sessionmovie 🎬

Turn your Claude Code session into a 45–60 second movie.

Your agent just spent two hours hunting a bug — exploring files, failing tests, finding the fix. That session has a story arc: struggle, insight, resolution. sessionmovie reads the transcript and renders it as a snappy, beat-synced, shareable film. No screen recording. No editing. One command.

you:    /movie
claude: 🎬 rendering "the-login-bug.mp4" ... done (58s, 9 files, +412/−118)

Status: v0.3 on npm (npx sessionmovie). The design docs in docs/ remain the source of truth for what's built next. Every feature launch ships with a movie of the feature being built — made by the tool, obviously.

How it works

There is no screen capture. Every frame is motion graphics rendered from transcript data — a stylized reconstruction of the session, like the replay screen in a sports game:

transcript JSONL
  → Parser        deterministic: events, diffs, test results (+ secret redaction)
  → Screenwriter  LLM: picks the beats, writes a screenplay (genre-neutral JSON IR)
  → Genre pack    strategy: renders the screenplay in its visual style
  → Remotion      React-driven video → MP4

Because frames come from data, not pixels: crisp typography at any resolution, vertical 9:16 that actually works, and total control of time — two hours compress into sixty seconds because we animate events, not video.

As a Claude Code skill, the LLM step runs inside your existing session — no API key, no server, no marginal cost. Rendering is local CPU via Remotion. A movie costs approximately $0.

What the movie looks like

There's no genre costume — one look, dark terminal canvas with a coral accent (see docs/visual-language.md). Title card, plainly stated (no cold open) → a few rounds of dialogue → one real artifact each, one code window / terminal / stat card chrome shared by every scene → an optional showcase of the biggest edit when the session earned one → a stats card built from real numbers: "2h 14m → 58s. 9 files. +412/−118."

Dialogue scenes star two pixel-art characters — you and Claude — with real (condensed) lines from the session and an emotion picked per beat. The agent is an original pixel-art homage to Claude Code; you're the same rig topped with your own GitHub avatar, pixelated and tinted to match. See docs/characters.md.

Genre packs — a future extension point

classic is the only shipped pack and the permanent fallback; there's currently no second genre to pick between. The GenrePack interface (a costume: React components + a music track + an SFX map + a caption persona) stays in the tree as the seam a future pack would plug into — see docs/genre-packs.md.

Usage

npx sessionmovie doctor                        # first-run check (downloads the render browser once)
npx sessionmovie path/to/session.jsonl         # vertical, ~50s, no genre to pick
npx sessionmovie session.jsonl --no-llm        # fast heuristic screenwriter, no claude call

Claude Code transcripts live in ~/.claude/projects/<project>/*.jsonl. The screenwriter step shells out to your existing claude CLI — no API key needed; --no-llm skips it entirely.

The Claude Code skill (/movie — movie of the current session) is the planned distribution: CLI first, skill as a thin wrapper — so any agent with transcripts is a future market.

Voiceover (optional)

Narration via ElevenLabs, on your own API key, of dialogue lines only — title/action/stats scenes stay silent (see docs/audio.md). Off by default — nothing calls ElevenLabs unless you pass --voiceover.

export ELEVENLABS_API_KEY=sk_...   # from elevenlabs.io → Developers → API keys
npx sessionmovie doctor            # validates the key before you spend a render on it
npx sessionmovie session.jsonl --voiceover

Key permissions matter. A default full-access key just works. If you create a restricted key, it needs at least Text to Speech (synthesis) and Voices: Read (what doctor probes) — a fresh key that fails doctor with HTTP 401 is almost always missing scopes, not a bad key.

Knobs: ELEVENLABS_VOICE_USER/ELEVENLABS_VOICE_CLAUDE give each speaker their own voice, ELEVENLABS_VOICE_ID is the single-voice fallback for both, ELEVENLABS_MODEL picks the TTS model. Synthesized audio is cached content-addressed, so re-renders of the same screenplay don't re-bill. Cost: roughly $0.10–0.30 per movie-minute of narration.

Docs

| Doc | What's in it | |---|---| | architecture.md | The four-stage pipeline and why the boundaries sit where they sit | | screenplay-format.md | The genre-neutral JSON IR — the project's load-bearing contract | | genre-packs.md | The GenrePack interface — a dormant extension point until a second pack ships | | visual-language.md | The no-genre look, the shared panel chrome, the energy kit, the stats card | | audio.md | Music/SFX rules (CC0, beat grids) and the dialogue-only ElevenLabs voiceover design | | characters.md | The two puppet characters — pixel-art mascot homage, avatar-head user, shared clip rig | | v1-storychange.md | Recognition first, dialogue is documentary, persona = tone, one voice at a time | | distribution-and-cost.md | CLI vs. skill, cost model, first-run UX | | security-and-privacy.md | Secret redaction — a v1 blocker, not a nice-to-have | | roadmap.md | The three-weekend v1 plan and the v1.1 candy |

License & credits

MIT. Bundled music/SFX are CC0 only, tracked in CREDITS.md.

Note: Remotion has its own license — free for individuals and companies of up to 3 people; larger companies need a Remotion company license to render.