Much of the dtwin ecosystem was built with AI agents doing a large share of the typing. What made that work was not one clever prompt. It was a division of roles, and the habit of writing my standards down so every agent follows them.
#Roles
| role | does | does not |
|---|---|---|
| orchestrator (main session) | plans, decomposes, makes the hard calls, writes risky code | read 40 files to find one line |
| deep reasoner | root-cause hunts across many files and logs, returns conclusions with evidence | edit anything |
| fast worker | mechanical, fully specified edits, runs builds and tests | make design decisions |
| critic | tries to break a plan or a diff, reports only issues that change behaviour | style nitpicks |
Independent work goes out in parallel. Anything subtle comes back to the orchestrator before it is written.
#Skills: standards that do not drift
Every time I correct an agent twice for the same thing, the correction becomes a skill: a short document the agent loads before that kind of task. Examples from my AI page:
- uch-d: my rules for 3D on the web (tiers, fallbacks, draw calls).
- artefakt: how to document a task so a reviewer can reproduce every claim.
- prezentatsiya: an offline, step-by-step HTML deck format.
#Guard rails
- Agents get read-only access to company systems through a gateway.
- Nothing irreversible (deploys, deletes, messages) happens without a human yes.
- Every claim in a report must point to a query, a log line or a file and line.
The speed comes from parallel work. The quality comes from the critic and from standards written as skills.