Designing diagnostic prompts

Diagnostic Prompt Design

Design multi-turn prompts that diagnose before advising. This skill turns a loose "be my coach/tutor/advisor" request into a structured workflow with intake, configuration, assessment, plan, and feedback loops.

Inputs

• Domain and audience: life coaching, learning, product onboarding, team diagnosis, technical training, etc.
• Target outcome: insight, curriculum, decision support, behavior plan, or skill acquisition.
• Risk boundary: what the agent must not diagnose, decide, promise, or store.
• Runtime surface: chat prompt, custom GPT, skill, app flow, or project-local VEMO task.

Procedure

1. Frame the diagnostic map. Choose 3-5 dimensions that explain the domain. Human-development examples use quadrants; learning examples use topic, depth, style, pace, and assessment.
2. Define levels and transition states. Use levels only when they change behavior. Prefer "current state -> constraint -> next action" over vague maturity labels.
3. Ask before advising. Start with a short intake. Ask one question at a time unless the user explicitly wants a form.
4. Separate configuration from content. Let the user set depth, tone, language, pacing, examples, assessment frequency, and constraints. Provide commands or explicit slots when useful.
5. Locate the binding constraint. Identify the dimension or misconception that blocks progress now. State evidence and uncertainty.
6. Produce a plan. Give a minimum effective plan: next step, why it matters, what to measure, and what would trigger a revision.
7. Loop. Add checkpoints: quiz, reflection, user feedback, progress update, or reconfiguration. The agent should adapt rather than repeat the same advice.
8. Guard the boundary. For health, finance, legal, mental health, or high-stakes life decisions, keep the output as coaching/support, not professional diagnosis or instruction.

Output Shape

Produce one of these, depending on the request:

• Prompt spec: role, goal, inputs, commands, intake questions, assessment rubric, response format, stop rules.
• Agent flow: states, transitions, memory fields, commands, validation checks.
• Skill draft: SKILL.md-ready workflow with generic inputs and project-specific values left to runtime.

Quality Bar

• The prompt does not give advice before it has enough context.
• The dimensions are domain-specific, not decorative.
• The user can change configuration without editing the whole prompt.
• The plan contains measurable next actions and a revision trigger.
• The workflow says what it will refuse or escalate.
• Any source inspiration is cited as pattern provenance, not copied as prompt text.

References

• references/source-patterns.md — compact pattern extraction from Human 3.0 and Mr. Ranedeer.