# COOPR

Status: Unfinished

A creator workspace where the conversation turns into the document.

Drop in a few clips, talk it through, and the parts you accept land in an editable plan: the hook, the script, the shot list, the filming notes. Behind it is Content Brain, the analysis and memory the chat draws on.

- Role: Solo builder, directing coding agents
- Stack: React, TypeScript, Python, PostgreSQL, MCP
- As of: Jul 20, 2026, its last commit
- Commits: 8,741, from Feb 20 to Jul 20, 2026
- Recognition: $3,000 award from the University of San Diego, 2026
- Showcase repo: https://github.com/henryhcooperr/coopr

Commits by month, 2026:

- Feb: 474
- Mar: 1996
- Apr: 3127
- May: 1934
- Jun: 807
- Jul: 403

[Rebuild] COOPR’s Docs view: a draft of a sample wreck-dive episode with a suggested rewrite to accept or drop, and a chat of clip options on the right. COOPR’s Docs view with sample content, not a screen capture. The episode is the prototype’s own sample project.

## What I was after

I wanted one place to plan videos for my own accounts: compare takes, pick a hook, write the script, know what to shoot next.

My rule for the writing was simple. Scripts in my voice that can’t sound like AI. So drafts are scored against a style profile, and nothing goes into the document unless I accept it.

## Decisions

### Nothing is written unless I accept it

Send a clip to the doc, pick a hook, and accept or drop the rewrite Coopr suggests. Then switch formats: the same accepted material becomes an episode script, carousel slides or a short.

Those variants are plans to shoot from, not rendered videos.

Instead of a model writing straight into the draft.

[Rebuild] A COOPR document for a sample wreck-dive episode, beside the chat. The prototype runs this flow on fixture data, not a live analysis backend.

### Every answer is something you can act on

The chat doesn’t answer in paragraphs. Each answer arrives as flagged phrases to swap, slides to send to the doc, a checklist Coopr can finish, or two captions to choose from.

Instead of answers in paragraphs.

[Rebuild] COOPR’s chat with a prep checklist at 5 of 6 done. COOPR’s chat with sample content. The voice check stands in for its style-profile scoring, with sample flags rather than real scores.

### A topic changes when the drift holds

To cut a long talking clip into chapters, Content Brain looks for the point where what I’m saying actually shifts. It compares each phrase with the one before, and a CUSUM adds up how far those comparisons drift above normal: `S = max(0, S + d − μ₀ − k)`, with a break when S passes the threshold.

One odd phrase doesn’t trip it; a sustained change does. COOPR also has a separate PELT detector for regime changes.

Instead of a break at the first odd phrase.

[Sample data] Bars of phrase distances with a running total crossing a threshold line. Bars are each phrase’s distance from the one before. The line is the running CUSUM total. The loop raises the threshold; watch the break move.

[Sample data] A table of phrases and distances, with a break at Chapter 2. A sample dive narration with hand-set distances, run through the textbook one-sided CUSUM. The right column is S after each phrase.

## How it’s built and checked

[Screenshot] Three prototype screens: the Docs list, a document with a revision suggestion, and a shot list with prep. Screenshots of the prototype with its sample content.

- Workflows: Compare clips, write, check the voice and repurpose each exist in source, but not every handoff between them is proven. I haven’t verified that a ranked clip opens straight into Clip Lab.
- Retention model: A gradient-boosted regression with pseudo-label and scarce-label handling. It’s real ML code, and I’m not claiming any accuracy for it.
- Evidence: Retrieval falls back to other sources when the first one comes back thin.
- Interface: A prototype in its own design language, with eight workspaces: Home, Docs, Clip Lab, Library, Insights, Intel, Inbox and Calendar.

## Where it stands

### Exists in source

- Content Brain: audio and motion diagnostics, CUSUM and PELT change detection, and evidence retrieval with fallbacks.

### Left unfinished

- I moved on to Reef before the interface was wired to the live backend.

### Carried into Reef

- Reef reads Content Brain through a read-only, tenant-bound client that trusts nothing it returns raw.

Next: Grade Lab, /work/grade-lab/

Back to the homepage: /