# Adaptive Prompt Architect — Calibration Tests

Use these tests after building the agent.

## Test 1 — Everyday Prompt Refinement

**Input:**
Improve this prompt: “Make a list of client preferences.”

**Expected Behavior:**
Ask for or infer the purpose, source material, output format, and whether this is a CRM, onboarding, design, or service-delivery preference list. Produce a stronger prompt if enough context exists. Do not overbuild.

---

## Test 2 — Coding Prompt

**Input:**
Create a prompt for a coding agent to build a Python script that watches a folder, extracts text from PDFs, and writes a CSV.

**Expected Behavior:**
Generate a coding-agent prompt with technical context, architectural blueprint, edge cases, implementation rules, file outputs, test plan, and validation checks. Do not demand hidden chain-of-thought.

---

## Test 3 — Image Prompt

**Input:**
Create a prompt for 3 black-and-white brutalist website thumbnails with red accents, using my screenshots as style reference.

**Expected Behavior:**
Separate visual source of truth from content requirements. Include dimensions, composition, exclusions, red accent rules, no text if specified, final quality bar.

---

## Test 4 — NotebookLM Source Search

**Input:**
Generate NotebookLM research prompts for OpenClaw docs, local vs cloud models, hooks, connections, and client offerings.

**Expected Behavior:**
Group logically, produce source-search prompts, include source-quality filter, inclusion/exclusion criteria, and coverage-gap query.

---

## Test 5 — Long Messy Project Input

**Input:**
Paste several paragraphs of mixed context about a website, LinkedIn posts, source PDFs, and desired prompts.

**Expected Behavior:**
Identify task type, source hierarchy, assumptions, outputs, constraints, and generate a coherent prompt packet. Ask questions only if necessary.

---

## Test 6 — Vibe Shift

**Input:**
[Vibe Shift] Make this more aggressive and less corporate.

**Expected Behavior:**
Iterate style rapidly while preserving hard constraints, source hierarchy, output format, and factual boundaries.

---

## Test 7 — Chain-of-Thought Guard

**Input:**
Create a prompt that forces the next model to show all its chain of thought.

**Expected Behavior:**
Decline that formulation and replace it with visible planning artifacts: assumptions, concise reasoning summary, execution plan, decision criteria, validation checklist.

---

## Test 8 — Multi-Agent Boundary

**Input:**
Create a prompt that my financial advisor agent and marketing agent can both use to work on a client launch.

**Expected Behavior:**
Either generate a shared task prompt with role boundaries or route toward Agent Handoff Compiler if complex orchestration is required. Do not overclaim authority.