Creative Claw

The unified AI media studio. Generates images, videos, speech, and branded graphics through one MCP server with 40 tools and one shared asset library. This skill is the single source of truth for every workflow — the MCP server provides the tools, this skill provides the knowledge.

Prerequisite

The Creative Claw MCP server must be connected. If generate_image, render_html_image, get_theme etc. aren't available, install:

/plugin install creative-claw@creative-claw-marketplace

Or connect the MCP directly: https://app.creativeclaw.co/mcp

Principles

1. Theme-first. If the user has a brand, every generation pulls from it. Call get_theme before any branded work. No theme + branded request → run the brand-theme workflow before generating.
2. The right engine for the job. Layout-driven content (text, logos, repeatable cards) → HTML render or template. Photoreal content (people, scenes, products) → AI model with a reference image. Code-driven motion (kinetic type, branded animation) → HyperFrames. Existing footage (cut, grade, subtitle) → edit-video workflow.
3. Reference images keep AI on-brand. Never run generate_image or generate_video for branded work without an image_url anchor — without it, models drift off-brand regardless of prompt quality.
4. Async means async. generate_video and render_html_video return job IDs. Always poll check_job. Never claim a result without status === "completed".
5. Assets are forever. Every result lives at a permanent R2 URL. Tag and name everything (update_asset) so future-you can find it (search_assets).

Hard rules (non-negotiable)

1. Always `get_theme` first for any branded generation. Skip only if the user has explicitly said "ignore my brand."
2. Always generate a reference image before any AI video. Even text-to-video. The starting frame controls quality and on-brandness.
3. `render_html_image`: never raw external URLs in `<img src>` or `background-image: url()`. Always use inline_images — substituted as data URIs through the SSRF-guarded cache. CORS, redirects, and expired CDN URLs are the #1 cause of broken renders.
4. Async tools require `check_job`. generate_video, render_html_video, generate_3d_model, transcribe all return { jobId, status: "queued" }. Poll until completion. See references/async-jobs.md.
5. HyperFrames knowledge lives in the `/hyperframes` skill. For any code-driven video composition, install and invoke /hyperframes (and /hyperframes-cli for local commands). Creative Claw provides the cloud renderer (render_html_video) and reusable templates (create_template / render_template) — the composition rules are owned by the HyperFrames skill.
6. Editing existing footage lives in the `/video-use` skill. Transcribe / cut on word boundaries / color grade / burn subtitles / 9:16 reframe — install and invoke /video-use. Use individual MCP tools (transcribe, trim_video, add_subtitles, etc.) only for one-shot ops.
7. Tag and name assets at generation time. Pass name and tags to every generate_* / render_* call. Use update_asset after if you forgot.
8. Don't quote pricing from memory. Use get_credits_balance for the user's balance. Hand them get_credits_link for top-up — you cannot complete checkout for them.
9. Never use the elicitation form / visualize widget when a question requires an image upload. Ask in plain chat instead so the user can attach the image directly to the conversation. The elicitation form is fine for everything else (text inputs, choices, parameter pickers) — this rule applies only when image upload is involved.
10. Pass thorough prompts verbatim — set `agentic_prompting: false`. generate_image / generate_video rewrite prompts server-side by default. When your prompt already contains literal control tokens (@Image1 / @Audio1 mention tags, exact timecodes, quoted dialogue, per-panel layout), disable the rewriter — paraphrasing @Image1 into "the reference image" silently breaks the model's syntax. Keep it on only for short, loose, single-line creative asks. Send it verbatim when you've done the directing; let the server direct when you haven't.

Routing

When the user's request matches a row, read that file before doing anything. The workflow files contain the model selection, prompting strategies, parameter sets, and anti-patterns that aren't reproduced here.

References index

Workflows (references/workflows/)

• onboard.md — first-time tour. The mission, four playful first-generations, the brand-theme unlock, recipes per goal.
• image-gen.md — AI image generation. Model selection picker (Nano Banana, FLUX, GPT Image, Recraft), editing vs. generating, branded vs. unbranded. The MCP server tailors prompts per model server-side.
• video-gen.md — AI video generation. Model selection picker (Veo, Sora, Kling, Hailuo, Seedance, HeyGen), reference-image-first rule, async polling, multi-segment planning, talking-avatar guidance, speech models list.
• storyboard-to-video.md — storyboard-first pipeline. Plan beats → generate a multi-panel review board (nano-banana-pro) → iterate by editing → extract CLEAN keyframes → hand off to Seedance 2.0 (@Image/@Audio refs, first/last frame, optional voice reference). Includes the "review board vs. generation frames" rule and when to disable the prompt rewriter.
• html-image.md — render_html_image deep dive. Full CSS surface, inline_images, fonts, dimensions, theme integration, when to use vs. AI gen.
• template-creation.md — Create a new template (both flavors: HTML render-templates AND AI-image prompt templates). Reference-image collection (existing files, URLs, or import_media drop-in), analysis, minimal dynamic params, render → load → iterate loop, prompt recipes for AI-generated background images inside HTML templates, "render one now" handoff.
• banner-from-template.md — Deeper HTML template mechanics: full HTML example, create_template parameter shape, batch rendering, the four templates worth building first.
• code-video-hyperframes.md — how Creative Claw's render_html_video and templates fit into a HyperFrames pipeline. Composition rules themselves live in the /hyperframes skill — install it before authoring.
• brand-theme.md — full theme lifecycle. Website extraction, local folders, URL lists, generating from scratch, updating, applying in generation.
• edit-video.md — pointer to the /video-use skill plus a list of the MCP tools available for one-shot edits (transcribe, trim_video, add_subtitles, etc.).

Model selection

Per-modality picker tables live inside the workflow files (image-gen.md, video-gen.md). The MCP server enhances user prompts server-side using per-model knowledge — write clear creative direction and the server tailors it for the chosen model. Use list_models and get_model_params at runtime to discover model IDs and extras.

Cross-cutting (references/)

• tool-catalog.md — MCP tools grouped (generation, templates, themes, assets, editing, models, jobs, credits).
• async-jobs.md — check_job polling pattern, parallel jobs, failure modes, timeouts.
• platform-dimensions.md — IG, LI, X, YT, TikTok, OG sizes + safe zones. Machine-readable copy at assets/platform-dimensions.json.

Assets (assets/)

• platform-dimensions.json — JSON copy of dimensions table.

Tool catalog quick reference

Full details and grouping in references/tool-catalog.md. Common ones:

Generate — generate_image, generate_video, generate_speech, render_html_image, render_html_video, render_template, compare_models Edit — remove_background, upscale_media, trim_video, scale_video, add_subtitles, extract_frames, merge_media, transcribe Brand — get_theme, list_themes, update_theme Templates — create_template, render_template, list_templates, update_template Assets — search_assets, update_asset, upload_asset, import_media, load_image Jobs — check_job (poll all async generations) Credits — get_credits_balance, get_credits_link