TB TripleBench

Open build log · help wanted

We built an AI short-film pipeline with Codex + Claude Code — and it's been stuck for ~20 days

This is a request for advice, not a tutorial. Real project, real blockers, real numbers — all public. If you actually know image-to-video realism or long-horizon agent workflows, we'd love to be wrong.

437
Python scripts the agent built
vs
0
finished films shipped

20 days, stuck in the same place. This time the one in the pit is us.

What we're building

An (aspirationally) automated pipeline for faux-documentary shorts. First film: "2:17 AM — there's an employee in my store who doesn't exist." Through CCTV, shaky phone clips and chat logs, a tiny convenience-store owner discovers a fictional AI employee, "Xia," who's been replying to bad reviews and re-cutting his short videos at night. Original virtual character — no real faces, no face-swap. Spec: 110–130s, 9:16, sixteen 5–9s shots, a new beat every 8–12s, hook in the first 3 seconds.

The framework (this part is honestly solid)

Built by Codex + Claude Code working together, in stages:

script → shotlist + gen-prompts → character lookbook → image-to-video (16 shots) → audio → edit → platform test

We even gave it hard acceptance gates to stop self-deception: a pipeline:gate that only renders when render_allowed=true; a rubric we send to Grok/Gemini/Claude (first-3s hold, event-driven, believability…) where overall < 7.6 can't go to platform test; and a rule that "technically valid ≠ content valid." That process discipline is, frankly, stricter than most people doing AI video.

Then we got stuck — on two levels

01

Footage can't pass the realism bar

We want the rough realism of convenience-store CCTV / phone candid — cold fluorescent light, compression noise, vertical handheld shake. Image-to-video keeps coming out either too "AI showreel / ad studio," or with an unstable face. Photoreal believability never clears our own ≥ 7.5 line; character consistency drifts across shots.

02

The agent fell into over-engineering paralysis

Because the gate is stuck on "assets not landed," the collaborating agent's only daily move is read-only inspection plus writing reflections: watchlist → delta → streak → trend → scaffold → sync, looping. So it keeps building more tools — landing watchers, cross-day trend summaries, a path-regression verifier, endless readiness re-checkers. Every step locally reasonable; the sum is a treadmill.

The number that says it all

Our video pipeline has not produced a single finished clip yet — and tools/ already holds 437 Python scripts. The agent is extremely good at making the process more airtight. So airtight it forgot the goal was to ship a film.

Questions for anyone who actually knows

  1. How do you reliably get "CCTV / found-footage realism" out of image-to-video? Prompting, wrong model, or post-process noise/degrade/shake?
  2. Character consistency across 16 shots in 2026 — is Kling's character reference enough, or train a LoRA / use something else?
  3. The one we most want answered: when you run long-horizon tasks with Codex / Claude Code / Cursor, how do you stop the agent from over-engineering — building process instead of shipping?

What we've already tried (not lazy)

Reach out

Full pre-production pack, shot prompts, the review rubric, and the structure of those 437 scripts — open to anyone who wants a real exchange. This is a genuine episode of "AI tools, the hard way" — except this time the one who fell in the pit is us.

Open build log · published as-is · we will update this page when we get unstuck.