krishanraja

video-engine

"Use this skill whenever the current user message, after trimming leading and trailing whitespace, equals 'Video engine' case-insensitively, so the skill can validate the launch guard. Discovery is not activation: after loading, activate only if that exact message is also the first user-authored message in a new chat. Terminal punctuation, additional words or lines, quoted mentions, later-turn uses, and $video-engine must not activate the engine. On a failed guard, stop before accessing the authority repository. Never activate for ordinary video requests."

krishanraja 0 Updated 2w ago

Resources

1
GitHub

Install

npx skillscat add krishanraja/ai-harness/video-engine

Install via the SkillsCat registry.

SKILL.md

Video Engine launcher

Use this as a thin launcher. GitHub krishanraja/content-engine main is the only authority for code, configuration, operational instructions, and durable learning.

Trigger contract

This section governs activation, not discovery or semantic retrieval. The catalog must select this skill whenever the current prompt itself is the exact phrase so this file can perform the history check; catalog routing must not try to infer whether a conversation is new. A client may also retrieve this guard while deciding whether a nearby phrase matches. Either read is not a launch. Activation begins only when the check below passes and the client accesses, fetches, or runs the authority repository or its engine surface. On a failed match, do not cross that boundary.

Activate a new Video Engine session only when both are true:

  1. The current message is the first user-authored message in a new chat.
  2. After removing leading and trailing whitespace only, its complete contents equal Video engine, case-insensitively.

Video engine, VIDEO ENGINE and Video engine pass. Everything else fails, including Video engine!, Video engine please, a message carrying another line, a quoted mention, $video-engine, and the exact words on a later turn.

This matches krishanraja/content-engine main exactly, which this file names as the only authority. The trigger was narrowed on 2026-09-08 while the same repository still used its former krishanraja/mindmake-video-studio identity, after the SURFACE reconciliation caught the standing collision. The 2026-09-15 rename changes the authority address, not what that older evidence proved.

  • Do not activate from phrases such as video edit, edit this video, generate a video, my video engine, how does the video engine work?, or any unrelated mention of video production.
  • If this skill was selected but the trigger contract is not satisfied, stop applying it immediately. Do not fetch the repository, run the CLI, inspect Video Engine state, or redirect the request. Handle the request normally with the relevant general capability.
  • After a valid launch, follow-up turns in that same chat may continue the active Video Engine workflow without repeating the launch phrase.

Start every session

  1. Obtain the latest main commit from GitHub into a disposable checkout under the current task's work/ directory. Never use an unrelated or stale local clone as authority.
  2. Read the checkout's AGENTS.md and .agents/skills/mindmake-video/SKILL.md completely before operating the engine.
  3. Use drive_root/Ventures/Active/Mindmaker/04_Content/Video Engine as the media inbox and base path, with drive_root resolved for your surface from contract/paths.yaml. Where that root is null the engine has no media surface: say so and stop rather than substituting a local directory. Media stays outside GitHub.
  4. Run studio doctor and the job-list form of studio status. Report blockers plainly; never bypass a hard gate.
  5. If the user only said Video engine, return a concise operating brief with current system health and the most useful next choices: weekly radar, create a video, resume/review a job, or import feedback/analytics. Recommend the strongest next action rather than asking a context-free question.

Authority and safety

  • Treat this installed launcher only as a pointer. When it conflicts with the repository, follow the latest GitHub main.
  • Fetch provider signals only through the repository's authenticated read-only adapters. Fail closed if credentials or schema versions are invalid.
  • Never put media, credentials, tokens, raw private records, OAuth state, jobs, or derived indexes in GitHub.
  • Never publish publicly. YouTube uploads are private-only and LinkedIn output is a local draft package.
  • Preserve the three approval gates and the feedback confirmation loop. No durable preference becomes active without explicit user approval.