完成任务16：编写 04_长期发展/环境保护.md

skills/video-use/install.md

@@ -0,0 +1,160 @@
+---
+name: video-use-install
+description: Install video-use into the current agent (Claude Code, Codex, Hermes, Openclaw, etc.) and wire up ffmpeg + the ElevenLabs API key so the user can start editing immediately.
+---
+
+# video-use install
+
+Use this file only for first-time install or reconnect. For daily editing, read `SKILL.md`. Always read `helpers/` — that's where the scripts live.
+
+## What you're doing
+
+You're setting up a conversation-driven video editor for the user. After install, the user drops raw footage into any folder, runs their agent (`claude`, `codex`, etc.) there, and says "edit these into a launch video." You do the rest by reading `SKILL.md`.
+
+Three things must exist on this machine:
+
+1. The `video-use` repo cloned somewhere stable.
+2. `ffmpeg` on `$PATH` (plus optional `yt-dlp` for online sources).
+3. An ElevenLabs API key in `.env` at the repo root (for Scribe transcription).
+
+And one thing must be true about the current agent:
+
+4. It can discover `SKILL.md` — either via a global skills directory (`~/.claude/skills/`, `~/.codex/skills/`) or via a `CLAUDE.md` / system-prompt import.
+
+## Install prompt contract
+
+- Do everything yourself. Only ask the user for things you cannot generate — the ElevenLabs API key, and confirmation before `brew install`.
+- Prefer a stable clone path like `~/Developer/video-use` (not `/tmp`, not `~/Downloads`).
+- The skill references helpers by bare name (`transcribe.py`, `render.py`). That works because SKILL.md and `helpers/` ship together — keep them as siblings when you register the skill.
+- After install, verify by running one real command against one real file. Don't declare success on file-existence checks alone.
+
+## Steps
+
+### 1. Clone to a stable path
+
+```bash
+test -d ~/Developer/video-use || git clone https://github.com/browser-use/video-use ~/Developer/video-use
+cd ~/Developer/video-use
+```
+
+If the repo is already there, `git pull --ff-only` and continue.
+
+### 2. Install Python deps
+
+```bash
+# Prefer uv if available; fall back to pip.
+command -v uv >/dev/null && uv sync || pip install -e .
+```
+
+`pyproject.toml` lists `requests`, `librosa`, `matplotlib`, `pillow`, `numpy`. No console scripts — helpers are invoked directly as `python helpers/<name>.py`.
+
+### 3. Install ffmpeg (+ optional yt-dlp)
+
+`ffmpeg` and `ffprobe` are hard requirements. `yt-dlp` is only needed if the user wants to pull sources from URLs.
+
+```bash
+# macOS
+command -v ffmpeg >/dev/null || brew install ffmpeg
+command -v yt-dlp >/dev/null || brew install yt-dlp     # optional
+
+# Debian / Ubuntu
+# sudo apt-get update && sudo apt-get install -y ffmpeg
+# pip install yt-dlp
+
+# Arch
+# sudo pacman -S ffmpeg yt-dlp
+```
+
+If `brew` / `apt` / `pacman` requires a sudo prompt, tell the user the exact command and wait. Do not invent a password.
+
+### 4. Register the skill with the current agent
+
+Figure out which agent you are running under, and register once. A symlink of the whole repo directory is the right shape — helpers/ needs to sit next to SKILL.md.
+
+- **Claude Code** (`~/.claude/` present):
+
+    ```bash
+    mkdir -p ~/.claude/skills
+    ln -sfn ~/Developer/video-use ~/.claude/skills/video-use
+    ```
+
+- **Codex** (`$CODEX_HOME` set, or `~/.codex/` present):
+
+    ```bash
+    mkdir -p "${CODEX_HOME:-$HOME/.codex}/skills"
+    ln -sfn ~/Developer/video-use "${CODEX_HOME:-$HOME/.codex}/skills/video-use"
+    ```
+
+- **Hermes / Openclaw / another agent with a skills directory**: symlink `~/Developer/video-use` into that agent's skills directory under the name `video-use`. If the agent has no skills directory, add a line to its system prompt / config pointing at `~/Developer/video-use/SKILL.md` (e.g. an `@~/Developer/video-use/SKILL.md` import in a `CLAUDE.md`-equivalent).
+
+If you can't tell which agent you're in, ask the user once: "which agent am I running under — Claude Code, Codex, or something else?" Then pick the right target.
+
+### 5. ElevenLabs API key
+
+Scribe (ElevenLabs) does all transcription. Without a key, nothing transcribes.
+
+1. Check existing state in this order and stop at the first hit:
+
+    ```bash
+    # a) env var already exported
+    [ -n "$ELEVENLABS_API_KEY" ] && echo "env"
+    # b) .env at repo root already has it
+    grep -q '^ELEVENLABS_API_KEY=..' ~/Developer/video-use/.env 2>/dev/null && echo "dotenv"
+    ```
+
+2. If neither is set, ask the user exactly once:
+
+    > I need an ElevenLabs API key for transcription (word-level timestamps, speaker diarization, filler tagging). Grab one at https://elevenlabs.io/app/settings/api-keys and paste it here — I'll write it to `~/Developer/video-use/.env`. Or if you already have it exported as `ELEVENLABS_API_KEY`, say "use env" and I'll skip.
+
+    When the user pastes a key, write it to `~/Developer/video-use/.env`:
+
+    ```bash
+    printf 'ELEVENLABS_API_KEY=%s\n' "$KEY" > ~/Developer/video-use/.env
+    chmod 600 ~/Developer/video-use/.env
+    ```
+
+    Never echo the key back in tool output. Never commit `.env`.
+
+3. Sanity check with a cheap, quota-free call:
+
+    ```bash
+    curl -s -o /dev/null -w '%{http_code}\n' \
+      -H "xi-api-key: $(sed -n 's/^ELEVENLABS_API_KEY=//p' ~/Developer/video-use/.env)" \
+      https://api.elevenlabs.io/v1/user
+    ```
+
+    `200` means the key works. `401` means the user pasted a wrong/expired key — ask once more and stop. Anything else (network, 5xx), move on and verify during first real transcription.
+
+### 6. Verify end-to-end
+
+Run one real thing. Prefer the lightest verification that still proves the pipeline is wired up:
+
+```bash
+python ~/Developer/video-use/helpers/timeline_view.py --help >/dev/null && echo "helpers OK"
+ffprobe -version | head -1
+```
+
+Full transcription test is optional at install time — it burns Scribe credits. Better to wait until the user hands you their first clip.
+
+### 7. Hand off
+
+Tell the user, in one short message:
+
+- Where the skill is installed (`~/Developer/video-use`).
+- That they should `cd` into their footage folder and start their agent there (e.g. `claude`).
+- That a good first message is: *"edit these into a launch video"* or *"inventory these takes and propose a strategy."*
+- That all outputs land in `<videos_dir>/edit/` — the repo stays clean.
+
+## Keeping the skill current
+
+- `cd ~/Developer/video-use && git pull --ff-only` pulls the latest code. The symlink auto-picks it up on the next run.
+- If `pyproject.toml` changed deps, re-run `uv sync` / `pip install -e .` after pulling.
+
+## Cold-start reminders
+
+- Symlink the **whole directory**, not just `SKILL.md`. The helpers need to sit next to it.
+- If `.env` exists but the key is empty, treat it the same as missing — don't assume existence means validity.
+- `ffmpeg` from static builds works fine. Any modern (≥ 4.x) build is enough.
+- `yt-dlp` is optional. Don't block install on it; install lazily the first time a user asks to pull from a URL.
+- Never run transcription as part of install verification unless the user explicitly asks — Scribe costs real money.
+- If the user is on Linux without a package manager Claude recognizes, print the manual `ffmpeg` install URL and wait rather than guessing.