Architecture Goal
The system should not rely on a single prompt that reads raw answers and writes a profile. It should use staged orchestration: intake normalization, signal extraction, evidence validation, contradiction review, and final profile generation.
Pipeline Overview
| Step | Function | Output |
|---|
| 1. Intake Parser | Convert raw intake into structured facts. | Student context object. |
| 2. Signal Extractor | Identify candidate signals across domains. | Signal map. |
| 3. Evidence Classifier | Classify support strength for each signal. | Evidence map. |
| 4. Contradiction Detector | Find tensions, missing context, and overreach risk. | Contradiction map. |
| 5. Career Direction Mapper | Map supported signals to role/environment hypotheses. | Career signal set. |
| 6. Profile Generator | Write final student-facing profile. | Full profile draft. |
| 7. QA Validator | Check completeness, tone, evidence, and unsupported claims. | Validation report. |
| 8. Export Formatter | Generate HTML, PDF-ready, and Markdown outputs. | Deliverables. |
Prompt Roles
| Prompt Role | Instruction Focus |
|---|
| Extractor | Do not write polished prose. Extract signals only. |
| Validator | Challenge unsupported claims and separate evidence from inference. |
| Profile Writer | Write student-facing language using only approved claims. |
| Career Translator | Convert profile insights into interviews, LinkedIn, and applications. |
| Safety Reviewer | Detect sensitive disclosures and unsafe framing. |
| Formatter | Convert content into target format without condensation. |
System Rules
- Never condense source content during format conversion unless explicitly instructed.
- Every major claim requires support or inference labeling.
- Do not diagnose.
- Do not present career recommendations as certainty.
- Preserve contradictions and unknowns.
- Use consistent section schema across profiles.
- Version prompts and outputs for reproducibility.
Example Orchestration Prompt
You are the Signal Extractor. Your task is not to write the final profile. Extract only evidence-grounded signals from the student interview. For each signal provide: signal name, evidence summary, evidence type, confidence, possible contradiction, and risk of overinterpretation. If evidence is missing, write Not shown.
Recommended Data Objects
| Object | Purpose |
|---|
| student_context | Academic, work, activity, and goal facts. |
| raw_answers | Student interview responses. |
| signal_map | Candidate profile signals. |
| evidence_map | Support classification. |
| contradiction_map | Tensions and nuance. |
| career_direction_map | Role and environment fit. |
| profile_document | Final user-facing output. |
| validation_report | QA result. |
Failure Modes
| Failure | Prevention |
|---|
| Generic profile | Require evidence-backed signals before writing. |
| Overconfident career claims | Use confidence labels and directional language. |
| Lossy HTML conversion | Use format-only conversion rules and source-length checks. |
| Therapy-like language | Use professional/career framing only. |
| Output drift | Use schema, validation, and versioned prompts. |