cnfjlhj/prompt-polisher

Prompt Polisher

Transform messy, unstructured input into polished, Claude-optimized prompts using Anthropic's best practices.

When to Activate

Activate this skill when:

• User provides voice transcription content (contains filler words, self-corrections, casual speech)
• User dumps stream-of-consciousness notes
• User pastes rough document content that needs structuring
• User explicitly asks to "polish", "clean up", or "turn this into a prompt"
• Input is clearly unstructured and intended as task instructions

Workflow

Execute these stages in order:

Stage 1: Voice Cleanup

If the input appears to be voice transcription or casual speech, clean it up:

Remove:

• Filler words: "um", "uh", "like", "you know", "basically", "actually", "literally", "right?"
• Self-corrections: "no wait", "I mean", "actually scratch that", "let me rephrase"
• False starts: repeated or abandoned sentence beginnings
• Verbal tics: "okay so", "let me think", "hmm"

Normalize:

• Convert spoken patterns to written form
• Fix run-on sentences
• Add proper punctuation

Stage 2: Intent Extraction

Parse the cleaned input to identify:

| Element | What to Find | |---------|--------------| | Core task | What does the user actually want done? | | Referenced files | Documents, project files, URLs mentioned | | Constraints | Limitations, requirements, boundaries | | Preferences | Style, format, approach preferences | | Success criteria | What does "done" look like? | | Scope | Single task vs. multi-step project |

Stage 3: Gap Detection

Scan for missing or ambiguous elements:

• [ ] Ambiguous requirements ("process them somehow")
• [ ] Missing success criteria (no clear definition of "done")
• [ ] Unclear scope (single task or multi-step?)
• [ ] Format unspecified (code? prose? structured output?)
• [ ] Model unknown (Opus 4.5 or Sonnet 4.5?)
• [ ] Technical choices undefined (which library/approach/pattern?)

If gaps exist, ask ALL clarifying questions at once:

Before I polish this prompt, I need to clarify a few things:

1. [First gap-specific question]
2. [Second gap-specific question]
3. Which Claude model are you using? (Opus 4.5 / Sonnet 4.5 / Not sure)

Wait for user response before proceeding.

Stage 4: Apply Best Practices

Reference reference.md for detailed guidelines. Apply these transformations:

Universal (Claude 4.x):

• Make instructions explicit and direct (no hints)
• Add contextual framing (explain WHY, not just WHAT)
• Use positive framing ("use prose" not "don't use bullets")
• Structure with XML tags for complex inputs
• Use action-oriented language ("Change this" not "Can you suggest")
• Include chain of thought prompts for complex reasoning
• Define clear success criteria
• Use normal tone (avoid aggressive CAPS/MUST/CRITICAL)

If Opus 4.5:

• Add anti-over-engineering guardrail: "Keep solutions simple. Only make changes directly requested."
• Avoid the word "think" → use "consider", "evaluate", "believe"
• Leverage planning strength: "Build an editable plan before executing"
• Keep prompts token-efficient (don't over-pad)
• Provide policy/constraint context when relevant

If Sonnet 4.5:

• Break complex tasks into explicit steps
• Optimize for speed/throughput
• Add more explicit step-by-step guidance for intricate tasks

Stage 5: Structure the Prompt

Generate the polished prompt using this template:

<context>
[Background info, relevant project context, referenced files, WHY this task matters]
</context>

<task>
[Clear, explicit statement of what needs to be done - action-oriented]
</task>

<requirements>
[Specific constraints, boundaries, must-haves, technical choices]
</requirements>

<success_criteria>
[What "done" looks like, how to verify completion]
</success_criteria>

<approach>  <!-- Optional: include for complex tasks -->
[Suggested approach OR "Build an editable plan before executing"]
</approach>

Formatting rules:

• Keep each section focused and concise
• Use bullet points only when listing multiple items
• Omit sections that aren't relevant to the task
• For simple tasks, a streamlined version without XML tags is acceptable

Stage 6: Preview for Approval

Present the polished prompt to the user:

Here's your polished prompt:

---

[POLISHED PROMPT CONTENT]

---

**Options:**
- **Execute** - Run this prompt now
- **Edit** - I'll modify something first
- **Redo** - Start over with different clarifications

Wait for user approval before executing.

Stage 7: Execute or Iterate

Based on user choice:

• Execute: Immediately act on the polished prompt as if the user had typed it directly
• Edit: User modifies the prompt, then re-preview
• Redo: Return to Stage 1 with new input or Stage 3 for new clarifications

Important Notes

• Always preview before executing - never skip the approval step
• If the input is already well-structured, acknowledge this and offer minor polish only
• Learn from user feedback - if they frequently edit a certain way, note the pattern
• For very short/simple inputs, use a streamlined format without full XML structure
• The goal is USEFUL prompts, not LONG prompts - be concise

See Also

• reference.md - Detailed Anthropic best practices for Claude 4.x, Opus 4.5, and Sonnet 4.5
• examples.md - Before/after transformation examples