I do all of my own development with AI agents, but with rules: every repository has a binding operating contract for the agents, architectural decisions are written down, and nothing is declared done before the tests have run. I set up the same in one of your repositories and teach your team to work with it.
These run in every project of mine and are published on my CV.
A single binding operating contract (AGENTS.md) mirrored to every AI coding tool (Claude Code, Copilot, OpenCode, Cursor, Windsurf, Codex, Aider) so the work stays repeatable regardless of which tool does it. On top of that a three-layer Obsidian knowledge vault (modules, sublists, specs) and architectural decisions as ADRs.
Prompts are versioned artefacts rather than strings buried in code: YAML metadata, a version history and test cases that a CLI runs as evals and compares across versions.
Implements the LATM loop ("LLMs as Tool Makers"): recognise → specify → build → test → document → register → use. At session start the tool catalogue is injected into context; at task end a cheap local heuristic decides for free whether some subtask was worth capturing as a tool, and if so, a detached builder agent writes it with tests and registers it.
The week starts with one repository and ends with your team using agents on its own tickets.
A 30 minute call: what slows the team down, which tools are already in use and what agents must never do.
An operating contract for agents (AGENTS.md and CLAUDE.md), a decision log, one command for the tests and a review checklist. All of it as a pull request in your repository.
Two hours on real tickets: how to hand over a task, how to check the result and when not to use an agent. I answer questions for a week afterwards.
Code that looks finished but that nobody checked: invented APIs, skipped tests and repetition. Rules and tests before done are the cure.
No. The same contract works in Claude Code, Cursor, Copilot and others. I set it up for your tools.
The work happens in your environment and on your accounts. The rules define what agents must never do.
Best of all: a team of two to ten developers benefits fastest, because the rules are shared by everyone.
The repository and the rules stay with you. If you want the same in other repositories or extra hands for projects, we agree separately.
Tell me the team size, the tools you use and one repository to start from. I answer the same working day.
jyri.gromyko2002@gmail.com
+358 44 978 2028