Prompts

# Prompts

The Claude system prompt for this workflow lives in
[`system.md`](system.md). When you change it, follow this loop:

## A/B testing loop

1. **Duplicate the Claude node** in the workflow and disable the original
   so they run side by side against the same input.
2. **Change exactly one thing** in the candidate node (e.g. add a
   constraint, tighten the tone list).
3. Run the workflow on 10 realistic briefs that cover edge cases (very
   short features, unusual category, luxury tone, technical tone).
4. For each output, score 0–3 on:
   - **Format compliance** — did it parse? exactly 5 hashtags?
   - **Tone match** — does it feel like the requested tone?
   - **Grounding** — did it invent specs not in the brief?
5. Keep the variant with the higher total. Document the change and
   result in a `CHANGELOG.md` in this folder.

## Why this matters

Prompt changes are easy to ship and hard to roll back. Without an A/B
discipline, you end up with "prompt drift" where the prompt accumulates
fixes for one-off complaints and gets worse on average.

A typical mature workflow goes through 5–15 prompt iterations in the
first month, then stabilizes. Keep the full history in Git.

## Model considerations

The workflow uses `claude-sonnet-4-20250514`. When to swap:

- **→ Haiku 4** for high-volume, simple briefs (> 1k/day). About 4× cheaper
  but format compliance drops from 99% to ~90% — plan a retry loop.
- **→ Opus 4** for complex or long-form content (press releases, product
  manifestos). 5× more expensive but noticeably better at narrative.

Tracking `usage.output_tokens` in your logs is the cheapest way to spot
when a prompt is getting verbose and triggering a model upgrade by
accident.