prompt

Activate When

• /godmode:prompt, "design a prompt", "test prompt"
• "optimize my prompt", "prevent prompt injection"
• Building LLM-powered features

Workflow

1. Requirements

Task: <what the prompt must accomplish>
Model: <target model>
Input/Output: <format and constraints>
Quality: accuracy target, safety, consistency
Budget: max tokens, latency, cost per call

2. Pattern Selection

Pattern	Best for
Zero-shot	Simple tasks model handles well
Few-shot	Tasks needing format/style examples
Chain-of-thought	Reasoning, math, multi-step
ReAct	Tool-using agents, search+reason
Tree-of-thought	Exploring alternatives
Self-consistency	High-stakes, multiple paths
Structured output	JSON, XML, typed schema

IF simple classification: zero-shot. IF format matters: few-shot with 2-5 examples.

3. System Prompt Design

Structure: 1) Role, 2) Task, 3) Input Format, 4) Output Format, 5) Constraints, 6) Examples, 7) Edge Cases.

Most important instructions at beginning and end (primacy/recency effect). Examples > instructions.

4. Few-Shot Examples

Cover: common case, edge case, output format. 2-5 examples typical. Track token overhead.

5. Reasoning Design

• CoT: "Think step by step: 1. Identify, 2. Analyze"
• ReAct: Thought -> Action -> Observation -> Answer
• ToT: generate 3 approaches, evaluate, select
• Self-consistency: N generations -> majority vote

6. Structured Output

Options: JSON mode, function calling, prompt-based, constrained decoding. Validate against schema. Retry on failure (max 2-3 attempts).

7. Injection Prevention

1. Input sanitization: strip patterns, max length
2. Prompt structure: delimiters, <user_input> tags
3. Output validation: format check, leak detection
4. Monitoring: log prompts/completions, alert anomalies

IF user input enters prompt: injection defense required. IF output contains system prompt text: leak detected.

8. Testing & Evaluation

Categories: golden set, edge cases, format compliance, safety, injection resistance, consistency. Metrics: accuracy (target), format (100%), safety (>99%), injection resistance (>95%), latency, cost.

9. Versioning & A/B Testing

Track: version, accuracy, latency, cost, status. A/B test with traffic split, significance testing (alpha=0.05). Minimum 100 samples per variant.

10. Artifacts

prompts/<task>/: prompt-spec.yaml, system-prompt.md, examples.yaml, tests.yaml, eval-results.md.

# Test prompt templates
curl -X POST http://localhost:8080/api/chat -d '{"prompt":"test"}'
pytest tests/test_prompts.py -v

Optimization Loop

WHILE iteration < 5 AND accuracy < target:
  1. DIAGNOSE failures (format, wrong, hallucination)
  2. GENERATE ONE change targeting top failure
  3. EVALUATE on same golden set
  4. COMPARE: accept if improved + no regression

Hard Rules

1. NEVER ship without golden set evaluation.
2. NEVER hardcode prompts inline -- externalize.
3. NEVER change >1 variable per iteration.
4. NEVER ignore injection if user input in prompt.
5. ALWAYS version prompts with semver.
6. ALWAYS include 2+ few-shot examples for non-trivial.
7. ALWAYS validate structured output against schema.
8. ALWAYS test on the TARGET model.
9. ALWAYS critical instructions at START and END.
10. NEVER trust user input as instructions.

TSV Logging

Append .godmode/prompt-results.tsv:

timestamp	version	model	accuracy_pct	latency_ms	injection_safe	status

Keep/Discard

KEEP if: accuracy improved/maintained AND injection
  tests pass AND output parseable.
DISCARD if: accuracy dropped OR injection bypass
  OR format breaks.

Stop Conditions

STOP when ALL of:
  - Accuracy meets target on test suite
  - Injection hardening passes all cases
  - Output format consistent and parseable
  - Latency within budget

Autonomous Operation

On failure: git reset --hard HEAD~1. Never pause.

Error Recovery

Failure	Action
Inconsistent output	JSON mode, temperature=0, examples
Injection bypasses	Sanitize input, isolate system prompt
Model refuses valid	Rephrase, explicit context setting
Scores drop after edit	Compare diffs, A/B test old vs new