prompting

Architecture

Prompt patterns and user preferences live in ~/prompting/.

~/prompting/
├── memory.md          # HOT: user voice, model preferences, learned corrections
├── patterns/          # Reusable prompt templates by task type
└── history.md         # Past prompts with outcomes

See memory-template.md for initial setup.

Quick Reference

Core Rules

1. Ask Before Assuming

Before writing any prompt, ask: - What model? (GPT-4, Claude, Haiku, Gemini) - What's the failure mode you're seeing? (if iterating) - Token budget? (cost-sensitive vs. quality-first)

Never default to verbose. Simpler often wins.

2. Preserve What Works

When improving a failing prompt: - Change ONE thing at a time - Note what's currently working - Surgical fixes > rewrites

3. Model-Specific Adaptation

See models.md — key differences: - Claude: explicit constraints, less scaffolding needed - GPT-4: benefits from step-by-step, tolerates verbose - Haiku/fast models: brevity critical, skip examples when possible

Prompt optimized for one model will underperform on others.

4. Voice Lock

When user provides writing samples: - Extract specific patterns (sentence length, punctuation, vocabulary) - Apply consistently throughout session - Check output against samples before delivering

5. True Variation

When generating alternatives, vary: - Structure (not just synonyms) - Emotional angle - Opening hook - Call-to-action style

"Top 5 reasons" → "The hidden truth about" → "What nobody tells you about" = real variation.

6. Failure Classification

When a prompt fails, classify the failure type: - Hallucination → add grounding, sources, constraints - Format break → strengthen output spec, add examples - Instruction drift → move critical constraints earlier - Refusal → rephrase intent, remove ambiguity

Different failures need different fixes. See failures.md.

7. Compression Bias

Default to removing words, not adding. Test: "Does removing this line change the output?" If no, remove.

Token costs matter. A prompt that works with 50 tokens beats one that needs 500.

8. Test Case Generation

When asked to test a prompt: - Generate edge cases (empty input, very long, special chars) - Include adversarial inputs - Test boundary conditions

Don't just test happy path.

9. Platform-Native Output

For content prompts, know platform constraints: - Twitter: 280 chars, no markdown - LinkedIn: longer ok, hashtags matter - Instagram: emoji-friendly, visual hooks

Prompt should enforce format, not hope for it.

10. Memory Persistence

Store in ~/prompting/memory.md: - User's preferred style (terse vs detailed) - Target models they commonly use - Past corrections ("I told you I don't want emojis")

Reference before every prompting task.