OR Key
drop another .md file to compare - side-by-side diff against video

video

Processes your videos for you.
description: "Triggers on prompt mention of 'video' as a processing skill."
personal 2 files

What it does for you

Processes your videos for you.

What it produces

A recent result, so you can see the kind of work it returns.

loading…

How to get it

These run inside the Snappy workspace. Want this working in your business? I set skills like this up with you, in one focused week.

Work with me
For developers how this skill is built, graded, and how it runs

at a glance- the short version

actorExported functions in state/lib/video.ts.
auditorNone wired yet - eval is manual (Robert review).
eval modeshape
categoryContent
stages1

what's inside - the parts that make up a skill 3/4 present

A skill is just a few plain-text files. Only the main one is required. The rest are optional, added as the work needs them. This is what the skill is made of; how it runs is just below.

The skill
state/skills/video/SKILL.md present
the skill itself, in plain text
The main file. It says what the skill is and lays out the steps in plain English.
Code
state/lib/video.ts present
code the skill can run
Reusable code this skill can call when it needs to.
Scripts
state/bin/video/ not present
helper scripts
Optional. Added when a skill has a few commands to run.
Loader
state/skills/video/AGENTS.md present
what the AI loads on the fly
Loaded automatically the moment this skill is needed. Kept short on purpose.

how it's graded - what counts as a good run 5 criteria · 5 judge

Each row is one thing a good run has to get right. deterministic means a quick check decides, pass or fail. judge means the AI reads the result and rates it. Grading each piece on its own (instead of one overall score) shows exactly where a run fell short, so the fix is obvious.

name
kind
check
transcription_succeeded
judge
The transcribed text from transcribe(transcribe_input) accurately reflects the audio content of the video.
caption_correctness
judge
The generated captions from caption(caption_input) are synchronized with the video and are grammatically correct.
clip_honesty
judge
The video segment produced by clip(clip_input) starts and ends precisely at the specified timestamps without unexpected artifacts.
resize_integrity
judge
The resized video from resize(resize_input) maintains its aspect ratio and visual quality at the new dimensions.
audio_extraction_quality
judge
The extracted audio from extractAudio(extractAudio_input) is clear and complete, matching the audio content of the input video.

how it runs - the shared frame every skill uses 4/5 present

Every skill runs the same way. One part does the work, a separate part checks it, and a short loader hands the AI exactly what it needs for the job. Anything this skill doesn't use shows a one-line note saying why, on purpose, not by accident.

makes the work The worker
present
Exported functions in state/lib/video.ts. the worker
Does the actual work. Whatever it produces is what gets checked next.
checks the work The reviewer
present
None wired yet - eval is manual (Robert review). the checker
A separate checker grades the work, so the part that made it can't approve its own work.
frame
learns Self-correction
not present

This skill doesn't fix its own gaps yet.

tidies up Background fixes
present
queued for rewrite runs in the background
Bigger fixes that can't be made on the spot get queued and rewritten in the background later.
remembers Run history
present
state/log/pending-eval.ndjson pending runs
Every run is written down here, then reviewed by hand each week.
Critical rules the things this skill must not get wrong
  1. Pipeline runs on Mac Mini via SSH - calls block on remote completion; do NOT inline-block the agent for long renders, background and report a job id
  2. ALWAYS use the typed lib state/lib/video.ts - do not shell out ffmpeg ad-hoc (memory: existing-infra-first)
  3. Eval is manual (Robert review) - every run files to state/log/pending-eval.ndjson
  4. Always background long operations — return a job id, do not block the turn
  5. Compose surface before asking for input — users fill fields in the surface, not in chat
  6. Transcripts render inline in the surface, not as a file attachment
  7. +1 more in AGENTS.md →

what it has learned - fixes written back in over time sample

When a run hits something this skill didn't handle, the fix gets written back into the skill so it doesn't happen again. FIXED means it was corrected on the spot. LOGGED means it's queued for a bigger rewrite. Either way, the skill gets a little better and never makes the same mistake twice.

  1. Loading feedback rows…

how the work flows- who makes it, who checks it

actor Exported functions in state/lib/video.ts.
auditor None wired yet - eval is manual (Robert review).
1 data
eval log
`state/log/pending-eval.ndjson` (manual)

SKILL.md- the skill, written out in plain English

video

Video processing pipeline via Mac Mini SSH.

Ported from kernel snappy-video in Phase 0.5. See state/lib/video.ts for the full API surface.

Steps

  • transcribe() - see state/lib/video.ts
  • caption() - see state/lib/video.ts
  • clip() - see state/lib/video.ts
  • resize() - see state/lib/video.ts
  • extractAudio() - see state/lib/video.ts

Eval

Actor: the exported functions in state/lib/video.ts. Auditor: none wired yet - eval is manual (Robert review). File a state/log/pending-eval.ndjson row on each run.

Score convention:

OutcomeScore
Pass on first try1.0
Failed first, auto-fix applied, re-check passed0.5
Still failing or unrecoverable0.0

Gotchas

via the Phase 0.5 driver. Only these rewrites were applied: already in state/lib/)

  1. realpathSync(process.argv[1]) CLI guard wrapped in try/catch
  • See the kernel SKILL.md for the original long-form guidance if you need it

(read-only reference at the kernel path above).

Graduation

This skill is prose. Graduate by defining a deterministic auditor and flipping eval: auto.

Rubric

criteria:
  - name: transcription_succeeded
    kind: judge
    check: "The transcribed text from transcribe(transcribe_input) accurately reflects the audio content of the video."
  - name: caption_correctness
    kind: judge
    check: "The generated captions from caption(caption_input) are synchronized with the video and are grammatically correct."
  - name: clip_honesty
    kind: judge
    check: "The video segment produced by clip(clip_input) starts and ends precisely at the specified timestamps without unexpected artifacts."
  - name: resize_integrity
    kind: judge
    check: "The resized video from resize(resize_input) maintains its aspect ratio and visual quality at the new dimensions."
  - name: audio_extraction_quality
    kind: judge
    check: "The extracted audio from extractAudio(extractAudio_input) is clear and complete, matching the audio content of the input video."

AGENTS.md- what the AI loads when this skill comes up

video - loader

Per-turn rules for the video skill. Full reference: state/skills/video/SKILL.md. Do not skip these.

Critical Rules

  • Pipeline runs on Mac Mini via SSH - calls block on remote completion; do NOT inline-block the agent for long renders, background and report a job id
  • ALWAYS use the typed lib state/lib/video.ts - do not shell out ffmpeg ad-hoc (memory: existing-infra-first)
  • Eval is manual (Robert review) - every run files to state/log/pending-eval.ndjson

Commands

| ui model | live composition via compose_inline, persisted as artifact lang_body, reopened with OpenArtifact | |library: state/lib/video.ts - transcribe(), caption(), clip(), resize(), extractAudio() |eval log: state/log/pending-eval.ndjson (manual)

Known Pitfalls

  • Long renders (transcribe, caption burns) can take minutes - background them immediately, return a job id, poll separately
  • Do not block the agent turn on SSH completion for anything over ~30s

Compose Surface

When the user asks to process, transcribe, clip, or caption a video, compose a surface immediately - do NOT exchange plain text questions for inputs.

Phase 1 - compose intake surface:

root = Stack([header, file_form, options, actions])
header = TextContent("Video pipeline", "medium-heavy")
file_form = Card([
  Stack([
    Input("<path or URL>", "Video file / URL"),
    Select("Operation", ["Transcribe", "Caption", "Clip", "Resize", "Extract audio"])
  ])
])
options = Card([
  Stack([
    Input("00:00:00", "Clip start (HH:MM:SS)"),
    Input("", "Clip end (HH:MM:SS)"),
    Input("1280x720", "Target resolution (resize)")
  ], "horizontal")
])
actions = Stack([run_btn], "horizontal")
run_btn = Button("Start", @Run("run_pipeline"), "primary")

Phase 2 - after job dispatched (async):

status = Stack([
  TextContent("Job dispatched", "medium-heavy"),
  StatCard("Job ID", "<id>"),
  StatCard("Operation", "<op>"),
  StatCard("Status", "running"),
  Button("Check status", @Run("poll_job"), "secondary")
])

Phase 3 - job complete:

  • Transcription: show transcript inline in a scrollable MarkDownRenderer block with copy button
  • Caption: show output path + offer to preview first 30s
  • Clip/resize: show before/after file sizes via StatCard, output path with copy button

Critical Rules

  • Always background long operations - return a job id, do not block the turn
  • Compose surface before asking for input - users fill fields in the surface, not in chat
  • Transcripts render inline in the surface, not as a file attachment
  • Log to state/log/pending-eval.ndjson after each dispatched job

Self-Test

An agent reading this should correctly:

  1. [ ] Use state/lib/video.ts rather than raw ffmpeg shells?
  2. [ ] Compose a surface with input fields rather than asking for path/op in plain text?
  3. [ ] Background long renders instead of blocking the agent on SSH completion?
  4. [ ] Show job status as a live surface the user can poll?
  5. [ ] File a pending-eval.ndjson row since the auditor is still manual?

Found a gap? Edit this file. <!-- footer-injection-point -->

api.ts- the code it can call

#!/usr/bin/env npx tsx
/**
 * snappy-video/api.ts -- Video processing pipeline via remote SSH render box.
 *
 * All heavy processing (Whisper, ffmpeg, caption-video.sh) runs on the remote host.
 * This module wraps SSH commands.
 *
 * Environment variables:
 *   SNAPPY_VIDEO_SSH_TARGET  SSH target for the remote render box (e.g. user@host.local).
 *                            Required for all remote operations. If unset, remote calls
 *                            return a structured error string instead of throwing.
 *   SNAPPY_ROBOT_ROB_DIR     Path to the robot-rob project on the remote host.
 *                            Defaults to ~/robot-rob if unset.
 *
 * Usage:
 *   npx tsx api.ts transcribe <path>
 *   npx tsx api.ts caption <path>
 *   npx tsx api.ts clip <path> <start> <duration>
 *
 * Or import as module:
 *   import { transcribe, caption, clip } from "./video.ts";
 */

import { execSync } from "child_process";
import os from "os";
import path from "path";
import { env } from "./env.ts";
import { realpathSync } from "fs";

const MAC_MINI = process.env.SNAPPY_VIDEO_SSH_TARGET ?? null;
const ROBOT_ROB = process.env.SNAPPY_ROBOT_ROB_DIR ?? path.join(os.homedir(), "robot-rob");

const REMOTE_UNAVAILABLE = "feature unavailable: set SNAPPY_VIDEO_SSH_TARGET to a valid SSH target";

function ssh(command: string, timeoutMs?: number): string {
  if (!MAC_MINI) return REMOTE_UNAVAILABLE;
  const full = `ssh ${MAC_MINI} "${command.replace(/"/g, '\\"')}"`;
  try {
    return execSync(full, {
      encoding: "utf-8",
      ...(timeoutMs != null ? { timeout: timeoutMs } : {}),
      stdio: ["pipe", "pipe", "pipe"],
    }).trim();
  } catch (err) {
    const e = err as { status?: number; stderr?: string; message: string };
    throw new Error(`Remote command failed (exit ${e.status}): ${e.message}${e.stderr ? ` stderr: ${e.stderr}` : ''}`);
  }
}

// --- Public API ---

export function transcribe(videoPath: string, model = "small"): string {
  if (!MAC_MINI) return REMOTE_UNAVAILABLE;
  const cmd = `cd ${ROBOT_ROB} && source venv/bin/activate && whisper "${videoPath}" --model "${model}" --output_format srt --output_dir /tmp/`;
  return ssh(cmd);
}

export function caption(
  videoPath: string,
  options: { style?: string; words?: boolean; clips?: boolean; outputPath?: string } = {}
): string {
  if (!MAC_MINI) return REMOTE_UNAVAILABLE;
  const outPath = options.outputPath || videoPath.replace(/(\.[^.]+)$/, "-captioned$1");
  const flags: string[] = [];
  if (options.style) flags.push(`--style "${options.style}"`);
  if (options.words) flags.push("--words");
  if (options.clips) flags.push("--clips");
  const cmd = `cd ${ROBOT_ROB} && ./caption-video.sh "${videoPath}" "${outPath}" ${flags.join(" ")}`;
  return ssh(cmd);
}

export function clip(videoPath: string, startTime: string, duration: string, outputPath?: string): string {
  if (!MAC_MINI) return REMOTE_UNAVAILABLE;
  const outPath = outputPath || `/tmp/clip-${Date.now()}.mp4`;
  const cmd = `ffmpeg -y -ss ${startTime} -t ${duration} -i "${videoPath}" -c copy "${outPath}"`;
  return ssh(cmd);
}

export function resize(videoPath: string, format: "9:16" | "16:9" = "9:16", outputPath?: string): string {
  if (!MAC_MINI) return REMOTE_UNAVAILABLE;
  const outPath = outputPath || videoPath.replace(/(\.[^.]+)$/, `-${format.replace(":", "x")}$1`);
  const vf = format === "9:16"
    ? "scale=1080:1920:force_original_aspect_ratio=decrease,pad=1080:1920:-1:-1:color=black"
    : "scale=1920:1080:force_original_aspect_ratio=decrease,pad=1920:1080:-1:-1:color=black";
  const cmd = `ffmpeg -y -i "${videoPath}" -vf "${vf}" -c:a copy "${outPath}"`;
  return ssh(cmd);
}

export function extractAudio(videoPath: string, outputPath?: string): string {
  if (!MAC_MINI) return REMOTE_UNAVAILABLE;
  const outPath = outputPath || videoPath.replace(/\.[^.]+$/, ".m4a");
  const cmd = `ffmpeg -y -i "${videoPath}" -vn -c:a aac -b:a 192k "${outPath}"`;
  return ssh(cmd);
}

// --- CLI ---

if ((() => { try { return import.meta.url === `file://${realpathSync(process.argv[1])}`; } catch { return false; } })()) {
  (async () => {
    const [, , cmd, ...args] = process.argv;

    switch (cmd) {
      case "transcribe": {
        const [path, model] = args;
        if (!path) { console.error("Usage: api.ts transcribe <path> [model]"); process.exit(1); }
        console.log(transcribe(path, model || "small"));
        break;
      }
      case "caption": {
        const [path] = args;
        if (!path) { console.error("Usage: api.ts caption <path>"); process.exit(1); }
        console.log(caption(path, { style: "bold", words: true }));
        break;
      }
      case "clip": {
        const [path, start, duration] = args;
        if (!path || !start || !duration) {
          console.error("Usage: api.ts clip <path> <start> <duration>");
          process.exit(1);
        }
        console.log(clip(path, start, duration));
        break;
      }
      default:
        console.log("Usage: npx tsx api.ts [transcribe|caption|clip] ...");
    }
  })();
}

scripts- helper scripts it can run

prose-only skill - 1 inline code block live in SKILL.md above (no state/bin/ sidecar yet).

how we check it- the checks, plus the last 10 runs

rubric shape schema-shape check (no inline rubric)
recent no runs actor/auditor: unverifiable
deps none declared

no recent runs logged - the eval contract is declared but nothing has been graded yet