菜单
Use when writing, reviewing, or improving prompts for agents, subagents, or LLM calls — detects anti-patterns, suggests fixes, and enforces prompt quality principles
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Use after completing a major work batch (plan execution, feature build, multi-task subagent run) to refresh code-graph-derived data in MEMORY.md/CONTEXT.md/SKILL_MEMORY, append dated narrative entries to TODO/CHANGELOG/lessons-learned, mine cross-project structural habits, and ratchet reusable permissions to global settings. Keeps project DNS in sync with reality.
Use when orchestrating multiple ghengis-skills across a single task -- coordinates pre-action validation, execution, post-action verification, and lesson extraction via a shared JSON scratchpad. TRIGGER when dispatching a subagent, finishing a major task, or when the user mentions "run the full chain", "validate this", "fire the pipeline", or explicitly invokes a chain by name.
Use to build and maintain an evolving understanding of user preferences, communication style, and working patterns — learns from interactions and adapts behavior over time
Use when starting a new project or organizing an existing one — creates self-documenting 4-layer project structure with MEMORY.md, CONTEXT.md, modular CLAUDE.md, and workspace routing
Use to accumulate and retrieve domain knowledge from past tasks — builds a grepable plain-text knowledge base that grows over time, enabling agents to learn from experience without vector databases
Use when reorganizing an existing project's file tree based on what actually connects to what — analyzes import graph, classifies layers, proposes an ideal target tree, validates the plan adversarially, then executes file moves with atomic per-language import rewrites. Always plan-first; never moves a file before the user confirms. TRIGGER on "reorganize this project", "clean up the file tree", "the structure is a mess", "make this match a real project layout", or after project-scaffold detects layout drift.
| name | pql-validation |
| description | Use when writing, reviewing, or improving prompts for agents, subagents, or LLM calls — detects anti-patterns, suggests fixes, and enforces prompt quality principles |
Prompt Quality Layer for Claude Code. Validates prompts before execution, catches anti-patterns, and auto-fixes common issues. Zero cost, instant feedback.
<task>, <context>, <constraints> structureAll 35 anti-patterns are documented in anti-patterns.md. Here is the category overview:
| Category | Count | What It Catches |
|---|---|---|
| Task | 7 | Vague verbs, dual tasks, no success criteria, over-permissive delegation, emotional language, scope too large, implicit references |
| Context | 6 | Assumed knowledge, missing project context, hallucination invitations, undefined audience, no failure context, missing constraints |
| Format | 6 | Missing output format, implicit length, vague aesthetics, no examples, ambiguous lists, no tone specified |
| Scope | 6 | No scope boundary, no stack constraints, unbounded iteration, no error handling spec, no performance target, no backward compat |
| Reasoning | 5 | CoT on reasoning models, contradictory instructions, roleplay without purpose, prompt injection risk, negative framing |
| Agentic | 5 | No starting state, no target state, no stop condition, unlocked filesystem, no human review step |
Severity levels:
warn -- Informational. Prompt will likely work but could be improved.critical -- Likely to cause poor output, unsafe behavior, or runaway execution.These are common fixes you can apply without an LLM call:
| Problem | Fix |
|---|---|
| Vague verb ("handle", "deal with") | Replace with specific action: "parse", "validate", "refactor", "delete" |
| Filler words ("basically", "essentially", "just", "very", "really") | Remove them. They add zero information. |
| Preamble ("Sure!", "I'd be happy to help", "Let me explain") | Strip from LLM output before passing to next stage |
| "Think step by step" on reasoning models | Remove. Models like o1, o3, DeepSeek R1, Qwen3 reason internally. |
| "Do whatever you think is best" | Replace with explicit constraints and boundaries |
| "Keep going until it's perfect" | Set a fixed iteration limit: "try 3 approaches, pick the best" |
| Emotional language ("terrible", "garbage") | Replace with specific technical observations |
| Negative framing ("don't make it ugly") | Reframe positively: "make it readable with consistent spacing" |
| Implicit length ("keep it short") | Specify exact length: "200 words", "5 bullet points", "3 paragraphs" |
| "You know what I mean" | State context explicitly -- the model has no memory unless you provide it |
Use these when structuring prompts for different task types.
Role: Who the AI should act as
Task: What the AI should do
Format: How the output should be structured
Best for: Quick questions, single-step tasks, code generation with clear specs.
Context: Background and situation
Objective: The goal of the task
Style: Writing style to adopt
Tone: Emotional register (formal, casual, empathetic)
Audience: Who will read this
Response: Expected format of the response
Best for: Emails, reports, proposals, documentation aimed at a specific audience.
Role: Who the AI should act as
Instructions: High-level directives
Steps: Ordered list of actions to perform
End Goal: What success looks like
Narrowing: Constraints and scope limits
Best for: Multi-step tasks, implementation plans, refactoring workflows.
Objective: The agent's mission
Starting State: Initial conditions and context
Target State: Desired end state
Allowed Actions: Tools and actions the agent may use
Forbidden Actions: Actions the agent must not take
Stop Conditions: When to halt execution
Checkpoints: Intermediate verification points
Best for: Agentic workflows, autonomous execution, Claude Code subagents.
File: Path to the file being edited
Function: Target function or method
Current Behavior: What the code does now
Desired Change: What should change
Scope: Boundaries of the edit
Constraints: Rules the edit must follow
Done When: Acceptance criteria
Best for: Targeted code changes, bug fixes, feature additions in specific files.
Wasted tokens = wasted money and context window. Apply these rules to all prompts and handoffs:
Filler words to remove: basically, essentially, actually, literally, obviously, clearly, simply, just, really, very, quite, rather, pretty much, kind of, sort of, in order to, due to the fact that, at this point in time, it should be noted that, it is important to note that, as a matter of fact, in terms of
Preamble patterns to strip from LLM output before passing downstream: "Sure!", "Okay!", "I'd be happy to help.", "Here is the result:", "Let me help...", "Let me explain...", "Let me show...", "Let me provide...", "I found that", "I discovered that", "I noticed that"
Structured handoffs: When passing data between agents or stages, use structured formats (JSON, XML tags) instead of prose. A 10-line natural language handoff can often be a 3-line structured block.
Thoughts that mean your prompt needs work:
| If you're thinking... | The problem is... | Fix it by... |
|---|---|---|
| "The model should know what I mean" | Assumed knowledge | Stating context explicitly |
| "It keeps giving me the wrong thing" | Missing constraints or format spec | Adding boundaries and output format |
| "Just make it better" | No success criteria | Defining measurable outcomes |
| "Do this and also that and also..." | Multiple tasks in one prompt | Splitting into separate prompts |
| "I'll know it when I see it" | Vague aesthetic requirements | Specifying exact visual properties |
| "Keep going until it's perfect" | Unbounded iteration | Setting a fixed iteration limit |
| "Use your best judgment" | Over-permissive delegation | Providing explicit constraints |
| "Fix it again, it still doesn't work" | No failure context | Including error messages and what you tried |
| "Write a blog post" (no audience) | Undefined audience | Specifying who will read it |
| "Make the API fast" | No performance target | Setting a measurable target (e.g., "under 200ms") |
This skill participates in skill-chain-supervisor chains via the shared scratchpad at <project>/.claude/ghengis-chain/context.json.
Role in chain: Pre-execution validator. Runs before any agent dispatch or major task.
Scratchpad subkey (namespaced writes): pql_validation.*
Reads (input scratchpad keys):
input.prompt_draftinput.user_requestWrites (output scratchpad keys):
pql_validation.score � float 0-1, overall prompt qualitypql_validation.anti_patterns � list of detected anti-pattern namespql_validation.suggested_fixes � list of fix suggestions (strings)pql_validation.tier_applied � 'tier1' | 'tier2' | 'none' � whether auto-fix was usedSuccess criteria: score >= 0.5 OR tier1/tier2 auto-fix applied
When invoked as part of a chain, this skill MUST:
pql_validation.* namespace only � never overwrite another skill's subkeypql_validation.error) rather than raisingWhen invoked standalone (not in a chain), scratchpad writes are optional but recommended for auditability.