Paper for people who write.
A scriptwriting hideout for people who make reels for a living,
where the machine does the reading and you do all the writing.
It's 1:47 a.m.
You have a car brand's content due Monday, and you know you saved the perfect reel three weeks ago. The one where the guy just stares at the dashboard for a full second before saying anything. You scroll through Saved. Four hundred videos. No search bar. No folders you actually used. Half the thumbnails are grey now because the creators deleted them.
So you do what everyone does. You open a blank doc and write the same "Stop scrolling if you drive a…" hook you wrote last week.
Foolscap exists so that never happens again.
Foolscap is the old paper size writers used for centuries, long before screens. It's also said to be named after the jester's cap that once watermarked the sheets. A page for people who write, with a fool's cap in the corner as a reminder that the funniest, sharpest stuff always came from a person willing to look a bit ridiculous.
That's the whole app. Hoard references like a lunatic, then write your script with your own hands, on a clean sheet.
Scrape the world. Write it yourself.
There is no AI in the writing. Not a toggle. Not a "just polish this". Not a sneaky autocomplete.
AI in Foolscap does the jobs nobody should be doing at 1:47 a.m.: saving, transcribing, measuring, remembering, noticing. It reads. The second a cursor blinks in the writing room, the machine is out of the building.
Why so strict? Because people are tired of slop. Gartner found half of US consumers say they'd rather buy from brands that keep generative AI out of customer-facing content. The scarce thing on the timeline now is a person who is visibly, stubbornly human. We're building for that person.
AI reads. Humans write.
Easiest way to explain it is to walk through a day.
9:00. New client: a car brand. You make a workspace, drop in their tone, the words they never say, the disclaimers legal makes you add, and the model specs. This is the Brand Bible, and it quietly checks every script you write against it. Plain rules, no AI judgment. The client's lawyer reads your script before you do.
9:30. You go hunting. You find a reel on Instagram, TikTok or YouTube, record it (or hit the one-tap record button), and it lands in Foolscap with the link attached. It keeps the footage, so when the creator deletes the original, you still have it. It pulls the transcript, the on-screen text, the caption. It even copes with that Urdu / Roman Urdu / English mashup that real people actually speak. Everything is searchable by anything said or shown. If you saw it, you own the receipt.
10:15. You open X-Ray on the reel you love. It breaks the video into beats: how long the hook lasts, how fast the cuts come, when the text appears, how many words a minute, where the pauses are. It tells you how far this video outran its own account's normal numbers. And there's one field it will never fill in: "Why it worked." That one is yours. Know exactly why it worked. Then do it differently.
11:00. You find the moment. Three seconds where the creator does a slow, deadpan blink at the camera and the whole video turns. You clip it into the Expression Vault, tag it deadpan, and pin it to line 4 of your script. Now your script doesn't just say what to say. It shows the face. Don't tell them what it should feel like. Show them the face.
11:30. The Writing Room. Left column is what we hear. Right column is what we see. Your reference sits pinned beside the draft like a friend at the next desk. A live timer reads your script at your speaking pace and tells you it's 34 seconds, and you wanted 30. You write ten hooks on the Hook Bench and keep the one that makes you grin. If your draft starts drifting too close to a saved reference, the Overlap Alarm taps your shoulder. You can steal a feeling. You can't steal the words. A blank page is scary. A blank page with autocomplete is worse.
14:00. Send the pitch. One link: script, reference clips, shot list. The client comments at the exact second they have a problem, you see a revision counter instead of a vibe, and "approved" actually locks the version. Works great where approvals live on WhatsApp. Approve it before you film it.
Friday. Shoot day. Set Mode is the script for the real world: huge type, works offline, mirrors for a teleprompter, scrolls as you speak. You mark good takes against the exact line they belong to.
Next month. The post goes live, and Foolscap ties it back to the script that made it. Day one, week one, month one. You write your own retro notes, and over time it shows you what actually wins for this client. Dead ideas go to the Script Graveyard, where they can be resurrected. Every script deserves an after-party.
Some things are further out but they're the ones I'm most excited about.
- Research Radar. Track competitors, dealers, creators, and neighbouring niches. Get pinged when something outperforms. Mine comment sections for the questions and objections customers keep repeating, then you pick which ones become angles. There's also a Format Arbitrage Board for formats crushing it in other niches that nobody in yours has touched yet. Let the machine do the scrolling. You do the thinking.
- Human Ink. A timestamped authorship history you can export as a one-pager for brands that need proof a person wrote it. Proof of pulse.
Here's the thing nobody tells you: a link is just an address. The machine can't watch an address. It needs the actual footage. So Foolscap takes reels in two ways, and keeps the link either way.
Record it. Play the reel, screen-record it with your phone's built-in recorder, and share the video into Foolscap. That's v1, and it needs zero custom magic. Later there's a one-tap Record button inside the app that does it for you.
Paste the link too. The link stays on file for the source, for spotting duplicates, and for showing the original embed. It just isn't what the analysis runs on.
A few things worth knowing:
- Android records through MediaProjection and iOS through ReplayKit. Recording other apps on iOS needs a Broadcast Extension.
- Audio is the real test. Android 10+ can capture it internally, but some apps can block it. Test on a real phone with real reels before promising anything.
- Gemini only samples one frame a second, so a screen recording is plenty for X-Ray. It's the Expression Vault clips that want the best quality.
- Recordings carry the app's buttons and captions, which muddy the on-screen text. A quick crop step fixes that.
- YouTube is the exception: public links work directly.
Record it. Keep the link. Own the receipt.
Plain-English version. The Flutter app catches the recording you share (or record) and uploads it. A plain Node.js server — just a regular backend, no Cloud Functions, running wherever's cheapest — picks it up, hands the video to Gemini, and gets back a strictly shaped bundle of facts: timestamps, hook type, cut counts, labels. All of it lands in Firestore. The writing room reads from there.
flowchart LR
A["📱 Flutter<br/>record or share"] --> B[("Cloud Storage<br/>snapshots")]
B --> C["⚙️ Node.js server"]
C --> D["🔭 Gemini<br/>analyst only"]
D --> E[("Firestore")]
E --> F["✍️ Writing Room<br/>zero AI"]
F --> G["📎 Pitch links"]
The AI firewall. The no-AI promise isn't a setting, it's architecture:
- Gemini only gets called from the backend server. The key never ships in the app.
- Every call uses a strict schema made of enums, numbers, timestamps and short labels. There is no field that can hold a generated script.
- The writing room doesn't import an AI client at all. Not hidden. Not off. Absent.
- Every edit is logged with its author, which is where Human Ink comes from.
Rough data shape:
workspaces/{wid}
├─ clients/{cid} brand bible: tone, banned words, disclaimers
├─ references/{rid} source, snapshot, transcript, beats, hook type,
│ outlier score, expression clips
├─ scripts/{sid} title, format, status, linked references
│ └─ versions/{vid} rows, author, timestamp
└─ pitches/{pid} script, references, comments, approval
Let's be real about what's hard. None of it scares me, and I'm going to work through every one of these.
- The video part is solved. Gemini can genuinely watch footage: files up to 2 GB through its Files API, small clips inline, public YouTube links directly. A minute of reel costs roughly 100 to 300 tokens per second depending on resolution, which is pocket change.
- The link part isn't. A TikTok or Instagram link is just an address. Gemini won't pull from arbitrary links, and Meta's embed API only hands back embed code. So v1 goes "record it, share the video in", and YouTube links work as they are.
- Bulk scraping is its own beast. Platforms ban automated access in their terms, so the Radar will run on licensed data or official routes, not a scraper held together with duct tape. Legal advice comes before launch, not after.
- Firebase wants the Blaze plan for Functions and Storage. Budget alerts go on day one. Future me says thanks.
- Roman Urdu mixed with English is the real exam. No standard spelling, and everyone switches languages mid-sentence. I'll test it on real clips before I trust it, and before I promise it.
- The Library. Capture, snapshots, transcripts, X-Ray, folders, search.
- The Room. Writing room, timers, direction tags, pinned references, Expression Vault.
- The Pitch. Shareable links, comments, approvals.
- The Guardrails. Brand Bible, linter, pillar heatmap.
- The Radar. Licensed data, outlier alerts, comment mining.
- The Set. Teleprompter, take log, results loop, Human Ink.
The machine reads. You write. The audience can tell.
📜