# Start with a conversation

An AI filmmaking workflow tailored to DJI drone footage. You direct; your agent handles creative proposals and command-line production. This is not positioned as a general-purpose video editor.

You need a local-capable AI Agent, Python 3.9+ and FFmpeg/ffprobe with libx264,
libx265 and AAC. No Python packages, video editing app, installable skill, model API
key or mandatory Windows server is needed by the scripts. Your Agent service has
its own access/account requirements. The repository is an execution guide, not an
AI service: without local file and command access the client cannot operate footage.

## First run

Tell the Agent: **“Read README.md and AGENTS.md in this repository. Help me prepare
the dependencies, then import this DJI card into a new project. I will choose the
usable clips and time ranges, and we will discuss the story before you edit.”**

If local system approvals are needed, the Agent should explain them. Installation
examples (inspect existing installations first; package IDs can change):

- macOS with existing Homebrew: `brew install python ffmpeg`
- Arch Linux: `sudo pacman -S python ffmpeg` (maintain a consistent system update state)
- Debian/Ubuntu: `sudo apt install python3 ffmpeg`
- Windows: install Python from python.org and FFmpeg from the builds linked by
  ffmpeg.org/download.html, or inspect `winget search Python.Python` and
  `winget search Gyan.FFmpeg` before installing a selected package. Reopen the terminal
  and use `py -3`/`python` if `python3` is unavailable. Confirm the real executable,
  not the Windows Store alias.

`python3 tools/aerial.py doctor` reports actual tools/filters/encoders and exits nonzero
when the required baseline is missing. No automatic system configuration is installed.

## Check a clean machine without private footage

```sh
python3 tools/demo.py /absolute/path/to/a-new-demo
python3 -m unittest discover -s tests -v
```

The destination must not already exist. The synthetic demo creates a pretend card,
imports unchanged media, records selections, renders a short preview and final
10-bit video with a generated test tone and selectable subtitles, and extracts a poster.
Use `--4k` for a real short 3840×2160 final. The test tone is not a soundtrack proposal.

## Your real project

Insert the card and ask the Agent to import it. Then tell it the filenames, time ranges,
scene context and any must-keep/excluded actions. Discuss the story, watch previews,
listen to music proposals and ask for changes. You never need to hand-edit JSON or
run the commands yourself; your Agent uses the execution contract (in starter: docs/agent-workflow.md).
Decisions and commands persist in the project, so another conversation can resume it.

Card insertion does not launch a background watcher. Card discovery lists DCIM-bearing
candidate mounts; you/your Agent still verify identity and import scope. DCIM/MISC
import is not a full disk image or proof of redundant backup. Originals are retained.

## Compatibility and limits

The scripts use Python standard library + FFmpeg and no branded Agent transport.
OpenClaw, GPT Dot, Grokbot, Claude Code and Codex CLI are candidate entry points only
when their configured local execution permissions support the contract. A brand name
alone is not a verified integration. See `compatibility.md` for the measured platform
scope, not an “all computers supported” promise.

The base renderer is an SDR cut renderer, not automatic grading, a music discovery
service or a universal Log/HDR converter. Advanced project-specific command-line work
is described separately. Software HEVC can be slow at 4K. Music/license and emotional
fit need Agent research and human listening. Optional archive/cinema (in starter: docs/optional-integrations.md)
never gates local creation. Code and docs are MIT; third-party assets are excluded.