A method for managing AI agents
Operchief
Project Method
Two Operchief postulates“AI systems have crossed the threshold of human cognitive capability.”“An LLM-based AI agent is a highly erudite probabilistic system in which context loss and plausible fabrication are intrinsic operating modes, not malfunctions.”Operchief Project Method follows from these two postulates. It is a way of working with an AI agent so that a person runs the project, not the agent.
Operchief Project Method (OPM) is a method for non-professionals who work with AI agents on personal and hobby projects of any kind. Professionals use similar approaches, but usually for one kind of work, one agent, or one environment. OPM is simpler and more general: it works with any current AI, in any setting.
At its core, OPM is a set of strict rules for the agent: do not change project files without permission, plan the work in detail, and record the results in logs. The aim is to turn work with an agent into a project a person can actually run: simple, inexpensive, and usable by non-professionals.
In that sense OPM goes against the mainstream of vibe coding. It is not “make me a good project from one prompt.” It is careful work shared by a person and an agent: the work is split into stages, each stage stays under control, and the result is brought to a state that can be checked.
OPM fits any work done with AI: build an Android app, automate a routine at work, write a marketing plan, run a YouTube channel, draft an article, put together an Arduino prototype, translate a book, find the best recipe for a Sunday roast, or carry any other personal project that needs several rounds of revision and checking.
The method can cut the cost of the work and make better use of an AI subscription. A long chat makes the model read the old context again and again. A person hesitates to open a new chat, for fear of losing the thread, and tens or hundreds of thousands of tokens — and the money with them — go to the tail of the conversation. Step away from the computer and the server cache may clear: in some environments that takes a few minutes, in others up to an hour. The model then reads the whole history again at full price. OPM moves the context into project files, so it is easier to start a new session, bring in more than one model, and keep planning separate from execution.
Four supports of OPM
- I
Authority stays with the human
The human is the only source of decisions in the project. In OPM this role is called the Operchief: the project's chief operating officer (COO). The agent analyzes, plans, proposes, and carries out what the Operchief has approved. The agent has no authority of its own.
- II
The PIN separates words from action
The agent works in two modes, and the line between them is absolute. The guard on that line is the PIN: a code the Operchief types in person. Discussion is the default. Reading, analysis, planning, a look at the risks, questions, and clarifications are allowed. Changing any file is not. Execution starts only after the PIN, and only inside the plan that was approved.
- III
Logs are the project's external memory
Decisions, plans, and the history of the work are kept in project documents and logs. A new chat restores the context from those files, so a person does not have to explain the point of the project, or the task in hand, to every new agent.
- IV
StepLoop
Any significant piece of work with an agent in the project goes through a cycle: state the task, look at it, discuss the plan, write the plan into the log, the agent carries the plan out, report, check the result, and either go back to planning or accept the result and archive the records.
The method calls this cycle a StepLoop. The Operchief is in it at every stage. Following the cycle is what keeps the project under control.
The Operchief takes different parts in that cycle: subject specialist, coordinator between agents, keeper of the protocol, the person who presses the buttons, or the one who cooks the Sunday roast.
When the Operchief cannot judge the result, or does not want to — the code is unfamiliar, the language of the text is not one they know, or there is no time — another agent can be brought into the cycle to check the work. The Operchief then simply accepts it by giving the PIN.
Document kit
The method can be used in two ways.
On your own computer, with an agent through any harness that runs there and can see the project files. Claude Desktop, Codex, Cursor, DeepSeek Harness, and other programs of that kind work this way.
In a browser chat: ChatGPT, Claude, Gemini, DeepSeek, and others.
Installing the method.
For work on your own computer:
Download OPM_SETUP.zip, put the archive in the project folder, and ask the agent to start installing OPM in the project. The agent will do the work and ask you about the project as it goes. If something is unclear, ask the agent: it will explain.
The project folder will fill up with directive files and folders, so learn the method gradually, by asking the agent. Every file is ordinary text in Markdown.
A project can have as many harnesses and agents as you need. Different pieces of work can be given to different agents through the exchange protocol built into the method. That is how the strengths of different agents, and paid and free plans, can be combined.
For a browser chat:
Download OPM_CHAT_EN.zip and upload it to the chat, asking the agent to set the method up in the chat. The agent will unpack the archive and set the log template up to suit you.
A working chat uses two files. OPM_CHAT_DIRECTIVES.md tells the agent how to work with you. It does not change from chat to chat. OPM_CHAT_LOG.md is the log. It grows and changes in each chat. The log is the project's external memory: a task can be carried into another chat by means of it.
To move to a new chat, upload the directives, the log, and the working files. At the end of the chat, ask the agent for the new edition of the log and for the working files.
File kit
| No. | Document | Role |
|---|---|---|
| 01 | START_HERE.md | how to begin; leaves the project root after setup |
| 02 | AGENTS.md | the constitution; the agent does not edit it |
| 03 | CLAUDE.md | the entry point for Claude Code; the agent does not edit it |
| 04 | agent brief and project brief | settings, the document map, the goal, and acceptance |
| 05 | logs | the working log, the reserve log, and the archive of closed cycles |
| 06 | directives | passing assignments, roles, and checking |
| 07 | PROJECT_EXAMPLES | optional practices, not rules |
Chat kit
| No. | Document | Role |
|---|---|---|
| 01 | OPM_CHAT_DIRECTIVES.md | tells the agent how to work with you |
| 02 | OPM_CHAT_LOG_TEMPLATE.md | a log template, not a finished log |
The file kit is in English. Its language of communication is chosen at the start and written into the brief. The chat kit's language of communication is English.