| name | prompt-enhancer |
| description | Use when the user provides a vague, ambiguous, or underspecified request that could go multiple directions. Also use when the user explicitly asks to enhance, sharpen, or improve a prompt before executing it. Triggers on requests like "fix the bug", "make it better", "write something about X", or any request where critical details are missing. |
Prompt Enhancer
Transforms vague requests into precise, actionable prompts. One pass, no back-and-forth.
When to Use
digraph when {
"User request received" [shape=box];
"Is it specific enough?" [shape=diamond];
"Could it go multiple directions?" [shape=diamond];
"Are critical details missing?" [shape=diamond];
"Enhance it" [shape=box];
"Execute directly" [shape=box];
"User request received" -> "Is it specific enough?";
"Is it specific enough?" -> "Execute directly" [label="yes"];
"Is it specific enough?" -> "Could it go multiple directions?" [label="no"];
"Could it go multiple directions?" -> "Enhance it" [label="yes"];
"Could it go multiple directions?" -> "Are critical details missing?" [label="no"];
"Are critical details missing?" -> "Enhance it" [label="yes"];
"Are critical details missing?" -> "Execute directly" [label="no"];
}
Enhance when: the prompt lacks specifics, has ambiguous terms, omits scope, or could be interpreted multiple ways.
Skip when: the request is already specific and actionable, is a follow-up in an established conversation (context is clear), the user prefixes with * (bypass), the user says "just do it" or "don't enhance", or the task is a single obvious action.
Edge case -- ambiguous but runnable: If a request is 80%+ clear, enhance silently in your reasoning and execute. Only stop to present when the ambiguity could send you in the wrong direction entirely.
The Enhancement Process
Step 1: Diagnose
Identify what's missing. Common gaps:
| Gap | Signal | Example |
|---|
| Missing target | Noun without referent | "fix the bug" (which bug?) |
| Missing scope | No boundaries defined | "refactor the API" (which parts?) |
| Missing criteria | No success measure | "make it better" (better how?) |
| Missing constraints | No rules given | "write an article" (for whom? how long?) |
| Ambiguous action | Verb could mean many things | "handle it" (handle how?) |
| Missing context | No prior state referenced | "update the design" (from what?) |
Step 2: Context (lightweight, only when needed)
Gather context only if the prompt references specific things that exist in the environment:
- References files/code: Quick Glob/Grep to find relevant files. Read 1-2 max. Stop.
- References prior work: Check conversation history or memory.
- General/creative request: Skip exploration entirely. Enhance from the request alone.
Do not explore "just in case." Exploration has a cost. Only explore when the prompt names something specific you can't resolve from context.
Step 3: Enhance
Restructure the prompt following this framework. Use only the sections that add clarity -- drop sections that would be padding:
[INTENT] -- One sentence: what the user actually wants accomplished.
[SCOPE] -- What's included and excluded. Named files, components, areas, or boundaries.
[APPROACH] -- How to do it, when it matters. Reference existing patterns, tools, or conventions when context was gathered.
[CONSTRAINTS] -- Must-follow rules: backwards compatibility, existing patterns, style requirements, platform limits.
[DELIVERABLE] -- What the output should look like: file changes, text format, structure, medium.
[DONE WHEN] -- Measurable success criteria. How does the user verify it works?
Step 4: Present
Show the user:
- The enhanced prompt (the sections above, as a single coherent prompt)
- A brief note on what changed and why (one line per major addition)
- Up to 3 critical questions if unknowns remain that block 80%+ of the work
Then wait. Do not execute until the user approves or modifies.
Enhancement Patterns
| Vague | Enhanced |
|---|
| "fix the bug" | "Fix [specific condition] in [scope] where [trigger] causes [error], preserving [existing behavior]" |
| "make it better" | "Improve [metric/quality] of [scope] by [approach], measured by [criteria]" |
| "add X" | "Implement [specific behavior] in [scope] following [pattern/convention], with [constraints]" |
| "write about Y" | "Write [format] about [specific angle] for [audience], [length], [tone]" |
| "refactor" | "Restructure [scope] to [goal] while preserving [invariants]" |
| "research Z" | "Investigate [specific question] about [domain], covering [aspects], deliver as [format]" |
Principles
Specificity over verbosity. Better structure, not more words. A 3-line enhanced prompt beats a 20-line one if it captures the same intent.
Preserve intent. Never change what the user wants. Only clarify how. If the user says "simple," don't add complexity. If the user says "quick," don't plan a deep approach.
Positive framing. Say what TO do, not what NOT to do. "Use tab indentation" beats "Don't use spaces."
Ground in reality. When context is gathered, reference actual files, patterns, and conventions. Not hypotheticals.
Show, don't ask. Prefer presenting an enhanced prompt with your best interpretation over asking 10 questions. Ask only when a critical unknown blocks most of the work.
Don't invent requirements. Enhance what's there. Don't add requirements the user didn't imply. "Fix the bug" doesn't become "Fix the bug and add comprehensive test coverage and documentation."
Common Mistakes
| Mistake | Fix |
|---|
| Over-exploring: reading 10 files for context | Read 1-2 files max. Stop. |
| Adding requirements the user didn't ask for | Only clarify, never expand scope |
| Making the prompt longer for the sake of it | If the original is already clear on a dimension, skip it |
| Asking 6+ questions before acting | Present your best interpretation, ask max 3 critical ones |
| Rewriting the user's style/tone | Preserve their voice; enhance structure and specificity |
| Enhancing when not needed | If 80%+ clear, just execute |
| Exploring for creative/general tasks | Skip exploration unless the prompt names something specific |
