01
Specs first, always
Before any work starts, there is a written spec: what we build, why, what "done" means and what must not change. Agents work from the spec, not from a chat history, so any agent can pick up the work and every result can be checked against it.
02
Skills for every kind of work
I give agents skills: written playbooks for a kind of work. Brainstorming and specs, planning, test-driven development, debugging, code review and verification before anything is called done. On top of those, domain skills for the stack: SwiftUI, Supabase, Vercel.
03
How I work with agents
- I set the goal and the limits. The agent does the work and shows evidence before it says "done": screenshots, test runs, checks on the live site.
- Nothing irreversible without my word: deleting a project, publishing copy in my name, spending money.
- Facts stay facts. The fit check is told never to invent experience. When it added a detail that is not in my CV, we tightened the rules and tested again.
- Agents in parallel. I run Claude Code, Codex and T3 Code side by side, each with its own task and its own test environment.
04
What shipped in that session
Live on agoshkov.comDeployed to Vercel from the command line, with the domain, www and DNS set up, and light and dark themes.
Check the fitPaste a job description or a link. Claude scores my fit on six angles using only my CV, with strengths, gaps, interview questions and a pitch for your team.
An MCP serverAt
agoshkov.com/mcp, so your own AI assistant can read my profile and run the same check.