10 Ways H3 Prompts Fail (And Easy Fixes)

Use this as a pre-flight list. Most “the model is broken” reports trace back to prompt structure, not weights.

---

1. One-line vibe prompts

Symptom: mushy subject, random camera, nonsense audio.
Fix: Add place, timed action, and two sound lines. See plain-English formula.

---

2. Reference files with no job description

Symptom: product morphs, wrong face, ignored logo.
Fix: [References] block — “@image1 = lock label text NOCTURNE.” Ref2VA guide.

---

3. One image asked to do everything

Symptom: face drifts when product is correct, or vice versa.
Fix: Split talent vs product vs mood across files.

---

4. Vague camera words

Symptom: chaotic or static when you wanted the opposite.
Fix: “Truck left 2m, slow push-in” beats “dynamic orbit.”

---

5. Expecting montage in FL2VA

Symptom: weird in-between frames.
Fix: FL2VA = continuous path between keyframes; use T2VA or Ref2VA for cuts.

---

6. Visual-only prompts

Symptom: random music, wrong foley.
Fix: Always fill soundscape + music (or N/A). Audio guide.

---

7. Dialogue overload

Symptom: lip-sync mush, rushed ending.
Fix: One short line per 10–15s clip.

---

8. Wrong mode for the asset pipeline

Symptom: label text wrong in T2VA; wasted time tuning text.
Fix: Pack shots → Ref2VA.

---

9. Copy-paste without editing references

Symptom: prompt mentions @image4, you uploaded two files.
Fix: Match reference count and rewrite the [References] block.

---

10. Chasing “official magic words”

Symptom: stuffing jargon without changing structure.
Fix: Keep timed actions + sound split; official field names are optional labels.

---

Healthy workflow

1. Pick mode
2. Draft structure in template Adapt section
3. Run once
4. Change one variable (camera or sound or timing)
5. Save your version in a doc — that’s your real library

Official docs: MiniMax prompt gallery · GitHub H3