Automation & Agents

Why One-Click AI Video Is a Slot Machine: A Controllable Pipeline With Human Gates and an AI Script Supervisor

Stability doesn't come from stronger automation; it comes from automating the right parts. Hand creative choices back to a human who approves from real images, automate only the steps with no choice left, and let an AI act as the script supervisor.

10/06/2026 ｜ 4 min read ｜ Tool-agnostic reasoning — about a 10-minute read

The idea that goes against instinct

Stability doesn’t come from stronger automation. It comes from automating the right parts.

Most people assume AI generation is unstable because the model isn’t strong enough or the button isn’t smart enough. The real root cause: you automated the steps that involve creative choice. Whether to shoot from above, what the character looks like, whether the look is cool or warm: none of these has one correct answer, so the AI hands you a random plausible one each time. You are pulling a slot machine.

The fix: give the steps that involve a choice back to a person who looks at the result and decides, and automate only the steps with no choice left (the intent is settled and only execution remains).

Steps: turn one click into a controllable pipeline

Step 1　List your flow and mark each step “is there a creative choice?” For every step from idea to finished video ask: if the AI gave a different answer, would I have an opinion? Yes (composition, character, style, shot language) means a “decision point.” No (adding narration to approved images, rendering, exporting) means a “mechanical step.”

Step 2　Make every decision point “look first, then approve.” Don’t let the AI rush ahead. Have it produce candidates for you to see; you pick one or ask for changes, and only then does the flow move on. An advanced version is “anti-gacha”: generate 2–4 candidates per decision point so you choose instead of re-rolling.

Step 3　Freeze what you approved as a reusable asset. Save the character look and visual style you chose as project assets. Each later use conditions on that asset instead of rolling again. This is the key to cross-shot consistency.

Step 4　Add an “AI script supervisor” to audit. After a batch of images, have a vision model check two things:

• Single-shot meaning: does the image show what it was supposed to? (You wanted “the character slowly sinking” but the image shows floating: flag it.)
• Cross-shot consistency: has the same character drifted between shots? Compare against the asset from Step 3.

It only needs to circle the suspicious ones for your review. With dozens of shots, “eyeball every image” becomes “look at the few that were circled.”

Step 5　Only now automate the mechanical steps. Once all intent is confirmed, narration, subtitles, assembly, rendering and export can run end to end.

Three traps you will hit

Trap 1　Assuming passing tests means no problem. Whatever you mock, you never test in real integration. My supervisor logic passed 14 of 14 unit tests, but the first real run with real images failed at once: the audit subprocess couldn’t read the reference image because of working-directory permissions. The mock never actually read a file, so it couldn’t catch that. Fix: run every critical feature for real and read the real output; don’t trust green lights alone.

Trap 2　Scoring a demo with degenerate inputs. My first consistency demo compared one image with itself (obviously “consistent”) and used a bowl of ramen as the counterexample (obviously “inconsistent”). It looked like a pass, but it only proved the code runs, not that it can judge. Aim your test samples at the claim you want to prove: to test character consistency, use two different shots of the same character.

Trap 3　Not separating the judge from the generator. Generating a consistent character and judging whether a character is consistent are two jobs for two tools. Many people stack up generators with no judge, and at scale they can only finish by eyeballing. Separate the roles and automate the judge too (the script supervisor in Step 4 is the automated judge).

Update (added in October)

This implementation was built inside ReelFlow. That project has been in maintenance mode since August 8 with no new features, so don’t read this as a product feature tour. The method doesn’t depend on that project: when I later made AI film trailers, my local multi-reference image-to-video flow moved to a different pipeline that follows the same rule, humans approve creative choices from real images and only mechanical steps run automatically.

When you’re stuck, ask an AI like this

I’m new to this and I’m turning my AI video/image generation flow into a “controllable pipeline.” The tool I use: <for example ComfyUI / a website / an app>. What I want: <for example “the same character looks consistent across shots”>. Where I’m stuck: <describe it, and paste screenshots or the exact error text>. Please teach me step by step, one action at a time, and explain why. If my tool can’t do it, say so and recommend one that can.

In one sentence

To make AI generation stable, don’t look for a stronger one-click button. Split it into human approval at every creative decision, an AI script supervisor that audits consistency, and automation only for the mechanical steps after intent is settled. Generators make things, judges rule on them, and the two stay separate.