I treat AI agents like a small team (how). These are the skills they load before a kind of task. Each one is a short written standard that came from correcting the same mistake twice.
#Skills
| skill | what it makes an agent do |
|---|---|
| uch-d | 3D on the web my way: GPU tiers, WebGL fallbacks, draw-call budgets, context-loss handling, real-GPU checks |
| artefakt | a task dossier a reviewer can reproduce: every claim comes with its query, log line or file and line |
| prezentatsiya | one offline HTML deck with step-by-step motion, separate technical and business paths, and speaker notes |
| overview | a product home page plus a story-style overview deck, in the style of dtwin.uz |
| sayt-pack | a router: which design, 3D, performance or testing skill to load for a website task |
| app-tezlik | a performance audit checklist across frontend and backend |
| system-architect | explain a legacy system with diagrams: context, sequences, failure modes |
| animatsiya-uslublari | visual styles for vertical explainer videos |
#Agent roles
- orchestrator: plans and makes the risky calls.
- deep reasoner: read-only root-cause analysis, returns evidence.
- fast worker: fully specified mechanical edits and test runs.
- critic: tries to break plans and diffs, reports only what changes behaviour.
#Rules every agent follows
- No secrets, internal hostnames or personal data in anything public.
- Nothing irreversible without a human yes.
- Prove it: no claim without a source you can rerun.