Skip to content

README.md

Latest commit

 

History

History
100 lines (78 loc) · 4.73 KB

README.md

File metadata and controls

100 lines (78 loc) · 4.73 KB
Raw

Watch Skill documentation

This directory contains setup guides, operating references, design notes, and workflow recipes. If this is your first visit, start with Getting started and then connect your client through the agent matrix.

Choose by task

I want to… Read
Install and index a first video Getting started
Connect an AI coding agent Agent compatibility
Look up a tool or endpoint MCP tool reference
Configure storage, models, or privacy Configuration
Fix an installation or runtime issue Troubleshooting
Understand cost and routing Cost policy
Decide whether an agent run actually succeeded Verification
Compare against other options Comparison
Move over from claude-video Migration guide
Build capture–critique–verify workflows THE LOOP
Apply Watch Skill to a job Use-case packs
Extend or review the engine Architecture and decisions
Find contribution opportunities Roadmap and contributing guide

Guides

  • YouTube analysis — captions, visual evidence, focused windows, and follow-up questions.
  • Arabic in, Arabic out — script-aware OCR and cross-language retrieval.
  • Live browser — pixels and structured page evidence at once, navigation policy, redaction, and what a page is never allowed to say.
  • The MCP App — the live workspace: what it shows, how to open it, the transport it negotiates, and what it will never render.
  • The Observer Loop — declare success first, correct under explicit approval, and let a separate process decide whether it worked.
  • THE LOOP — capture, criteria, iteration, and proof artifacts.
  • Lessons and savings — corrections, evaluation, caching, and token accounting.
  • How self-improvement works — the complete local lesson lifecycle without anthropomorphic claims.

Use-case packs

Packs combine existing tools into repeatable workflows. They do not add a second API.

Runnable counterparts live in the example catalog.

Reference and design

Architecture

  • Architecture — boundaries, data flow, and extension points.
  • Engineering decisions — the trade-offs behind those boundaries, with the measurements that settled them.

Browser and verification

  • Browser Runtime — operator and observer modes, target resolution, action receipts, recovery, and the false-success benchmark.
  • Verification — contracts, deterministic checks, assurance levels, attestations.
  • Live watching — sources, bounded pipelines, cursors, finalisation.
  • Capture capabilities — what a given machine can actually record, and how each answer was established.

Interfaces

  • MCP tools — all 37 tools and their REST/CLI twins.
  • Configuration — the source of truth for WATCHSKILL_* settings.

Evidence

  • Release proof — test results, browser benchmark, packaging and security checks, and the limits of each.
  • VLM performance — the measured local vision model: revision, latency, memory, and the failure modes a live session exposed.
  • Benchmarks — methods, fixtures, and committed results.

Development

  • Testing tiers — what the offline suite proves, and what only a real model can.
  • Roadmap — what ships today, what is missing, and what is planned.

Documentation standards

Commands should run from the repository root unless a guide says otherwise. Integration statuses must use the definitions in the agent matrix. Measured claims need a dated benchmark or reproducible example, and configuration blocks must parse under templates/agent-adapter/validate.py.