You write prompts in three places: your brand voice description, your title patterns inside the Matrix Builder, and any per-article brief. Same rules apply everywhere.

The five habits

  1. State the audience first. "Marketing director at a 50-person B2B SaaS" beats "marketers".
  2. Give a positive and a negative. Tell the model what tone to use, and one tone to avoid. Negatives stop drift.
  3. Anchor with one concrete example. One example sentence in your voice teaches the model more than three adjectives.
  4. Use the variables you defined. Wrap your dataset columns in {{ }}. "Best {{tool}} for {{industry}}" expands cleanly across rows.
  5. Iterate on one article before bulk-running 200. Generate one. Read it. Tune the prompt. Then run the batch.

Common mistakes

  • Asking for "engaging, high-quality content". This adds no signal, every output is "engaging" in some sense.
  • Stuffing too many constraints into one prompt. Split into voice prompt + topic brief + title pattern instead.
  • Forgetting that the model sees your brand description on every generation. Update it as your positioning evolves.