enhance-prompt

Enhance Prompt

Overview

Transform raw, vague, or incomplete prompts into professional-grade, context-rich instructions that produce optimal AI output.

Core principle: A bad prompt produces bad output. Every minute spent on prompt quality saves ten minutes of wrong output.

Announce at start: "I'm using the enhance-prompt skill to transform your prompt into a professional-grade instruction."

Do NOT output an enhanced prompt without first scanning the codebase for context. Even if the prompt seems clear, codebase context ALWAYS improves it. No exceptions.

The Iron Law

NO ENHANCED PROMPT WITHOUT CODEBASE CONTEXT

Skipping the codebase scan? The enhanced prompt will be generic. Generic prompts produce generic output. That defeats the purpose.

When to Use

Always:

• Prompt under 10 words
• Vague action verbs ("fix", "improve", "update", "optimize")
• Missing file paths, function names, or technical details
• Non-English prompt
• Code snippet with unclear instruction

Especially when:

• User pastes raw prompt and asks to improve it
• /enhance-prompt slash command is invoked
• User says "write a prompt for..." or "prompt to..."

Checklist

You MUST complete these steps in order:

1. Score the raw prompt — 7-dimension scorecard (see references/prompt-patterns.md)
2. Auto-detect category — match intent to template (see references/auto-detection.md)
3. Check for multi-intent — if 2+ concerns detected, trigger chain decomposition (see references/prompt-chain.md)
4. Scan the codebase — follow references/codebase-analysis.md procedure
5. Ask clarifying questions — MAX 2, only if critical ambiguity remains after scan
6. Apply 7-Layer Enhancement — Clarity → Specificity → Context → Structure → Constraints → Output → Verification
7. Format output — follow references/output-format.md template
8. Present before/after — show scoring improvement and copy-ready version
9. Support iteration — if user requests refinement, follow references/iteration-mode.md

Process Flow

digraph enhance_prompt {
    rankdir=TB;
    receive [label="Receive\nraw prompt", shape=box];
    score [label="Score\n7 dimensions", shape=box];
    detect [label="Auto-Detect\ncategory", shape=box, style=filled, fillcolor="#ffffcc"];
    multi [label="Multi-Intent?", shape=diamond];
    chain [label="Chain\nDecomposition", shape=box, style=filled, fillcolor="#ffffcc"];
    scan [label="Scan\ncodebase", shape=box, style=filled, fillcolor="#ccffcc"];
    ambiguous [label="Critical\nambiguity?", shape=diamond];
    ask [label="Ask MAX 2\nquestions", shape=box];
    enhance [label="Apply 7-Layer\nEnhancement", shape=box, style=filled, fillcolor="#ccccff"];
    format [label="Format\noutput", shape=box];
    present [label="Present\nbefore/after", shape=doublecircle];
    feedback [label="User\nfeedback?", shape=diamond];
    iterate [label="Iteration\nMode", shape=box, style=filled, fillcolor="#ffccff"];

    receive -> score;
    score -> detect;
    detect -> multi;
    multi -> chain [label="yes (2+)"];
    multi -> scan [label="no (1)"];
    chain -> scan;
    scan -> ambiguous;
    ambiguous -> ask [label="yes"];
    ambiguous -> enhance [label="no"];
    ask -> enhance;
    enhance -> format;
    format -> present;
    present -> feedback;
    feedback -> iterate [label="refine"];
}

The 7-Layer Enhancement

Each layer is mandatory. Skipping layers produces weak prompts.

Layer	What	Red Flag If Missing
1. Clarity	Remove ambiguity, define terms	Multiple interpretations possible
2. Specificity	File paths, function names, line numbers	"Fix the thing" without WHERE
3. Context	Tech stack, patterns, dependencies from scan	AI has to guess the project
4. Structure	Context → Task → Constraints → Output sections	Wall of text
5. Constraints	DO/DON'T rules, boundaries, edge cases	No guardrails = wrong output
6. Output Format	Code block, diff, step-by-step, docs	AI chooses random format
7. Verification	Testable success criteria	"Done" is subjective

Details and examples: See references/prompt-patterns.md

Red Flags — STOP and Redo

• Enhanced prompt has no file paths → Redo Layer 2
• No tech stack or dependencies mentioned → Redo Layer 3
• Output is a wall of text without sections → Redo Layer 4
• No DO/DON'T constraints → Redo Layer 5
• No verification criteria → Redo Layer 7
• Prompt is longer than the solution → Over-engineered, simplify
• Enhanced prompt contradicts codebase patterns → Rescan codebase
• Same information repeated across sections → Consolidate, remove redundancy
• All "After" scores are 5/5 → Likely inflated, re-evaluate each dimension honestly
• Verification criteria are subjective ("works well", "looks good") → Replace with testable commands/thresholds

Common Rationalizations

Excuse	Reality
"Prompt is clear enough"	If it were, you wouldn't need this skill
"Codebase scan takes too long"	30 seconds of scanning saves 10 minutes of wrong output
"Context is obvious"	Obvious to you. Not to the AI reading it cold
"I'll add details later"	You won't. Enhanced prompt is the details
"Just make it longer"	Length ≠ quality. Specificity = quality

Multilingual Handling

When prompt is not in English:

1. Detect language automatically
2. Preserve the user's intent — translate meaning, not words
3. Use English for ALL technical terms (database, authentication, frontend)
4. Output enhanced prompt in user's language with English technical terms

References

• references/prompt-patterns.md — 7-Layer framework with examples and scoring
• references/codebase-analysis.md — Multi-level scanning with monorepo and framework detection
• references/output-format.md — Structured output template with risk assessment
• references/prompt-library.md — 6 core prompt templates
• references/prompt-library-extended.md — 7 extended prompt templates
• references/auto-detection.md — Auto-category detection and template selection
• references/prompt-chain.md — Prompt chain decomposition for multi-intent prompts
• references/iteration-mode.md — Enhancement iteration and refinement tracking

Key Principles

• Codebase context is non-negotiable — Always scan before enhancing
• Auto-detect, don't assume — Use signal matching to identify prompt intent before selecting a template
• One concern per prompt — Multi-intent prompts get decomposed into focused chains
• Specificity beats length — src/auth/login.ts:42 > paragraph of description
• Score before and after — Prove the enhancement worked
• Score honestly — Inflated "After" scores help no one; every 4-5 must be justified by specific content
• Iterate with precision — Apply targeted changes on feedback, don't rewrite everything; max 3 iterations
• One question at a time — Don't overwhelm with clarification requests
• Copy-ready output — End with clean version ready to paste anywhere
• Output efficiency — Enhanced prompt must be concise and non-redundant; if it's longer than 700 words for a simple task, trim it