Launching today

Cinematographly Studio
AI director & producer for video studios, with a spend gate
0 followers
AI director & producer for video studios, with a spend gate
0 followers
Cinematographly takes short AI video from brief to export in Claude Code: script, storyboard, clips, voice, music and a timeline for Resolve, Final Cut or Premiere. A gate checks every paid call. Free for teams of 3 or fewer; studios pay $99 per seat.



Launch Team

AWS Startup AdvisorPersonalized AWS guidance that meets you where you’re already building
Promoted
Hi Product Hunt! I'm the maker of Cinematographly. I run a small video studio, and we use it on our own client work.
What it does
It's a Claude Code plugin that directs and produces short AI video. It writes the script and storyboard, makes start frames, clips, voice and music, and hands you a timeline for DaVinci Resolve, Final Cut Pro or Premiere Pro (experimental).
The spend gate
The part I care about most. Every paid Higgsfield or Kling call goes through a hook that needs an open job, a linted prompt and room in the budget. It also needs the stage before it approved in a person's own words. If the gate can't run, the call's refused.
For studios
A profile per client
A price for each job before anything's made
A monthly credit ceiling
Review packs and delivery folders per platform
Try it free
Install it and run /cinematographly:demo: nothing is spent. It's free for teams of 3 or fewer, and $99 per seat above that. In Claude, Cursor and ChatGPT it runs as a skill, but the gate isn't automatic there yet.
I'd love feedback on the gate, and on what a studio would need next.
Under the hood, for the technical reader
Parts: one skill, 110 Python scripts, 10 slash commands, 8 subagents and a set of hooks. SKILL.md is the agent's rulebook, and 58 reference files are loaded stage by stage, only when that stage needs them.
The hook: the spend gate runs as a PreToolUse hook. hooks.json routes every Higgsfield, Kling, Scenario and DaVinci Resolve tool, plus any claude.ai connector matched by its id, to gate.py. Each hook's got a 20-second timeout and onFailure set to block, and gate.py exits 2 for older Claude Code. So a crash, a timeout or an empty answer refuses the call.
Four checks before a paid call: a job must be open (a checkpoint file that gate.py open has validated); each stage before it must be approved in the approver's own words, found in this session's transcript; the prompt must have passed prompt_lint.py; and the call must fit the job's credit envelope.
Other rules: reads pass silently. Publishing needs the export approval. A Resolve script that starts a render or calls Resolve's AI voice or auto-captions isn't allowed. After each generate call, a PostToolUse hook records the request by hash, so an identical retry within 30 minutes asks first.
Ceiling and sessions: the monthly credit ceiling sums one file per spend event, grouped by UTC month; a call past it needs a person to confirm it. Each Claude session keeps its own job pointer and transcript path, so two windows never share a job.
Pipeline: 15 stages, from intake to delivery. Each writes an artifact into a fixed job folder, and the next gate checks that artifact. Picture lock's a one-way door. Independent units fan out to the subagents: research, drafts, critiques, shot prompts, scene continuity, take review and sound-effect spotting. Agents don't spend, approve or touch Resolve.
Quality checks measure the files: clip_qa.py, motion_qa.py, look_qa.py, mix_qa.py for loudness, lipsync_check.py, title_qa.py and srt_qa.py. A contact sheet is built before any draft is shown.
Edit: cut_plan.py places each cut from the measured motion of the take. timeline_export.py writes FCPXML 1.10 for DaVinci Resolve and Final Cut Pro, and FCP 7 XML (xmeml v5) for Premiere Pro, which is still experimental. On Resolve Studio 21.1.2 every clip imported frame-exact at 23.976 fps, and the exporter supports seven frame rates without drop-frame timecode.
Costs: prices come from a dated rate card. Each client has a cost file with a retake factor, a job is priced before anything is spent, and closing a job writes a scorecard: credits per kept take and per kept second.
Tests: the test runner checks 109 scripts: 5 suites and 89 self-tests. The public CI runs four jobs: the gate on the macOS system Python with the standard library only, the full suite on Python 3.12, claude plugin validate --strict, and REUSE licence checks.