| name | prompt-cleaner |
| description | Transforms raw unstructured prompts into CLAUDE.md-aligned structured prompts using format selection rules (Markdown/YAML/JSON). Use when user submits a messy or incomplete prompt before sending to Claude Code, or when you need to standardize a prompt's structure with goal, context, constraints, and output_format components. |
Prompt Cleaner — Architecture Design
Mission
Transform raw prompts (Vietnamese/English) into structured prompts aligned with CLAUDE.md format selection rules. Multi-format output: Markdown for explanations, YAML for constraints, JSON for structured data.
Workflow Progress Tracker
### [prompt-cleaner] Progress:
- [ ] Phase 1: COLLECT
- [ ] Phase 2: ANALYZE
- [ ] Phase 3: RESTRUCTURE
- [ ] Phase 4: OUTPUT
Boot Sequence
Read Tier 1 files at boot:
Phase 1: COLLECT
Receive raw prompt. Assess confidence level.
Confidence Scoring Heuristic:
Components detected (Goal, Context, Constraints, OutputFormat):
- 4/4 = ≥80% confidence
- 3/4 = 60-79% confidence
- 2/4 = 40-59% confidence
- 1/4 = 20-39% confidence
- 0/4 = <20% confidence
Decision Matrix:
| Confidence | Action |
|---|
| < 40% | AskUserQuestion — 3 specific questions to clarify |
| 40-70% | Proceed to Phase 2 + Optional subagent explore |
| ≥ 70% | Proceed to Phase 2 directly |
Interaction Point #1: If confidence < 40%, ask user:
- What is the main goal you want to achieve?
- What context or information do you already have?
- What output format do you expect?
Phase 2: ANALYZE
Extract 4 required components:
| Component | Description | Prompt Question |
|---|
| GOAL | Primary objective (1-3 lines) | "What should the AI do?" |
| CONTEXT | Available information (sufficient/insufficient?) | "What context exists?" |
| CONSTRAINTS | Rules, limitations | "What restrictions apply?" |
| OUTPUT_FORMAT | Expected output format | "What format should output be?" |
Context Augmentation (Optional):
- If confidence < 70% and related to codebase
- Use
delegate_task subagent explore (max 5 files)
- Must cite source:line for all additions
- Forbidden: Adding information without source
Interaction Point #2: After extraction, if confidence 40-70%, present 4 components for user confirmation.
Phase 3: RESTRUCTURE
Step 3.1: Select Format
Read data/format-selection-rules.yaml.
Apply format selection matrix:
Fallback: If format does not match any rule, default to Markdown.
Step 3.2: Apply Template
Load appropriate template. Fill placeholders:
{{GOAL}} — from Phase 2 extraction{{CONTEXT}} — from Phase 2 extraction{{CONSTRAINTS}} — from Phase 2 extraction{{OUTPUT_FORMAT}} — from Phase 2 extraction
Step 3.3: Validate
Read loop/clean-checklist.md.
Run validation. Pass threshold: ≥5/6 items.
Phase 4: OUTPUT
| Condition | Action |
|---|
| Prompt already structured | Return as-is + note "Already well-structured" |
| Confidence < 40% | Return AskUserQuestion first |
| Validation PASS (≥5/6) | Deliver cleaned prompt |
| Validation FAIL (<3/6) | Retry restructure with checklist feedback |
| Validation PARTIAL (3-4/6) | Proceed but note improvement areas |
Guardrails
| ID | Rule | Description |
|---|
| G1 | No hallucination | Must cite source:line for all context additions |
| G2 | Over-clean guard | If prompt already structured → return as-is + note |
| G3 | Format whitelist | Only use formats in data/format-selection-rules.yaml |
| G4 | Goal-first | <goal> MUST appear first in output |
| G5 | Ask when unsure | Confidence < 40% → AskUserQuestion, don't guess |
| G6 | Length budget | Cleaned prompt ≤ 3x original length |
Placeholder Syntax
{{GOAL}} — Primary objective
{{CONTEXT}} — Background information
{{CONSTRAINTS}} — Rules and limitations
{{OUTPUT_FORMAT}} — Expected output structure
{{CONFIDENCE}} — Confidence score (0-100%)
{{CITE}} — Source citation (file:line)
Notes
- XML format has been deprecated per design.md §1 — use Markdown/YAML/JSON only
- Templates have backup fallback to Markdown if format does not match
- Subagent explore limited to 5 files maximum
