The night the prompt became a production system
I wanted to test metaprompting outside a programming demo, so I built a music-video workflow in which the model wrote prompts, generated images, called video tools, sampled frames, and judged whether a shot should move forward. It even helped write the song through Suno. The bill was aggressive. The lesson was worth it.
Metaprompting in plain English
A normal prompt asks for an output. A metaprompt asks the model to design, inspect, or improve the instructions that produce the output. Add a quality gate and the workflow becomes a loop: define intent, draft prompt, generate, inspect, revise, then ask a human for the final decision.
Build the smallest useful loop
- Write a one-paragraph creative brief with audience, feeling, format, and constraints.
- Ask the model to turn the brief into a shot specification.
- Generate one cheap still image before paying for video.
- Inspect identity, framing, lighting, unwanted text, and continuity.
- Ask the model to explain the largest mismatch and revise only that part of the prompt.
- Stop after a fixed number of attempts and let a human choose.
The quality gate
Do not ask “is this good?” Give the reviewer a rubric: story clarity, subject consistency, composition, motion feasibility, safety, and cost. Require a pass, revise, or reject decision with evidence. The reviewer should never automatically publish or spend without limits.
What I would change
Use low-resolution tests, cap retries, and promote only approved stills into video. Metaprompting is powerful precisely because it can create many iterations; that is also how it can quietly set money on fire.
Software, links, and setup
- Claude Code for workflow orchestration.
- Suno for music ideation.
- OpenAI image generation or another image API.
- Python plus FFmpeg for frame sampling and contact sheets.
Prompt pack
Prompt architect
You are the prompt architect for a short visual sequence. Convert my brief into a standalone generation prompt. Include subject identity, setting, wardrobe, composition, lens, lighting, palette, action, and exclusions. Preserve facts from the brief and label any creative assumption.
Quality reviewer
Review these sampled frames against the approved brief. Score story clarity, identity continuity, composition, motion coherence, artifacts, and brand fit from 1 to 5. Return PASS, REVISE, or REJECT. Identify the single highest-leverage prompt change. Do not approve if the subject changes identity or if readable text is malformed.
Cost guard
Before generating, estimate the number of calls and maximum cost. Use one draft and no more than two revisions. Ask for human approval before any higher-cost render.
What to remember
- Use AI to improve the process, not only the final artifact.
- Review against a rubric instead of taste alone.
- Test with stills before paying for motion.
- Put hard limits around retries and spend.
Then test it, document it, and make the next version easier for someone else to reproduce.