v2.7.0 - Ultimate Seedance 2.5 skill

Added

• A second shipped skill: seedance. ByteDance's Seedance video models run on Dreamina, Jimeng AI,
  Doubao and the BytePlus API as well as inside ComfyUI, so their prompting knowledge could not live
  inside a ComfyUI skill without being wrong half the time. It ships as its own skill in three layers:
  SKILL.md routes, reference.md is normative and carries only ByteDance primary sources plus
  parameters read from ComfyUI's node code, and supplementary.md holds third-party observations in a
  subordinate layer with an explicit conflict ledger. Bundled for the Claude Code plugin by
  tools/build_plugin.py.
• The official Seedance 2.5 mechanics, from the Dreamina Seedance 2.5 User Guide (ByteDance, last
  modified 2026-07-31). The 2.5 prompt formula is [creatives description] + [one-sentence summary] + [specific plot description] + [overall supplement]. The three-module long-video formula includes an
  explicit anti-collapse block whose "prohibited" slot is where sound, subtitle and known-fragile items
  go. The real-person character formula runs seven dimensions with a sub-formula each, including the
  official anti-plastic trick: the literal suffix "retaining the true micro pores and skin texture".
• Timestamp direction, which reverses the 2.0 advice. Seedance 2.0's official guide calls precise
  timing unstable and tells you not to force it. 2.5 shipped timestamp control as a headline feature,
  built for exactly that request, with the documented syntax 0s-3s: / 3s-8s:. Both versions are
  now stated separately so an agent does not apply 2.0's warning to 2.5.
• Exact 2.5 limits and the official stability guidance that contradicts "more references is better":
  30 images, 10 videos and 10 audio clips (single 2 to 30s, 30s total each), audio-only input now allowed
  where 2.0 required at least one image or video, duration [4, 30] which is 97 to 721 frames, and output
  at 480p or 720p (the third-party "native 4K" claim is refuted; 4K is an input resolution only).
  ByteDance's own numbers: 1-5 subjects work well, 6-10 lose stability, and above 5 subjects you should
  split viewpoints into separate images rather than supply one multi-view image.
• The modes, with the ceiling that is easy to get wrong. Long Video generates 30 to 180 seconds in
  one shot. Extend Video accepts only sources under 30s, adds 4 to 30s per operation and nests, so the
  extreme case is a 30s original extended by 30s for 60s, and that is the ceiling on that path.
• The Blender route, which turned out to be official. Clay Renderer is a Dreamina plugin for Maya and
  Blender: set the camera route in your DCC, export the white-model video, upload through the plugin, and
  Seedance references the white-model action and camera movement.

Fixed

• Two plugin manifests had drifted and nobody was watching them. .claude-plugin/marketplace.json
  and claude-code/.claude-plugin/plugin.json both advertised 67 model prompt recipes against an
  actual 75, and plugin.json still declared version 1.9.0 while the repo shipped 2.6.0. Same class
  as the GitHub About box: a surface that no file edit touches and no count sync reached. Both corrected,
  and the manifest counts are now part of the release checklist.

Added

• Seedance 1.5 Pro, a whole tier the recipe had been silent about. docs/MODEL_INDEX.md already listed
  it; MODELS.md did not mention it once, so an agent reading the recipe would never reach it. The weekly
  cycle could not catch this because check_updates.py reports only the delta since the last run and 1.5
  predates the window. Now documented with the parameters read from comfy_api_nodes/nodes_bytedance.py,
  including three facts no prompt guide carries: the node enforces a 4 second minimum even though the
  slider allows 3, generate_audio is honoured only for 1.5 Pro and doubles the price, and
  raise_if_text_params rejects settings written into the prompt text.
• Seedance 2.5 status, recorded so nobody builds against it. It launched 2026-07-31 and has no
  ComfyUI nodes; the recipe now says so explicitly and points at 2.0 as the newest available in ComfyUI.
  The 2.5 deltas that are confirmed (30 images + 10 video + 10 audio per pass, 30s per generation with
  multi-round extension, against 9+3+3 and 15s for 2.0) come from the official ByteDance Seed blog.
• Routing to the new seedance skill for the family's full prompting mechanics: the three task types
  and the one word that switches between them, the @Image 1 label syntax, the （）<>{}【】 symbols,
  shot sequencing, the counterintuitive official rule to use 4 to 5 assets rather than the 50-asset
  ceiling, and the failure table for face drift, duplicate characters, unwanted subtitles and extension
  seams. Distilled from the official BytePlus ModelArk prompt guide.