Prompt Iteration Skill

Iteratively improve teaching agent prompts using automated evaluation with local Qwen 2.5-7B as judge.

When to Use

• •After identifying functional issues with the agent
• •When prompt changes need validation before deployment
• •For systematic prompt engineering with measurable improvement

Pre-flight Checks

Before starting, verify:

1. •

   vLLM is running:

   bash

   curl http://localhost:8004/v1/models
   

2. •

   Golden dataset exists:

   bash

   ls tests/golden/
   

3. •

   Create iteration branch:

   bash

   git checkout -b eval/prompt-iteration-$(date +%Y%m%d-%H%M)
   

Guardrails (USER-DEFINED)

Iteration Loop

Step 1: Baseline Evaluation

Run evaluation on all dimensions:

python -m tests.evaluation.cli --dimension all --output baseline.json

Record the baseline pass rate for each dimension in a TodoWrite checklist.

Step 2: Identify Worst Dimension

From baseline.json, find the dimension with lowest pass rate. Focus on one dimension at a time.

Step 3: Analyze Failures

Run verbose evaluation on the failing dimension:

python -m tests.evaluation.cli --dimension <failing_dimension> --verbose

Review failed test cases to understand the failure pattern:

• •What is the agent doing wrong?
• •Which prompt file controls this behavior?
• •What specific language could fix it?

Step 4: Edit Prompt

Modify the relevant prompt file based on analysis:

Make targeted, minimal changes. Do not rewrite the entire prompt.

Step 5: Re-evaluate

Run evaluation on the dimension you're fixing:

python -m tests.evaluation.cli --dimension <dimension> --output iteration-N.json

Step 6: Check Circuit Breakers

Compare iteration-N.json to baseline.json:

1. •

   Did any dimension drop >20%?

   • •YES: git checkout src/backend/app/prompts/ and STOP
   • •NO: Continue
2. •

   Did the target dimension improve?

   • •YES: Commit and continue
   • •NO: Increment no-improvement counter
3. •

   3 consecutive no-improvement?

   • •YES: STOP and report
   • •NO: Continue

Step 7: Commit if Improved

If pass rate improved:

git add src/backend/app/prompts/
git commit -m "prompt: improve <dimension> pass rate N% -> M%

- [Description of what changed]
- [Why this should help]

Tested with: python -m tests.evaluation.cli --dimension <dimension>
"

Step 8: Repeat or Stop

Continue loop until one of these conditions:

• •SUCCESS: All dimensions >= 80% pass rate
• •MAX ITERATIONS: 5 iterations completed
• •REGRESSION: Any dimension dropped >20% from baseline
• •STALLED: 3 consecutive iterations with no improvement

Output Format

After loop completes, produce this summary:

## Evaluation Summary

**Baseline**: 45% overall
**Final**: 82% overall
**Iterations**: 4
**Status**: SUCCESS / STOPPED (reason)

### Dimension Results

| Dimension | Baseline | Final | Change |
|-----------|----------|-------|--------|
| language_choice | 40% | 85% | +45% |
| tool_usage | 60% | 95% | +35% |
| output_cleanliness | 50% | 90% | +40% |
| confusion_recovery | 30% | 70% | +40% |
| persona_consistency | 40% | 80% | +40% |
| curriculum_alignment | 50% | 75% | +25% |

### Commits Made

- `abc123` prompt: improve language_choice 40% -> 65%
- `def456` prompt: improve tool_usage 60% -> 95%
- `ghi789` prompt: improve output_cleanliness 50% -> 90%

### Next Steps

- [If success] Ready to merge to main
- [If stopped] Manual review needed for dimension X

Quick Reference

# Full evaluation
python -m tests.evaluation.cli --dimension all

# Single dimension
python -m tests.evaluation.cli --dimension language_choice --verbose

# Check judge availability
python -m tests.evaluation.cli --check

# List test cases
python -m tests.evaluation.cli --list

# A/B comparison (if testing prompt variants)
python -m tests.evaluation.cli --compare prompts/v1.md prompts/v2.md

Prompt Files Reference

src/backend/app/prompts/agent/
├── base.md           # Core rules, language handling, confusion recovery
├── mode_help.md      # Vocabulary helper mode behavior
├── mode_practice.md  # Conversation practice mode behavior
└── tools.md          # Tool definitions (speak, render_vocabulary, etc.)