Skip to content
SJ
All projects

Private

Routecut

Writes a trip's shooting script leg by leg, and only writes shots the gear in your bag can take.

Python / FastAPI / Flutter / SQLite / Claude


How the gear bag decides which shots get written

Routecut writes the shooting script for a trip leg by leg, then the shot list, the teleprompter and the YouTube pack. The built-in writer only writes a shot if something in the gear locker can take it: each shot style maps to a capability, so with the drone left behind no aerial shot or high reveal appears, and without a microphone nothing depends on clean dialogue audio.

The forecast feeds the same decision, so a leg with a 60% chance of rain gets a covered backup. Model-written legs get both rules in the prompt, and a shot naming gear that is not in the bag is re-tagged before storage.

What happens when no model is available

Three interchangeable script writers, Anthropic, Amazon Bedrock and OpenAI, sit behind a preference order, and behind all of them is a built-in deterministic writer. With no credentials configured the app still produces a script, offline, and the tests and evals run without spending anything.

How model output is checked before it is stored

Everything a model returns passes a guardrail layer before storage, and each finding is either repaired or fatal. Beat lengths outside 3 to 180 seconds are clamped and runaway output is trimmed, while hype phrases, leaked prompts, the wrong script for the trip's language (Devanagari in a Hinglish script, or the reverse) or a leg more than 2.5 times its target runtime drop that leg back to the built-in writer. Trip names and notes are treated as data, never as instructions.

What the client and the audit cover

The Flutter client has nine screens across phone and desktop layouts. Its teleprompter reads each beat aloud in the trip's language at the pace the script was written for, or scrolls at that pace without a voice, and the export screen downloads a script PDF, a shot-list CSV and an SRT per leg. An audit recorded 18 security, performance and technical findings, each reproduced rather than inferred: 17 were fixed in code, including missing authentication, CSV formula injection in the shot-list export and an N+1 query on the trips list, and one was closed as a product decision.

Writing that draws on this project

Start a project

Describe what you are building and where it is stuck. Rough notes are fine.