AGENT J / WORKFLOWS
Story Animation
Turn a story worth telling into a film worth finishing.
Start with this request
Install Story Animation from the Plaza and guide me through setup…
Help me install and configure Story Animation from the Agent J Plaza: ID story-animation, pinned version 1.2.3. First read https://agentj.app/workflows/story-animation/?lang=en plus the Plaza description and README, and check my download access. If unavailable, explain the limit and offer https://github.com/preangelleo/story-animation as a public-source option, without purchasing anything. Preview agentj plaza install story-animation --version 1.2.3 --workspace <a new directory I confirm> --harness codex; follow actual local help and native confirmation. Explain the uncertified community status; do not skip signature checks, overwrite a workspace or purchase a seat for me. Read AGENTS.md, SKILL.md, documentation/WORKFLOW.md, documentation/PROVIDERS.md, documentation/FONTS.md and documentation/LEARNING.md, run doctor and modules status, and report actual capability gaps. Help select my author (Codex, Claude, Ark or a custom adapter, writing from frozen sources without network or tools), voice, image tool and licensed fonts. If a recommended provider is unavailable, explain and verify the required adaptation before substituting it. Configure the subtitle font explicitly with me: STORY_SUBTITLE_FONT points to a font file I am licensed to use, STORY_SUBTITLE_FAMILY is its Fontconfig family, STORY_SUBTITLE_FONT_INDEX is set for collections, and the font is registered with Fontconfig in the runtime environment and actually prechecked; no font is bundled, so never assume a default font for me. SQL memory, embeddings, comment intake, comment replies, topic suggestions and dedup are all off by default: preview first and let me decide each one. Comment replies stay read-only until I explicitly enable them; without a real embedding endpoint, do not claim semantic search works. Collect keys through Agent J secret cards or my secret manager, never chat, logs or the repository. Confirm the topic, 1–3 takeaways proposed after research, visual style, spending budget and publishing needs. Validate a small sample, including the first-frame cover and chapter-card frames free of body captions, within my approved budget before producing the full story. YouTube authorization, paid calls, uploading and scheduling each need my explicit approval. Normal upload, readback, recovery and final visibility checks never download caption text: they verify durable successful caption IDs and live captions.list serving status, reconcile unknown outcomes without blind reposting, and use --diagnostic-caption-download only when I explicitly ask for a diagnostic. Report completed setup and remaining gaps; doctor passing is not proof of a fully reviewed film.
Spend your time on the story
An independent animated storytelling workflow: research, narrative, consistent characters, voice, aligned captions, chapter transitions and a finished film. Configure it, adapt its providers, or change the source.
The guide, pilot and GitHub source are public; you may also obtain the source from the public repository. Full native Plaza downloads require an Agent J account with a paid seat. Third-party model and media usage is billed separately. Installation does not authorize spending or automatic publication.
Featured · Community workflow · Not officially certified
Watch the complete pilot
The approximately 18-minute historical pilot, produced with 1.1.0 and not re-rendered with any 1.2.x release (1.2.0–1.2.3): vintage rubber-hose visuals, Chinese narration, restrained sound design and chapter transitions. This direct version has burned-in Simplified Chinese captions and can be downloaded for local viewing. Download the pilot · Five selectable subtitle tracks on YouTube
No autoplay · No third-party player · Full film
From a subject to a finished story
Research and intent: collect sources and propose 1–3 takeaways for the audience.
Write and review: the orchestrator does the research; your configured author (Codex, Claude, Ark or a custom adapter) receives only frozen sources, 1–3 takeaways and compact dedup advice, writes a fresh spoken narrative without tools, and an independent editor reviews it.
Characters and scenes: character reference sheets guide scene images. Still images stay still; key moments can use optional video inserts.
Voice and captions: narration, script-based forced alignment, multilingual subtitles and a reusable local sound library.
Chapters and delivery: chapter clips join at scene boundaries, and every decoded output frame is checked against the original clips. Body captions are masked on the exact chapter-card frames without rewriting the source SRT, forced alignment or the five language tracks. Review the final film before choosing whether to publish.
The tested production stack
- Orchestration and thumbnail
- Codex; thumbnail role configured as gpt-6.1-sol. Actual built-in image_gen availability must be checked.
- Chinese narrative (1.1.0 pilot)
- The pilot used doubao-seed-2.1-pro on Ark Agent Plan, invoked through headless OpenCode with research separated from writing. 1.2.x does not call Ark by default: the installer configures the author, and Ark is one adapter you can explicitly choose.
- Narration and alignment
- The pilot used an owner-authorized ElevenLabs cloned voice with eleven_v4; the public package defaults to configurable eleven_multilingual_v2. Captions use ElevenLabs Forced Alignment.
- Images and optional video
- Codex built-in image_gen; the exact underlying image model is not exposed. Video inserts used Seedance 2.5 through Replicate.
- Local rendering
- ffmpeg / ffprobe, Fontconfig and HyperFrames 0.8.225. This tested combination does not imply every provider replacement has been verified.
- 1.2.3 package and verification boundaries
- This page describes the v1.2.3 distribution manifest and bundled documentation. Learning modules are all off by default; the package includes configurable authors, first-frame covers, chapter-card caption masks, explicit font setup and caption-ID/serving checks. The repository includes synthetic-media and mock tests; this page does not claim actual YouTube comment intake or replies, caption upload or download, real embedding quality or accepted real films. The historical showcase was not remade with this package.
What capabilities do you need?
| Stage | Recommended capability | Replacement and limitations | Current status |
|---|---|---|---|
| Research and script | The orchestrator researches with available web search; the author uses your configured Codex, Claude, Ark or custom adapter | author.provider accepts codex, claude, ark or custom. Without author.command it only produces a frozen handoff packet for an independent tool-free session. Kimi, Qwen or other endpoints need an adapter; verify authentication and plan eligibility, as a subscription does not imply arbitrary API access. | No required vendor; Ark was tested in the 1.1.0 pilot and is not the 1.2.x default |
| Narration and authorized voice | ElevenLabs: your API key and an authorized voice | Volcengine or Fish Audio can be alternatives. Their adapters are not bundled; Agent J must adapt the API, produce a listening sample and check costs. | ElevenLabs adapter included |
| Script-based forced alignment | ElevenLabs Forced Alignment: script + audio → timestamps | Volcengine offers transcript-assisted timing. It is not a drop-in adapter: map word/sentence timing and rerun text, segmentation and clock checks. | ElevenLabs tested; Volcengine integration required |
| Images and character consistency | Codex with working built-in image_gen | Agent J can integrate a third-party image service while preserving character references, output manifests and image review. No universal third-party image adapter is bundled. | Recommended path tested; replacements need code adaptation |
| Optional video, sound and music | Replicate video; ElevenLabs sound effects / music | Video inserts can be disabled. Sounds first use a lowercase-name local cache, then request generation only when missing. Licensed local assets or other adapted providers can be used. | Optional; confirm spending and rights before generation |
| Fonts, rendering and publication | Python 3.11+, Node, ffmpeg, Fontconfig, HyperFrames; licensed fonts, with the subtitle font explicitly configured and registered through Fontconfig | Local film delivery is sufficient. YouTube is optional: your Google Cloud OAuth app, a separate channel and explicit profile; caption upload requires youtube.force-ssl permission. In 1.2.3, normal publishing never downloads caption text; it checks successful caption IDs and live captions.list serving status. | Fonts, keys, voices and channel tokens are not included |
| Optional semantic search (embeddings) | Your own OpenAI-compatible embedding endpoint and model; the key is referenced only by the STORY_EMBEDDING_API_KEY variable name | No vendor is required, and you can skip it entirely. Without it, dedup uses SQL/lexical matching only and makes no semantic-recall claim. Failed vector writes stay observable and retryable in the SQL outbox; SQL is authoritative and vectors are rebuildable, not a cross-database atomic transaction. 1.2.3 sends each record as one unsplit input; many embedding models cap single-input length and may silently truncate the rest, so semantic coverage of long full texts is unverified and must not be treated as working full-text semantic search. Actual cost comes from real bills. | Off by default; not yet verified with a real provider |
| Optional comment feedback and topic suggestions | Your own verified channel and OAuth; comments are read as untrusted data | Comment intake, replies, topic suggestions and the dedup adviser each have their own switch. Replies are preview-only until the installer explicitly enables them; sensitive, contentious, unclear or promotional comments wait for human review. Quota or API errors never count as zero comments. | All off by default; no comments actually read or answered |
What can you configure directly?
These are existing surfaces: installation-template parameters, JSON / presets, CLI flags and environment settings are labeled separately. They are not all installation flags. Subtitle size, text color, background box and margins currently require source changes; they are not parameterized.
Expand all parameters and configuration surfaces
| Configuration surface | Editable fields | Purpose and boundaries |
|---|---|---|
| Install template | project_name | The sole project-template parameter. Choose your working folder, AI tool and pinned version through the native Plaza CLI. |
| documentation/configuration.json · brand | brand.name, brand.english_name, brand.ceo, brand.promise | Brand name, English name, CEO label and channel promise. |
| documentation/configuration.json · channel | channel.id, channel.youtube_profile, channel.slug, channel.timezone, channel.publish_hour, channel.publish_minute, channel.lead_days, channel.language, channel.production_hour | Separate channel identity, profile, timezone, production and release time. Default: two days ahead for 20:00 Tokyo. Configuration does not enable automation. |
| documentation/configuration.json · content | content.language, content.narration_first, content.standalone, content.target_minutes.min, content.target_minutes.max, content.video_inserts.max_count, content.video_inserts.target_seconds_each, content.video_inserts.placement, content.video_inserts.default_model, content.video_inserts.model_version | Language, narration-first policy, target length, video insert count, duration, model and version. Inserts can be disabled; verify model access and price. |
| documentation/configuration.json · budget | budget.per_episode_usd, budget.monthly_usd, budget.max_retries | Episode and monthly budgets and retry limits. Configured amounts are not spending approval; never blindly retry an unknown outcome. |
| documentation/configuration.json · storage | storage.database, storage.artifacts, storage.vector_collection, storage.namespace | The work ledger is always local SQLite (versions, budgets, publication safety) and is separate from optional long-term fact memory. Also sets the artifact directory, vector collection name and namespace; a collection name does not mean semantic search is enabled. |
| documentation/configuration.json · voice | voice.provider, voice.voice, voice.model, voice.language | Voice provider, voice label, model and language. Changing a provider label does not implement an adapter. |
| documentation/configuration.json · render | render.width, render.height, render.static_image_motion, render.first_frame.enabled, render.first_frame.frame_count, render.first_frame.fit | Final canvas size and still-image policy; images stay still by default. The first-frame cover is on by default: it replaces frame 0 only (frame_count supports only 1; fit=strict never crops or stretches), adds no duration, and captions resume from frame 1. |
| documentation/configuration.json · quality | quality.require_visual_manifest, quality.require_semantic_caption_units, quality.min_script_chars, quality.min_duration_seconds | Minimum content length and acceptance requirements. Preserve visual manifests and semantic-caption checks. |
| documentation/configuration.json · publication | publication.default_dry_run, publication.allow_profile_fallback | Dry-run and no cross-channel profile fallback; preserve identity checks and publishing approval. |
| documentation/configuration.json · learning (SQL memory) | learning.enabled | Module name sql; off by default. When off, the work ledger still runs; there is just no cross-episode memory of accepted scripts, topics and summaries. Every other learning module requires sql first. |
| documentation/configuration.json · learning.vector | learning.vector.enabled, learning.vector.endpoint, learning.vector.model, learning.vector.dimensions, learning.vector.batch_size, learning.vector.api_key_env, learning.vector.collection | Off by default; endpoint, model and dimensions default to null, batch_size 16, collection story-animation. api_key_env holds only a variable name (default STORY_EMBEDDING_API_KEY); keys never go in config. Dimensions are checked against the actual response. Configuration is not proof of working semantic search; failed writes wait in the SQL outbox for retry. |
| documentation/configuration.json · learning.comments | learning.comments.enabled, learning.comments.auto_reply, learning.comments.lookback_hours, learning.comments.catchup_days, learning.comments.max_pages, learning.comments.not_before | Comment intake and replies are off by default. auto_reply (module reply) defaults to false: preview only, the installer must explicitly enable it, and it requires comments. Defaults: 24-hour lookback, 7-day catch-up, 5 pages max. A gap longer than the catch-up window is blocked with no API call and no watermark advance. not_before is unset by default. |
| documentation/configuration.json · learning.topics / dedup | learning.topics.enabled, learning.dedup.enabled | Off by default. When off, topics follow a manual frozen plan and dedup is a manual editorial check. When on, feedback only supplements future planning, and the dedup adviser reads SQL first, then any available semantic candidates. |
| documentation/configuration.json · author | author.provider, author.network_tools, author.command, author.profile, author.timeout_seconds | provider defaults to the configured AI tool; codex, claude, ark or custom are accepted, with no required vendor. network_tools must be false. Without command, only a frozen handoff packet is produced; command is an installer-supplied argv adapter (no shell, no inherited secrets, tools_used must be empty). Output is always a candidate that needs independent editing. |
| Learning modules · CLI | modules status, modules configure --enable, sql,vector,comments,reply,topics,dedup, --embedding-endpoint, --embedding-model, --embedding-dimensions, --author-provider, --execute | --enable is the complete desired set, not an addition; empty disables everything. Preview by default; add --execute to save. Saving makes no provider calls, replies, uploads or timers. |
| Learning · CLI | learning drain EP, learning feedback-packet, learning comments, --lookback-hours, --catchup-days, --max-pages | drain runs permitted pending index and comment work and reports each status; unknown cost stays null, and uploads are never retried. Scan flags override one scan only (1–8760 hours, 1–365 days, 1–1000 pages). feedback-packet needs independent feedback analysis before assessing or sending exact approved replies. |
| First-frame cover · CLI | thumbnail context EP, thumbnail prepare EP, --image, --review, thumbnail verify EP, thumbnail consume EP | The cover gate before the final render. The cover must match the final canvas exactly; review.json binds image_sha256, an independent reviewer (never the episode author), references and the current context. prepare previews by default; add --execute to save. verify and consume re-check read-only. |
| JSON file routing | style_registry_file, series_mapping_file | Choose style and series mapping file paths. schema_version, stage and revision_approval_gates are schema/governance fields, not creative switches. |
| config/styles.json | index, id, name.zh, name.en, description, prompt, palette, lighting, source | Edit or extend 26 styles. source is provenance metadata, not a composition instruction. styles resolve supports --style-index, --style-file and --mapping-file. |
| config/weekly_series.json | timezone, key, display_name, weekday, default_style_index, description | Weekday series and default styles. An episode --style-index takes precedence over weekday defaults. |
| Chapters · colors · preset / CLI | background_color, title_color, number_color, line_color | Defaults: #141414 background, #faf9f5 title and number, #d97757 accent line. Use chapters render --preset PATH --output PATH; add --execute to render. Each field has a --kebab-case override. |
| Chapters · text and layout · preset / CLI | title_text, chapter_number, layout, group_x, group_y, group_width, text_align, group_padding_left | Title, chapter number, a/b layout, position, width and alignment. Use chapters render --preset PATH --output PATH; add --execute to render. Each field has a --kebab-case override. |
| Chapters · typography · preset / CLI | title_font, title_family, number_font, number_family, font_weight, title_size, number_size, title_line_height, number_line_height, title_letter_spacing, number_letter_spacing, number_gap | Your licensed font files and actual families, size, leading, letter spacing and gap. font_weight currently only accepts 400. Use chapters render --preset PATH --output PATH; add --execute to render. Each field has a --kebab-case override. |
| Chapters · line · preset / CLI | line_enabled, line_width, line_height, line_gap, line_x, line_y | Enable, dimensions, gap and offsets. Use chapters render --preset PATH --output PATH; add --execute to render. Each field has a --kebab-case override. |
| Chapters · timing and entrance · preset / CLI | duration, width, height, fps, entrance_y, entrance_duration | Duration, size, frame rate and text entrance of the standalone transition; this does not add motion to still images. Use chapters render --preset PATH --output PATH; add --execute to render. Each field has a --kebab-case override. |
| Chapters · sound · preset / CLI | sound_path, sound_name, sound_volume, muted | Local sound path, cache name, gain and mute. The public preset is muted and does not include the creator's sound files; generate with your own rights and budget. Use chapters render --preset PATH --output PATH; add --execute to render. Each field has a --kebab-case override. |
| Subtitle font · environment | STORY_SUBTITLE_FONT, STORY_SUBTITLE_FAMILY, STORY_SUBTITLE_FONT_INDEX | Public installs must set these explicitly: STORY_SUBTITLE_FONT points to a font file you are licensed to use, STORY_SUBTITLE_FAMILY is its Fontconfig PostScript family, and STORY_SUBTITLE_FONT_INDEX selects the face in a collection (default 0). Register the font with Fontconfig for the same user or container that runs the workflow (fc-cache, or FONTCONFIG_FILE pointing to an explicit configuration). No font or default family is bundled; a missing family fails before any output is written, and a mismatched Fontconfig selection stops instead of falling back. Chapter body-caption burns switch to tools/burn_subtitles_masked.py, which receives --font / --font-family / --font-index from these variables and binds the font SHA into its fingerprint. Precheck with fc-query / fc-match and burn a short sample first. |
| Multilingual subtitles · CLI | multilingual prepare --languages, zh-Hans, zh-Hant, en, ja, ko | Select Simplified/Traditional Chinese, English, Japanese and Korean. Review translations and preserve the aligned timing map. The R2 pilot on this page has burned-in Simplified Chinese captions only. |
| Ark adapter · environment and CLI (tools/chinese_author.py) | STORY_AUTHOR_MODEL, STORY_AUTHOR_ENDPOINT, --min-han-chars, --max-han-chars, --timeout | Applies only when you explicitly use the Ark adapter; the default author route never calls Ark. STORY_AUTHOR_MODEL defaults to doubao-seed-2.1-pro. Default minimum: 3500 Han characters; max 0 means unlimited. Preserve substance rather than compressing to a target duration. |
| ElevenLabs · CLI | speak, align, sfx, music, --model, -v, --seconds | Model, your authorized voice and duration. Inject the API key through a secret manager, never a public parameter file. |
| Local sound library · CLI | sound get NAME, --category, --prompt, --seconds, --episode, --scope | Lowercase names reuse cached assets. Missing generation requires --execute --confirm-cost --estimated-usd; preserve spending approval. |
| Sound cues · JSON | scene_id, name, source_audio_in, source_audio_out, source_timeline_in, final_timeline_in, gain_db, fade_in_seconds, fade_out_seconds, ducking.sidechain, ducking.attenuation_db, ducking.attack_seconds, ducking.release_seconds, audio_sha256 | Scene placement, gain, fades and narration ducking. audio_sha256 is an integrity constraint, not a value to fabricate; validate before mixing. |
| YouTube · explicit configuration | STORY_YOUTUBE_TOKEN_DIR, profile, channel_id, title, description, privacyStatus, playlist | Separate OAuth profile, channel, title/description, visibility and series playlist. Required scopes, channel readback and publishing approval still apply. Before caption upload, local five-track text, timing, hashes and independent review must pass; afterwards only a durable receipt with a genuinely successful nonempty caption ID counts, checked against live captions.list for a unique ID, video, language, name and serving status, with videos.list confirming the video and channel. |
| Pilot release and caption readback · CLI | pilot-release upload / thumbnail / captions / playlist / readback / privacy / recover, --plan, --execute, --privacy, --operation, --remote-id, --reason, --diagnostic-caption-download | Write stages default to dry-run. readback/recover make real read-only API checks, and successful recover also writes the local reconciliation ledger. Platform writes require --execute; visibility additionally requires an explicit --privacy. Captions in processing keep waiting and block final visibility; failed is rejected; 403, errors, unknown or corrupt receipts never count as success and are never blindly resent; reconcile them with recover. Normal results record upload_receipt_verified / metadata_serving_verified with content_matches=null, which is not a remote full-text comparison. Only a person explicitly adding --diagnostic-caption-download to readback or recover downloads remote captions for a cue-by-cue comparison; write stages reject the flag before any API call, and no config or background job can enable it. |
Parameters cover common changes. The workflow source can also be edited: ask your Agent J to adapt any stage in your workspace and add verification. Preserve provider permissions, spending approval and human review.
Seven series. A visual style for each.
Edit the weekly-series JSON and the registry of 26 visual styles, or choose a style index for an individual episode independently of the weekday.
- Monday · Fate / inspirationEuropean Clear-line Comic
- Tuesday · Mystery / suspenseVintage Rubber-hose Cartoon
- Wednesday · Wealth / entrepreneurshipShonen Action Anime
- Thursday · History / biographyInk-wash Animation
- Friday · Love / relationshipsHand-painted Pastoral
- Saturday · WomenRounded 3D Feature
- Sunday · OtherHandmade Clay Stop-motion
Before you begin
Can I use different providers?
Yes, by adapting the workflow. The recommended stack is the tested starting point. Ask Agent J to adapt the relevant provider, then validate inputs, timing, sound, reference-image support and costs with a sample. A possible replacement is not a verified integration.
What if parameters do not cover my needs?
You receive readable, editable workflow code. The table documents existing configuration surfaces; ask your Agent J to change source and tests for anything else. Preserve key isolation, budget controls, identity checks and human review.
Will it spend money or upload every day?
Installation does not start paid production or publishing. Producing two days ahead for 20:00 Tokyo time is an editable default. You confirm the production budget, channel authorization and automation activation.
How are quotations and dramatization handled?
Research separates checkable facts, attributed accounts and dramatized scenes. Preserve sources for quotations from real people. The film is storytelling; names, amounts, pronunciation and material claims still need human review.
Will it reply to comments automatically?
No. Comment intake and replies are both off by default. Even with intake on, replies stay preview-only until the installer explicitly enables reply. External comments are untrusted data; sensitive or unclear ones wait for human review, and an unknown reply outcome is reconciled against receipts before any retry.
Do I need a particular embedding vendor?
No. Embeddings are off by default and optional; without them, dedup uses SQL and lexical matching only. To enable them, configure your own compatible endpoint, model and key variable name. Configuration is not verification, and this page claims no semantic results from any real provider. Note that 1.2.3 does not split long texts, so anything beyond a model's single-input limit may be truncated.
Does the public package include a subtitle font?
No. The public package bundles no font files and has no default font family. Use a font you are licensed to use, set STORY_SUBTITLE_FONT and STORY_SUBTITLE_FAMILY (plus STORY_SUBTITLE_FONT_INDEX for collections), and register it with Fontconfig in the runtime environment. A missing family fails before any output is written, and a mismatched font selection stops production.
Does publishing download caption text from YouTube to verify it?
No. Normal upload, readback, recovery and final visibility checks never download captions: local five-track text, timing, hashes and independent review pass first, and afterwards only durable successful caption IDs and live captions.list serving status are checked, with content_matches=null and no claim of remote full-text equality. Only when you explicitly ask for a diagnostic does readback or recover with --diagnostic-caption-download download captions for a cue-by-cue comparison; neither background jobs nor default configuration can enable it. The public package's diagnostic path was tested only against a fixture transport, not a real YouTube download.
Source and installation
The guide, pilot and GitHub source are public; you may also obtain the source from the public repository. Full native Plaza downloads require an Agent J account with a paid seat. Third-party model and media usage is billed separately. Installation does not authorize spending or automatic publication.