Project 04

Agent Skills,
Packaged & Shared.

Two small open-source AI Agent projects that came from my own engineering-learning friction. video-transcribe-skill uses Python, FFmpeg, and local Whisper to turn unsubtitled tutorials into transcript evidence before Codex summarizes, extracts, or answers questions. codex-agent-custom-instructions documents the working rules behind that process: clarify first, make small edits, keep deterministic steps in scripts, and verify the result.

Codex Skill Python AI Agent GitHub
video-transcribe-skill custom instructions
2 public GitHub repositories used as evidence of self-driven AI Agent practice
Python used for deterministic video/audio routing and local transcription tooling
MIT license used for the small open-source skill and instruction resources
What I built

video-transcribe-skill

I built this for a real engineering learning problem: many useful mechanical, robotics, and software tutorials have no subtitles, so an AI agent needs transcript evidence before it can summarize or answer questions reliably. The skill resolves video sources, extracts audio with FFmpeg, runs local faster-whisper transcription, and outputs transcript plus metadata files.

custom instructions

A bilingual checklist that turns my own coding-agent practice into reusable rules: state assumptions, clarify ambiguity, keep edits small, put routing and retry logic in deterministic code, surface failures, and verify the result instead of only claiming completion.

Self-driven loop

The value is the loop: I found a repeated bottleneck in my own engineering learning, built a small tool around it, separated scripts from model judgment, wrote down the method, and made the result public. That process is stronger evidence of self drive than a sentence claiming I learn fast.

Workflow architecture
Design principle

Keep fragile steps out of model judgment.

The video skill was designed around one rule: deterministic code should do platform routing, download attempts, audio conversion, path checks, and transcript writing; the model should only reason after transcript text exists.

01

Input routing

  • Bilibili URLs and BV ids use public Bilibili APIs.
  • YouTube and generic URLs fall back to yt-dlp.
  • Local files go straight into the conversion path.
02

Local processing

  • FFmpeg normalizes audio for transcription.
  • faster-whisper runs locally without cloud transcription keys.
  • The output includes transcript.txt and metadata.json.
03

Agent use

  • Codex reads the transcript instead of pretending to watch the video.
  • Summaries mention quality caveats when ASR is imperfect.
  • The same route can support notes, extraction, or Q&A.
Skills learned

What these projects show

01

Packaging repeatable workflows

I learned to separate deterministic tooling from model judgment: scripts handle download, routing, conversion, transcription, and file checks; Codex reads the transcript for summaries, extraction, and Q&A only after evidence exists.

02

Writing for reuse

I practised turning a personal setup into README-first documentation, install steps, command examples, repository layout, limitations, and source notes that another user can inspect and adapt.

03

Self-directed iteration

Both repositories came from repeated friction in real learning and coding work, then turned that friction into public tools. The projects are small, but they show the habit I want to keep building: notice a workflow problem, make the tool, document the method, and verify the result.

Source repositories: video-transcribe-skill and codex-agent-custom-instructions under github.com/UKplus8HRS.