juneyaooo/cheap-claude-video/capsule-cinema

Capsule Cinema

Use Capsule Cinema as a recipe-driven local video studio. Keep planning, generation, targeted rework, assembly, QA, and reusable production knowledge auditable in one workspace.

Resolve the runtime

Resolve the runtime before reading references or running commands:

1. If {baseDir}/scripts/capsule.py exists, use {baseDir} as CAPSULE_CINEMA_ROOT. This is a standalone Codex, Claude Code, or OpenClaw installation.
2. Otherwise, if {baseDir}/../../scripts/capsule.py exists, normalize {baseDir}/../.. and use it as CAPSULE_CINEMA_ROOT. This is an OpenClaw plugin or source checkout.
3. Otherwise run bash {baseDir}/scripts/bootstrap-runtime.sh. Use the absolute path printed on stdout as CAPSULE_CINEMA_ROOT. The bootstrap clones the public runtime into the installed skill without reading or moving credentials.
4. Change to CAPSULE_CINEMA_ROOT before using relative paths in this skill.

If the runtime was just cloned, verify the prerequisites and install Python dependencies before production:

python3.12 --version
python3.12 -m pip install -r lib/requirements.txt
PYTHONPATH=lib python3.12 scripts/capsule.py list

FFmpeg is optional for recipe browsing and storyboard-only work, but required before rendering, audio mixing, subtitle burn-in, or final assembly.

Route the request

Read references/production-guide.md before planning or generating media. Use it to route to the narrower references for storyboard craft, production patterns, channel policy, tool recipes, capsule packages, assembly pitfalls, review gates, and workflow artifacts.

Classify the request before choosing tools:

• New AI video or storyboard
• Post-production or assembly from local media
• Reference-video analysis or remake
• Existing capsule production
• Targeted scene, voice, subtitle, BGM, or assembly repair
• Action transfer, digital human, lip sync, or music-video workflow
• Capsule creation, update, validation, packaging, or installation
• Release QA or blocker diagnosis

Use only workflows and providers registered in the runtime. Do not invent an unsupported route or silently replace the promised production method.

Follow the production contract

• Load capsules/<name>.capsule/ before planning whenever a capsule is selected. Treat archived or legacy single-file recipes as non-active sources.
• Read lib/config/tool_capabilities.yaml and lib/config/tool_registry.yaml before choosing or invoking providers.
• Confirm the final image, video or motion, TTS or voice, BGM or music, SFX, subtitle, compositing, and local-script tool chain before paid or batch generation. Explain relevant alternatives, missing capabilities, and any downgrade.
• Define the delivery promise before generation: motion-led, source-led, TTS-led explainer, reference remake, capsule preset, or specialized route.
• Generate and inspect one representative hard scene before batching a new AI-video direction.
• Never silently downgrade after a provider or tool failure. Retry within the approved policy, or stop for approval when a substitution changes the promised result.
• Read contracts/content_scope.yaml before creating, updating, or materializing a capsule. Keep episode-specific facts, people, projects, accounts, metrics, prices, narration, and temporary assets out of reusable capsule defaults. Preserve only repeatable structure, identity, method, and QA knowledge.
• Review semantic conflicts before updating an existing capsule. Run scripts/capsule_package_validate.py after each create or update: it now includes the consistency lint, and drifted capsules fail validation, so the update command itself stops until the drift is fixed.
• Keep a capsule in preset mode unless a mature deterministic local renderer has successful-run evidence, cross-topic verification, parameterized inputs, and explicit deterministic steps.
• Treat lifecycle blocked as blocked and review_required as pending. Do not report either state as complete.
• Keep final deliverables under output/. A release must include its artifact manifest, applicable QA reports, repair plan when needed, and release/release_checkpoint.json.

Review every writing pass

For every new or revised narration, script, title, cover line, or publishing copy, read references/natural-copy-review.md. Review theme, viewer benefit, and evidence before language. Apply the same pass after duration edits and before TTS; keep revision notes out of audience copy. This also applies to manual production outside the runtime wrappers. Preserve user-supplied verbatim text when requested.

Protect credentials and local state

• Read credentials only from agent-managed secrets, system environment variables, or a user-created root .env for standalone local execution.
• Never ask the user to paste complete secret values into chat. Never print, commit, migrate, or place secrets in prompts, capsules, manifests, logs, or generated copy.
• Never create lib/.env. Keep lib/.env.example free of secrets.
• Preserve user-owned .env, local-channels/, capsules, generated output, and private provider adapters during upgrades.

Use the command entry points

Run commands from CAPSULE_CINEMA_ROOT with PYTHONPATH=lib where shown.

Inspect each command's --help before constructing an unfamiliar invocation. Prefer the high-level runners over calling provider classes directly.

Work in reviewable stages

For serious generation, use this sequence unless the user explicitly requests a narrower stage:

1. Summarize the viewer experience, route, tool chain, expected limitations, representative-scene gate, and release QA bar.
2. Produce and validate the storyboard.
3. Wait for direction approval before paid or batch media generation.
4. Produce one representative hard scene and inspect it.
5. Generate the remaining media after the sample passes.
6. Build and validate work/edit_plan.json.
7. Assemble narration, video, subtitles, BGM, and SFX.
8. Run technical QA, visual checks, quality scoring, and visible-copy checks required by the route.
9. Repair blockers or report them honestly.
10. Create the release checkpoint and return only verified deliverables.

For targeted feedback, regenerate only the affected scene or layer and then rebuild dependent artifacts. Do not restart the entire production unless the change invalidates the whole plan.

Manage capsules safely

Active capsules live at capsules/<name>.capsule/. Read references/capsule-package-format.md before creating or changing their structure.

Package and install a capsule with:

python3.12 scripts/capsule_package_pack.py capsules/<name>.capsule --out /path/to/packages
python3.12 scripts/capsule_package_install.py /path/to/packages/<name>.video-capsule.zip --out capsules

Before writing reusable lessons, distinguish stable production knowledge from one episode's literal content. Promotion is a move, not a copy: once a lesson lands in a recipe or QA rule, delete it from learning/promoted_lessons.yaml in the same update. Show conflict points to the user before overwriting an established capsule rule.

Diagnose honestly

When a workflow cannot continue, identify the exact layer:

• Missing prerequisite or dependency
• Missing provider credential or entitlement
• Unsupported capability or unregistered provider
• Invalid storyboard, capsule, media path, or edit plan
• Failed representative scene or quality gate
• User approval still required
• Release checkpoint blocked

Return the smallest safe next action. Do not describe partial media, an unvalidated timeline, or a blocked lifecycle as a finished video.