Video Edit Skill

We read the skill at commit fca19ae of prime-skills/runcomfy-agent-skills and verified our catalog copy matches it byte for byte. One thing the gate turned up: the identical file also ships from three other places, and the skill's own install command points at one of them rather than at the repository this listing represents.

video-edit at a glance

1. It is a router, not an editor, and the routing table is the product.
2. Three paid models sit behind it, each with its own schema and constraints.
3. Every link carries campaign tracking, so the skill doubles as an attribution channel.
4. Preservation comes first in every prompt, which is the real craft lesson in the file.
5. One edit direction per call, because compound edits drift.
6. The same file ships from four places, and its install command points at one of the empty ones.

1. It is a router, not an editor

The description says so plainly: this skill "is a smart router that matches the user's intent to the right edit model in the RunComfy catalog". The body opens with the pitch:

Video edit, intent-routed. This skill doesn't lock you to one model [...]

Source: SKILL.md, opening

What follows is a nine-row table mapping user intent to model, with a stated reason per row: a talking-head restyle goes to Wan 2.7 for identity and motion preservation, motion transfer goes to Kling 2.6 Pro, a costume swap goes to Lucy Edit Restyle. The instruction after the table tells the agent exactly what to do with it:

The agent reads this table, classifies the user's intent, and picks the matching subsection below.

Source: SKILL.md, routing instruction

The table ends with a default row, Wan 2.7, "Most versatile, highest resolution", so an unclassifiable request still lands somewhere rather than stalling.

2. Three paid models, three schemas

Each route documents its model string, a full parameter table, and runnable invocations. Wan 2.7 takes a prompt, a video, an optional reference image, resolution, aspect ratio, duration and an audio setting, with hard limits stated: source video "2-10s, ≤100MB". Kling's motion control has a different shape entirely, including a character_orientation enum where the choice changes your maximum output length, ten seconds for image orientation and thirty for video.

This is the part a general-purpose agent cannot improvise. Model endpoints, enum values and per-model size caps are not guessable, and getting them wrong on a paid endpoint costs a failed run rather than a syntax error.

3. Every link carries campaign tracking

The header links out five times, and four of the five carry campaign parameters:

runcomfy.com

Source: SKILL.md, header links

utm_source=skills.sh, utm_medium=skill, utm_campaign=video-edit. Only the GitHub link is untagged. Campaign parameters let RunComfy attribute traffic to the skill registry; whether they also tie it to paid usage is not something the file tells us.

What makes this worth a section is the tension inside the file itself. Its Security and Privacy section is explicit about what the tooling does not do:

Outbound endpoints: only model-api.runcomfy.net (request submission) and *.runcomfy.net / *.runcomfy.com (download whitelist for generated outputs). No telemetry, no callbacks.

Source: SKILL.md, "Security & Privacy"

Both things are true at once. The skill phones nothing home, and its documentation links are attributed. Reading only one of those gives you the wrong picture of how it earns.

4. Preservation comes first in every prompt

The genuinely useful craft in this file is its prompting discipline, and the pattern is stated as a template:

Preservation goals first: "Preserve [face / pose / motion / framing / lip movement]; [then state the change]".

Source: SKILL.md, "Prompting tips"

Both Wan 2.7 examples follow it exactly. The background swap reads "Preserve the speaker's face, pose, and lip movement; change the background to a modern office with neutral lighting." The packaging swap reads "Maintain the original framing and hand movement; replace the packaging design using the reference image."

The other three do not. The Kling example's prompt is "A young american woman dancing", with no preservation clause at all, and both Lucy examples state the change before the preservation. So the rule is stated once and then followed in two of five worked examples, which is worth knowing if you are copying an example rather than the rule. Copy the rule: naming what must not change, before naming what should, transfers to any video model, including ones this skill does not route to.

5. One edit direction per call

The constraint that saves the most money is one line:

One edit direction per call. Compound edits drift on motion.

Source: SKILL.md, "Prompting tips"

Ask for a background swap and a color grade together and the model gives you drift on both. The file also gates reference_image behind justification ("Don't pass refs for general restyle") and recommends audio_setting: "origin" for talking heads so the soundtrack is not regenerated. Each of these is an iteration you do not pay for twice.

6. The same file ships from four places, and the install command picks the emptiest

Our catalog lists this skill under prime-skills/runcomfy-agent-skills. The skill's own header links somewhere else, and so does its install command:

npx skills add agentspace-so/runcomfy-skills --skill video-edit -g

Source: SKILL.md, install command

We went looking and found four byte-identical copies of this file, published under prime-skills, agentspace-so, doany-ai and runcomfy-com. Their install counts are wildly lopsided: roughly 406,000 on the copy this page tracks, and single or double digits on each of the other three. The command printed inside the skill sends you to one of the near-empty ones.

None of this is new to us. Our own clone census, taken across 79,848 listings in August 2026, found 3,443 verified copies holding 15.8% of every install counted, and named this exact generative-media family as the single largest cluster in the dataset. The census is the rigorous version of the observation, with its methodology and raw data published; this dive is just where you can watch it happen to one file.

The practical reading for you: the number on this page is the adoption of this listing, and for a skill published in four places that is not the same as the adoption of the skill.

Tools it drives: the RunComfy CLI (npm i -g @runcomfy/cli), authenticated with runcomfy login or a RUNCOMFY_TOKEN for CI, calling runcomfy run <vendor>/<model>/<endpoint>.

Behavior change, before and after:

The fine print worth knowing:

• The models are paid and the CLI needs an account, so this skill spends money by design. The routing and prompting rules exist partly to make you spend less of it.
• Every outbound link is campaign-tagged for RunComfy's analytics, and the install command sends you to a different GitHub organization than the one this page lists.

Should you install video-edit?

Install it if you edit video with an agent and you are already on RunComfy, or willing to be. The routing table and the preservation-first prompting are real expertise, compressed, and both survive being read even if you route to a different vendor. Skip it if you have no RunComfy account, since every route terminates in a paid endpoint. And read it once as a document regardless: it is the clearest worked example we have found of a vendor turning a model catalog into an agent-facing product.

npx skills add prime-skills/runcomfy-agent-skills --skill video-edit -g

(That installs the copy this page tracks. The command printed inside the skill points at agentspace-so/runcomfy-skills instead, which is the same file with a fraction of the adoption.)

Quotes verified against prime-skills/runcomfy-agent-skills at commit fca19ae on August 20, 2026. Related: the agent skill clone census. Previous: how supabase-postgres-best-practices actually works. All dives: the series index.