Checks at every stage

For Legislative Awareness Week 2026, I needed short videos pointing chapters to the Playbook. We had two people and a simple camera setup.

I limited the editing work to free tools. The videos also had to stay non-partisan, so my brief told AI to flag any shot that raised a concern instead of using it.

A typical first prompt sounds like this illustration: "Make me an animation showing off our tool and what it can be used for." I set up the prompt so I could check at every stage that my vision came through.

1. Write rough concepts and turn them into shot-by-shot storyboards.
2. Inventory the footage and log usable moments.
3. Build a paper edit, then stop for my approval.
4. Build the draft, then stop for my notes.
5. Render the final, with measured checks on sound, captions, and text placement.

Compliance: no partisan or endorsement framing (midterms Nov 3); service members in uniform must not imply Department of War endorsement; flag any shot that raises a concern instead of using it. [...]

PROCESS — do these in order and stop where indicated: [...]
3. Paper edit: a table mapping each narrative beat to clip, in/out timecodes, on-screen text, and music beat. STOP and wait for my approval. [...]
5. Draft: render a low-resolution preview plus a contact sheet of the result. STOP for my notes. [...]

QA BEFORE YOU HAND ANYTHING TO ME — report results, not assurances: [...]
- Captions checked against the audio; any uncertain words listed
- Everything readable with sound off

Tell me plainly anything you can't do with these tools and the closest workaround.

I used Claude Design with Opus 5.5. I chose Opus because it was efficient with my usage limits and delivered the results I needed.

My review notes on the animation flagged zooms cutting off key text, on-screen text timing that needed fixing, the wrong clock format, sound effects that were too loud or harsh, and white lines glitching in one scene. I also rewrote the on-screen message in one scene myself.

I also use AI for bill analysis, hearing prep, member materials, and our weekly newsletter. On policy, I directed the Duty Status Explorer with AI, working from the bill text and the RAND report. On communications, I draft our weekly Military & Government Brief with AI from a master file, built to paste into VoterVoice as HTML. Before it goes out, I check the facts and make sure it stays non-partisan. The video prompt set stages I could approve, and my Playbook prompt set outside tests as pass criteria.

AFSA animation on Facebook

What I did: I wrote the brief, reviewed each stage, and made the final calls.

What happened: The final video took five drafts and runs about 53 seconds.

The point isn't catching AI making a mistake. It's setting up the work so I can check that I've communicated my vision and AI can help efficiently.