55 lines
3.2 KiB
Markdown
55 lines
3.2 KiB
Markdown
# AI Skill Prompt Files — Index
|
|
|
|
15 prompt files for all LLM-using AI skills. Each file contains: system prompt, user prompt template
|
|
with `{{variable}}` placeholders, output schema, confidence scoring guidance, and examples.
|
|
|
|
Regex-only skills (SKILL-001 Classifier, SKILL-008 MismatchDetector, SKILL-008a VersionDetector)
|
|
do not need prompt files — they use pure pattern matching.
|
|
|
|
## Prompt File Inventory
|
|
|
|
### Tier 1 — High Impact
|
|
|
|
| File | Skill | Model | Notes |
|
|
|---|---|---|---|
|
|
| `SKILL-001a-DocumentScopeDetector.md` | Scope detection (customer/fair/job) | gpt-4o-mini | Receives upload context + entity IDs |
|
|
| `SKILL-002-ContactExtractor.md` | Contact extraction | gpt-4o-mini | Regex primary, LLM fallback |
|
|
| `SKILL-003-DeadlineDetector.md` | Deadline detection (German focus) | gpt-4o-mini | Regex primary, LLM fallback for context |
|
|
| `SKILL-004-ChecklistGenerator.md` | Action item extraction | gpt-4o-mini | German output, 20-item limit, source quoting |
|
|
| `SKILL-004b-MultiFairDetector.md` | Multi-fair detection | gpt-4o-mini | Returns fair array from consultation docs |
|
|
| `SKILL-004c-BudgetExtractor.md` | Budget/cost extraction | gpt-4o-mini | Regex primary, LLM for context classification |
|
|
| `SKILL-004d-PainPointSummarizer.md` | Pain points + expectations | gpt-4o-mini | German output, structured summary |
|
|
|
|
### Tier 2 — Validation
|
|
|
|
| File | Skill | Model | Notes |
|
|
|---|---|---|---|
|
|
| `SKILL-005-StandDataExtractor.md` | Stand dimension extraction | gpt-4o | Handles multi-booth specs + hall info |
|
|
| `SKILL-006-ConfirmationParser.md` | Booking confirmation parsing | gpt-4o-mini | Service-type detection, discrepancy flagging |
|
|
| `SKILL-007-BrandConsistencyChecker.md` | Brand vs. photo comparison | gpt-4o (vision) | Compares brandbook palette to booth photos |
|
|
|
|
### Tier 3 — Enhancement
|
|
|
|
| File | Skill | Model | Notes |
|
|
|---|---|---|---|
|
|
| `SKILL-009-MaterialListExtractor.md` | Material/furniture extraction | gpt-4o-mini | Categorizes items into 4 types, German names |
|
|
| `SKILL-009a-LeadMagnetIdeator.md` | Activation idea generation | gpt-4o | Creative output, temperature 0.7, 3 ideas |
|
|
| `SKILL-010-AcceptanceReportGenerator.md` | Acceptance report generation | gpt-4o | Structured JSON for Gotenberg PDF template |
|
|
| `SKILL-011-SetupProgressAnalyzer.md` | Photo progress comparison | gpt-4o (vision) | Detects construction changes, suggests checklist |
|
|
| `SKILL-012-WeeklyProtocolSummarizer.md` | Protocol summarization | gpt-4o-mini | German output, timeline-friendly structure |
|
|
|
|
## Usage
|
|
|
|
Each prompt file is a Markdown document containing:
|
|
1. **Front matter** (YAML): skill ID, name, tier, model, temperature, response format
|
|
2. **System Prompt**: instructions for the LLM
|
|
3. **User Prompt Template**: the message sent to the LLM, with `{{variable}}` placeholders
|
|
that are populated at runtime by the PHP handler with extracted document text,
|
|
entity context, and metadata
|
|
4. **Output Schema**: the expected JSON structure
|
|
5. **Example** (some files): sample input/output pairs for testing
|
|
|
|
These files are the canonical prompt source. The PHP handler (`AiSkillTriggerListener`)
|
|
loads the prompt, replaces `{{placeholders}}` with runtime data, sends to the LLM API,
|
|
and parses the response against the schema.
|