Content & Media Creation · prime-skills/runcomfy-agent-skills

wan-2-7

Name: wan-2-7
Author: prime-skills/runcomfy-agent-skills

Generate text-to-video with Wan 2.7 (Wan-AI's flagship motion model) on RunComfy. Documents Wan 2.7's strengths (multi-reference conditioning, audio-driven lip-sync via `audio_url`, smoother transitions, prompt expansion), the duration / resolution / aspect-ratio schema, and when to route to HappyHorse 1.0 / Seedance 2.0 / Kling / LTX 2 instead. Calls `runcomfy run wan-ai/wan-2-7/text-to-video` through the local RunComfy CLI. Triggers on "wan", "wan 2.7", "wan-2-7", "wan video", or any explicit ask to generate video with this model.

L4 High-Privilege · Graded auth & full auditV1 Commit-PinnedExecute CommandNetwork AccessWrite FileDual-pass Audit

Control levels describe skill capabilities and suggested control measures; they do not represent quality or safety certifications, nor do they dictate usability. Usage decisions rest entirely with users and administrators. Different agent hosts apply varying constraints; do not assume hosts provide automatic protection. Third-party skill copyrights belong to their original authors. Use at your own risk.

1How It Works & Mechanism

Implemented via 'zero-code + external CLI delegation': wan-2-7/ contains only a single SKILL.md without scripts; at runtime, the host agent executes `runcomfy run wan-ai/wan-2-7/text-to-video --input '<JSON>' --output-dir <path>` via Bash, packaging prompts and parameters into JSON and delegating to the out-of-repo CLI. The skill's value-add lies in the documented schema and prompting methodology rather than any local inference capability (video generation itself occurs in the RunComfy cloud).

wan-2-7/SKILL.md

The skill invokes `runcomfy run wan-ai/wan-2-7/text-to-video` with a JSON body matching the schema.

git ls-tree HEAD verifies that wan-2-7/ has only a single entry, SKILL.md; all 31 skills across the repository share this structure, and runcomfy-cli/SKILL.md serves as the standalone manual for this CLI suite.

The core differentiating capability is 'audio-track-driven lip-sync': the audio_url field accepts WAV/MP3 files (3–30s, ≤15MB) to drive lip-syncing; if omitted, background music is automatically generated. This serves as the key differentiator between Wan 2.7 and other video models in this family (others either self-generate speech or cannot accept external audio tracks).

wan-2-7/SKILL.md

| `audio_url` | string | no | — | WAV/MP3, 3–30s, ≤15MB. **Drives lip-sync.** Omit → background music auto-generated. |

The documentation adds an operational discipline: 'Match audio length to clip duration.', and treats out-of-spec assets as anti-patterns that trigger immediate rejection ('Audio outside the 3–30s / 15MB / WAV-MP3 spec → rejected.').

The input schema comprises 8 fields covering duration, aspect ratio, resolution, negative prompts, prompt expansion, and reproducibility: prompt (~5000 characters / ~1500 tokens), aspect_ratio (5 values, defaults to 16:9), resolution (720p/1080p, defaults to 1080p), duration (2–15 seconds, defaults to 5), negative_prompt (≤500 characters), enable_prompt_expansion (defaults to true), and seed.

wan-2-7/SKILL.md

| `enable_prompt_expansion` | bool | no | true | Auto-rewrites short prompts. Disable for literal control. |

The schema contains no image/video reference input fields, yet the same document claims support for multi-reference conditioning under 'Where it shines' and 'Limitations' (see the discrepancy item in verification.second_pass).

The prompting guide translates video generation into actionable rules: front-load camera angles and motion in plain English, specify one primary action, restrict negative_prompt to concrete issues, disable the default-on prompt expansion for brand-sensitive literal copy, and fix the seed to perform variant comparisons.

wan-2-7/SKILL.md

**Use `negative_prompt` for concrete issues.** Good: "no subtitles, no watermark, no flicker". Bad (vague): "no bad lighting".

This 'good vs. bad writing comparison' constitutes the document's primary value—formalizing the model's tolerance for negative prompts into executable criteria so agents avoid crafting ineffective prompts.

The routing table cross-routes between this skill and other video models in the family: use Seedance 2.0 Pro for self-generated speech, HappyHorse 1.0 for top-tier leaderboard quality, Kling for motion editing on existing footage, and LTX 2 for ultra-fast iteration; if the user explicitly specifies Wan, unconditionally route back here.

wan-2-7/SKILL.md

| Lip-sync video to an audio track you supply | **Wan 2.7** (`audio_url`) |

The opening section also notes: 'If the user said "Wan" / "Wan 2.7" / "wan-ai" / "alibaba video" explicitly, route here regardless.' Among the suggested targets, Seedance 2.0 and HappyHorse 1.0 have matching skill directories in the repo; 'Kling Video O1' and 'LTX 2' have no corresponding skills in the repo (see second_pass).

Limitations and failure modes are explicitly defined: maximum duration of 15s (longer clips require stitching multiple invocations), maximum resolution of 1080p, strict audio specification constraints, a limit of 5 reference assets, and the operational boundary that 'this model does not generate speech; it only accepts audio tracks.'

wan-2-7/SKILL.md

- **For in-pass voice generation (no separate audio track), use Seedance 2.0 Pro** — Wan accepts audio rather than generating it.

Pairing 'capability boundaries + alternative model recommendations' is a consistent design pattern throughout this Pro Pack documentation; an exit code table (0/64/65/69/75/77) defines retry semantics.

2Core Capabilities

01Text-to-video generation (prompt ≤ ~5000 characters / ~1500 tokens; degrades if too long)

02External audio track-driven lip-sync (audio_url: WAV/MP3, 3–30s, ≤15MB; automatically scores music if omitted)

03Duration, resolution, and aspect ratio control (2–15s, 720p/1080p, 5 aspect ratios; defaults to 5s/1080p/16:9)

04Negative prompt filtering for concrete defects (negative_prompt ≤500 characters; only specify specific flaws)

05Prompt expansion toggle (enabled by default to rewrite short prompts; can be disabled for literal brand-sensitive copy)

06Seed-based reproducibility for variant comparisons (identical prompt and seed for comparisons; change seed for diversity)

07Multilingual voiceover variant workflow (same prompt paired with different audio_url sources to generate multilingual versions)

08Exit code semantics and retry decision logic (exit code 75 is retryable; 77 requires re-login)

3External Dependencies

Kind	Dependency / Endpoint
cli	runcomfy（RunComfy 官方 CLI）
package	@runcomfy/cli（npm 全局包）
package	npx skills（skills.sh 安装器）
api	RunComfy Model API（Wan 2.7 text-to-video）
network	RunComfy 产物下载白名单
network	用户传入的音轨/参考 URL（由 RunComfy 服务器抓取，非本机）
network	RunComfy 站点/文档（文档引用，非运行时调用）
network	GitHub（技能源码链接）

4Recommended Governance & Evidence

Suggested Controls for Safe Operation

For Agent Runtime

Graded/per-use authorization, sandbox recommended

For Administrators

Approval workflow, rollback support, full audit

Evidence & Triggers (File Paths & Excerpts)

scripts/: Executes local script or command: (The skill itself contains no scripts) The host agent executes the external CLI via Bash: `runcomfy run <vendor>/<model>/<endp`
网络端点: Declares external network access: POST https://model-api.runcomfy.net/v1/models/wan-ai/wan-2-7/text-to-video (
本地工作目录: File read/write and artifact output: Generated artifacts are downloaded and written to `--output-dir` by the external CLI (documentation notes that single files > 2 GiB will abort to prevent disk exhaustion)

Audit Notes: Based on the unified rating criteria (the sibling image-edit skill in the same repo was rated Amber): At runtime, it drives a third-party npm package/CLI `@runcomfy/cli`, which accesses user RunComfy credentials (~/.config/runcomfy/token.json or RUNCOMFY_TOKEN) on each call—triggering the 'invokes third-party CLI/remote package and accesses credentials/quotas' tier. Claims regarding CLI implementation, Model API behavior, token file permissions (0600), the 2 GiB download cap, and exit code semantics all reside outside the repository, with no source code to verify within this pin; per the guideline 'when in doubt, bump up one tier', it is rated Amber. Additionally, video generation is a high-cost inference service billed per run, where retry code 75 might multiply consumption; audio tracks and (claimed) reference assets are sent off-premises to a third-party cloud, leaving the confidentiality of unreleased voiceovers/audio up to a third party. The skill itself consists purely of markdown with no scripts, obfuscation, or hidden endpoints, thus avoiding the Red tier.

5Second-Pass Verification

[ok] skill 目录内容 = 仅 SKILL.md — git ls-tree HEAD 与 ls -a 双向核实；仓库根仅 LICENSE + README.md + 31 个技能目录，无 scripts/、plugins/、.claude/。
[ok] 全部外部依赖条目回查调用点真实存在 — runcomfy CLI 调用点 4 处（默认/带音轨/字面控制示例）、npm 安装 1 处、npx skills 1 处、model-api 端点 1 处（How it works）、下载白名单 1 处（Security & Privacy）、文档/模型页/GitHub 链接多处，均实读命中。
[discrepancy] 『multi-reference conditioning / 最多 5 个参考素材』与输入 schema 一致 — description 写着 multi-reference conditioning，Where it shines 写 'Up to 5 reference media (image / video / voice)'，Limitations 写 'Reference media cap 5 (image + video + voice combined)'；但『Endpoints + input schema』表只有 prompt / audio_url / aspect_ratio / resolution / duration / negative_prompt / enable_prompt_expansion / seed 八个字段，没有任何 image/video 参考输入。结论：文档宣称了 schema 无法表达的能力（要么 schema 表不完整、要么能力被夸大），照 schema 写调用的 agent 拿不到该能力。
[discrepancy] 路由建议的目标模型在本仓有对应技能 — 'Cinematic motion editing on existing footage | Kling Video O1' 与 'Ultra-fast iteration | LTX 2' 两个建议目标在仓库内没有同名技能目录（全仓无 ltx-* 目录；kling-3-0/SKILL.md 自称 'Kling 3.0 (also called Kling V3.0)'，未使用 'Kling Video O1' 这一名称，该名称只出现在 wan-2-7/seedance-v2/happyhorse-1-0/ai-video-generation 四份文档里）。路由表能指向自家技能（Seedance/HappyHorse），但这两条是悬空引用。
[unlocatable] Security & Privacy 的技术断言可在本仓证实 — token 权限 0600、2 GiB 下载上限、CLI 不做 shell 展开、下载白名单、退出码语义均为仓外 @runcomfy/cli 行为；仓内无 CLI 源码或版本 pin，只能作为文档自述引用。同仓 runcomfy-cli/SKILL.md 是同一套自述（同源转述，非独立证据）。
[discrepancy] 安装一行与 GitHub 链接指向旧仓 agentspace-so/runcomfy-skills — SKILL.md 安装行写 `npx skills add agentspace-so/runcomfy-skills --skill wan-2-7 -g`，GitHub 链接指向 agentspace-so/runcomfy-skills；本仓储物 pin 在 prime-skills/runcomfy-agent-skills（README 亦用 runcomfy-agent-skills）。gh api：agentspace-so/runcomfy-skills 仍存在（5★、最后推送 2026-04-29、含 wan-2-7 目录），是另一份较旧仓；prime-skills/runcomfy-agent-skills 与 agentspace-so/runcomfy-agent-skills 为同一仓（gh 对两者返回相同 stats，full_name=prime-skills/runcomfy-agent-skills）。按文档安装可能装到旧版。
[ok] 音轨规格纪律在正文各处一致 — schema 行（WAV/MP3, 3–30s, ≤15MB）、Prompting 段（'`audio_url` must be 3–30s, ≤15MB, WAV/MP3. Out-of-range files reject.'）、Limitations 段（'Audio specs — 3–30s, ≤15MB, WAV/MP3 only.'）三处口径一致。
[ok] 时长/画幅/分辨率口径一致 — duration 2–15（default 5）与 Limitations 'Duration cap 15s' 一致；aspect_ratio 5 值与 'only the 5 documented values' 一致；resolution 720p/1080p 与 'No native 4K — 1080p ceiling' 一致。

6Verdict & Scenarios

Thoroughly explains the differentiating capability of 'external audio track-driven lip-sync': field specifications (3–30s/≤15MB/WAV-MP3), behavior when omitted (automatic music scoring), discipline of matching voiceover length, and multilingual voiceover workflows, allowing correct invocation on the first attempt.

Front-loads all parameter conventions and defaults into a single table: 8 fields, default 5s/1080p/16:9, expansion enabled by default but toggleable, eliminating the need for agents to consult external documentation.

Prompting methodology provides actionable criteria (camera language front-loaded, single primary action, concrete negative prompts, trade-off between expansion toggle and literal control) accompanied by three prompt examples.

Clearly delineated capability boundaries: no speech generation, capped at 15s/1080p, strict audio rejection rules, and explicit recommendations for alternative models, minimizing trial-and-error costs.

Zero code, zero assets: A single SKILL.md permits complete auditing without script execution surfaces.

Best for: Suitable for: Multilingual video variants requiring custom voiceovers/narration (swapping audio_url with the same prompt for multilingual releases), ad and explainer segments where lip-sync must match a designated audio track, and automated video pipelines relying on fixed seeds for variant comparison. Prerequisite: The runcomfy CLI must be installed and logged in (or configured with RUNCOMFY_TOKEN).

Not for: Not suitable for: Scenarios requiring the model to generate its own dialogue speech (switch to Seedance 2.0 Pro), continuous shots exceeding 15s (split into multiple scenes and stitch), 4K delivery requirements (1080p ceiling), or any scenario requiring data to remain entirely on-premises—all inference runs on the RunComfy cloud, with only JSON and generated artifacts passing through the local machine. Also unsuitable for workflows expecting multi-image or multi-video reference conditioning (the schema does not support it; see Risk #1).

Install Agent-Ready

① Verified Mirror 2026-09-15

Option A · Direct DownloadDownload wan-2-7.tar.gz

sha256: 63f9073a45acf23a…

Option B · Copy Guide for Agent

install.json

Agents automatically download and verify mirrors from this JSON guide.

② Upstream GitHub

GitHub Repo ↗

Pinned to commit fca19ae084.