontmatter metadata, not full-body reads:

Then list only the selected skill chain. If full skill bodies were read, name which ones and why; otherwise state "metadata-only routing was sufficient."

Only list skills that are actually useful for this task. Do not pad the table. State when conductor was deliberately skipped because scope is TRIVIAL/LOW.

Section 3: Optimized Prompt — Full Version

Present the complete optimized prompt inside a single fenced code block. Must be self-contained and ready to copy-paste into Codex. Include:

• Clear task description with context
• Tech stack (detected or specified)
• Skill invocations at the right workflow stages (write as: "Use the plan skill to...")
• Acceptance criteria
• Verification steps
• Scope boundaries (what NOT to do)

Do NOT use slash-command syntax — Codex activates skills via conversation, not slash commands. Write skill references as natural language: "Use the plan skill", "Run tdd skill", "Apply verify skill".

Section 4: Optimized Prompt — Quick Version

Compact version for experienced users:

Section 5: Enhancement Rationale

Footer

Not what you need? Tell me what to adjust, or make a normal task request if you want execution instead of prompt optimization.

Examples

Example 1: Vague Chinese Prompt (Project Detected)

User input:

帮我写一个用户登录页面

Phase 0 detects: package.json with Next.js 15, TypeScript, Tailwind CSS

Optimized Prompt (Full):

使用项目现有技术栈（Next.js 15 + TypeScript + Tailwind CSS）实现用户登录页面。

技术要求：
- 沿用项目现有组件结构和路由约定
- 表单验证使用项目现有方案（Zod/Yup/其他，先搜索确认）
- 认证方式沿用现有方案（如无则默认 JWT）
- 包含：邮箱/密码表单、验证、错误提示、加载状态、响应式布局

工作流：
1. 使用 `plan` skill 规划组件结构和认证流程，参考现有页面模式
2. 使用 `tdd` skill 先写失败测试（表单单元测试 + 认证集成测试）
3. 实现登录页面和认证逻辑
4. 使用 `code-review` skill 审查实现
5. 使用 `verify` skill 确认所有测试通过且页面正常渲染

安全要求：密码不明文传输、防暴力破解、XSS 防护、CSRF token

验收标准：
- 所有测试通过，覆盖率 80%+
- 移动端和桌面端正常渲染
- 登录成功跳转 dashboard，失败显示错误

不要做：不实现注册页面、不实现忘记密码、不修改现有路由结构

Example 2: English Prompt (Go Project)

User input:

Add a REST API endpoint for user profile updates with validation

Phase 0 detects: go.mod with Go 1.22, Chi router

Optimized Prompt (Full):

Add a REST API endpoint for user profile updates (PATCH /api/users/:id).

Tech stack: Go 1.22 + Chi router (detected from project)

Requirements:
- PATCH /api/users/:id — partial update (name, email, avatar_url, bio)
- Input validation for all fields
- Auth: require valid token, users can only update own profile
- Return 200 with updated user on success; 400 for validation errors; 401/403 for auth

Workflow:
1. Use `plan` skill to design endpoint structure, middleware chain, and validation logic
2. Use `tdd` skill — write table-driven tests for success, validation failure, auth failure, not-found
3. Implement following existing handler patterns in the codebase
4. Use `code-review` skill + `security-review` skill
5. Use `verify` skill — run full test suite, confirm no regressions

Do not: modify existing endpoints, change database schema, add new dependencies without
first using `deepsearch` skill to check existing ones.

Example 3: EPIC Scope

User input:

Migrate our monolith to microservices

Optimized Prompt (Full):

Use the `conductor` skill to create a durable track for:
"Migrate monolith to microservices architecture"

Before executing, the track spec and plan must answer:
1. Which domain boundaries exist in the current monolith?
2. Which service to extract first (lowest coupling)?
3. Communication pattern: REST, gRPC, or event-driven?
4. Database strategy: shared DB initially or database-per-service?
5. Deployment target: Kubernetes, Docker Compose, or serverless?

Track phases:
- Phase 1: Identify service boundaries, create domain map
- Phase 2: Set up infrastructure (API gateway, CI/CD per service)
- Phase 3: Extract first service (strangler fig pattern)
- Phase 4: Verify with integration tests, extract next service
- Phase N: Decommission monolith

Each phase = 1 PR with `verify` skill gate.
Use `project-session-manager` skill between phases to preserve context.
Use `using-git-worktrees` skill for parallel service extraction.

Recommended: Opus 4.7 for conductor/spec planning, Sonnet 4.6 for phase execution.

Installed Skills Reference

Before using this reference, refresh it from .agent/skills/local and .agent/skills/upstream and prefer local overrides. Key skill groups commonly available in this oh-my-codex installation:

Planning & Architecture: plan, brainstorming, architect-planner, planning-methodology, planning-with-files, ralplan, ultrawork, ralph

Implementation: tdd, tdd-workflow, build-fix, refactor-clean, coding-standards, backend-patterns, frontend-patterns, frontend-design, api-design

Testing: tdd-generator, test-gen, test-coverage, e2e, e2e-testing, bdd-generator, verify, verification-loop

Review & Quality: code-review, security-review, aireview, quality-check, quality-validation, requesting-code-review, receiving-code-review

Research: research, deepsearch, iterative-retrieval, multi-model-research, documentation-lookup

Documentation: update-docs, update-codemaps, writer-memory

Debugging: systematic-debugging, debug-analysis, trace, log-analyzer

Orchestration: team, swarm, conductor, pipeline, autopilot, ultraqa

Session Management: project-session-manager, checkpoint, strategic-compact

Frontend: frontend-ui-ux, ui-ux-pro-max, frontend-slides, web-clone, nextjs-turbopack

Infrastructure: mcp-setup, mcp-server-patterns, start-dev, release, git-master

Skill Tools: skill-create, skill-development, skill-tester, skill-quality-analyzer