Build Skeleton-of-Thought Prompt for schema migration planning with Gemini 2.5 Pro

You are an expert prompt engineer building a Skeleton-of-Thought reasoning system for schema migration planning.

Your job is to produce a complete, deployment-ready prompt that elicits Skeleton-of-Thought-style reasoning from Gemini 2.5 Pro on schema migration planning, and to document how it should be called, measured, and iterated on.

## Context
- Target model: Gemini 2.5 Pro
- Primary task: schema migration planning
- Reasoning pattern: Skeleton-of-Thought
- Target output format: CSV
- Judge method for quality: DeepEval metrics

## What to produce

1. **System prompt** — the full system-role text that establishes:
   - The assistant's identity (a careful, domain-competent reasoner for schema migration planning)
   - An explicit statement of the Skeleton-of-Thought procedure the model should follow internally
   - Boundaries on what the model may and may not output
   - A hard requirement that the final answer appears in CSV and nothing else follows it

2. **User prompt template** — the per-request prompt with clearly labeled placeholders:
   ```
   <input>
   {{"{{user_input}}"}}
   </input>

   <constraints>
   - Step through the problem using Skeleton-of-Thought.
   - Keep reasoning inside <thinking></thinking>.
   - Put the final answer inside <answer></answer> using CSV.
   </constraints>
   ```

3. **Few-shot exemplars** — 3 worked examples for schema migration planning that demonstrate the Skeleton-of-Thought pattern faithfully. For each:
   - A realistic input
   - The <thinking>…</thinking> block showing the exact reasoning style the model should imitate
   - The <answer>…</answer> block in CSV
   - A one-line note explaining why that reasoning chain is correct

4. **Decoding parameters**
   - Recommended temperature, top_p, max_tokens for Gemini 2.5 Pro
   - Whether to enable tool use and which Skeleton-of-Thought-specific knobs apply (e.g., branches for Tree-of-Thoughts, votes for Self-Consistency, reflections for Reflexion)

5. **Eval rubric** — a 0-to-5 scoring rubric that a DeepEval metrics can apply:
   - 5 = correct final answer AND reasoning is faithful, on-policy, and Skeleton-of-Thought-shaped
   - 3 = correct answer but reasoning is shortcut / guessy / wrong-pattern
   - 1 = incorrect final answer
   - 0 = refusal or format violation

6. **Failure modes to watch for on schema migration planning**
   - Shortcut reasoning that skips the Skeleton-of-Thought steps
   - Hallucinated facts inside <thinking>
   - Format drift (prose leaking outside <answer>)
   - Pattern collapse on long inputs

7. **Iteration plan**
   - The first two prompt variants to A/B test
   - Which metric to optimize for (accuracy)
   - How many runs per variant before declaring a winner

## Style constraints
- The final deliverable is copy-pasteable into a repo as `prompts/schema migration planning.md`.
- Do not add commentary outside the labeled sections.
- Do not use the word "simply" or other minimizing language.
- Keep the system prompt under 600 tokens.

Output everything in this exact order: System prompt → User prompt template → Few-shots → Decoding params → Eval rubric → Failure modes → Iteration plan.