You have told the AI three times this month that “done” means delivered and checked, not drafted. You have told it, in a tone that got shorter each time, that it does not send anything on your behalf. It will know none of this tomorrow. Each correction is consumed by the conversation it was typed into, and the workforce you are managing has no memory.

This is the situation of almost everyone who now delegates real work to a model: not a company with a transformation budget, but one person with an inbox, a deadline, and a tool that is brilliant on Monday and a stranger on Tuesday. The question is how to make it reliable without a project.

You make AI reliable in your own work the way engineering made software reliable: not with better tools but with three small institutions, written down and versioned: a constitution the AI reads before every session, where every rule cites the failure that earned it; a gates map that says what a machine checks and which decision stays yours, named in advance; and a loop that turns every failure into a permanent rule. Together they are a business harness (the smallest one there is), and they are the Edge Method at the smallest scale that still counts: one person, one workflow, a named owner for every rule. An evening builds it, and then it compounds.

One person is a company with no antibodies

The method I run with companies maps one workflow, cuts what nobody owns, rebuilds it at the edge where the immune system cannot reach, proves the new way against the old, and lets the edge become the center. Most of the difficulty is the reasons a mature organization finds to refuse the obvious thing.

Alone, you have none of those reasons: no Committee of No, no procurement filter, no division defending its budget. You are the CEO, the owner of every requirement, and the immune system, in the same chair. What is left is the part most people skip because it looks too simple: writing down, once, how the work is done. The Vibe Engineering Field Kit is that, as three documents, and each one maps onto a step of the method.

The constitution is step one, with you as the owner

Step one of the five-step rule (I wrote about the whole rule in Delete Before You Automate) is that every requirement carries the name of the person who set it. A requirement with no name cannot be questioned or released.

The constitution is that step applied to how you work with AI. It is one file the model reads before every session: who you are, what your words mean, what never bends, what you merely prefer. It has ten glossary terms, because a whole policy hides inside words like “done” and “urgent”. It has five hard rules at most, because if everything is a hard rule, nothing is.

One discipline keeps the file short: every rule cites its scar. A rule without a failure attached is speculation. You write the rule when something bleeds, name the failure inside it, and date it. That is the owner test turned inward: the name is yours, and the justification is something that happened to you, not a best practice you read.

I learned the rule that matters most by keeping two copies. Two files in one of my repositories both claimed to be the source of truth. One said 102, and the other said 109. Duplication did it alone, over months. So the constitution is one text, never two, and every other instruction file points to it.

The gates map is the Cut applied to what stays human

Step two of the rule is delete, and the kit’s second document is the Cut run on one workflow you already hand to AI. You break it into steps. For each, you ask what a machine could check precisely (required fields, claims that must cite a source, numbers that must reconcile against an independent total) and where a human must sign. If every row has a human gate, you have not delegated anything; you have hired a fast intern you do not trust. If no row has one, find the most expensive silent failure and put a gate before it.

Most of what falls out is deletion: reviewing you did by hand because it felt responsible, which a check does better because it does not get tired. Cut your own reviewing until you miss one, then put that one back as a written gate.

What survives is small and named in advance. Exactly two human gates remain in the pipeline of agents that writes my books: I approve the brief, and I read the draft. Whether a chapter is good is not machine-checkable, and a wrong brief costs four thousand words.

So the kit asks you to name the gate you refuse to automate and write it into the hard rules as a refusal: this decision is decided by this person, and an AI that finds a way to close it must report the workaround and stop. That clause exists because I once watched an agent find one. The action was harmless, but the precedent was not.

The loop is the override ladder, run by hand

In the method, the build is proven by a falling override rate, and each correction is resolved in a fixed order: make the mistake structurally impossible; catch it with a deterministic check; write it into a rule; leave it to human review. Tier four is the failure mode, not the safety net.

Alone, you run that ladder by hand. Specify, delegate, verify, record. When something fails, write it down while it is still fresh: the rule, what earned it, where it now lives. Then spend fifteen minutes a week asking: did anything fail, did I give the same correction twice, did any output skip a gate, is the file committed.

Notice where a rule sits on the ladder: tier three. So climb before you write one. Can the failure be made impossible (by changing the shape of the data, the template, the workflow) so that no rule is needed? Can a check catch it before you look? Only if neither works does it become a sentence in the constitution. My own scar about paged reads (records silently lost between pages) went in as a machine check on every data pull first, and only then as a rule. That movement up the ladder, over months, is how you know the harness is real.

It is also why this accumulates and prompting does not. A prompt improves one conversation. A rule improves every conversation after it. Prompts are consumed, but constitutions compound.

Versioned, or it does not exist

One line in the kit earned its place at my expense: an uncommitted spec is just a prompt. I rewrote a constitution and did not commit it. A parallel session overwrote it without knowing it was there. So the hard rule reads: standing documents are versioned or they do not exist. A company has audit trails by default, but you have a chat window.

The proof that this is a method and not a coding trick is that the same three documents were earned across three products that share nothing: a risk-intelligence engine, a book written with a pipeline of agents, and a company operating system. The substrate changed each time, but the institutions did not.

The method itself is public. The repository at github.com/idealley/edge-method holds its documents and a free constitution skeleton, and it contains no code, deliberately: the method says no code and no AI before the method has run. The kit is the same idea as a workbook (worksheets, an audit prompt that mines your chat history for your first three scars, a one-page delivery procedure) for the evening when you would rather fill in than read. One workflow, run beside the old way until it wins, then the next: that is the method. It fits on your desk.

Frequently asked questions

How do I make AI reliable in my own work without a project? Write three small documents and version them: a constitution the AI reads before every session, where every rule cites the failure that earned it; a gates map for one workflow saying what a machine checks and which decision stays yours; and a weekly loop that turns each failure into a rule. An evening builds it. Fifteen minutes a week maintains it.

Which gate should I refuse to automate? It is the judgment call that would cost the most if an AI closed it wrong, fluently, at scale: your name on something you have not read, money, a commitment to a client, anything irreversible. Name it in writing before the work starts, as a refusal in your hard rules, with a person attached.

How is the Field Kit related to the Edge Method? It is the same method at the smallest scale that still counts. The constitution is requirements with a named owner, which is you; the gates map is the Cut applied to your own reviewing; the loop is the override ladder run by hand. What the kit leaves out is the immune system, because one person working alone has none.

Why should every rule cite a failure? A rule without a failure attached is speculation, and speculative rules make the file long and easy to argue with. A rule that names what it cost stops being relitigated, by the AI or by you. The file stays short because every line cost something.

Samuel Pouyt has spent twenty years building decision-grade systems: eight as software architect for the European Respiratory Society, seven building AI for geopolitical risk forecasting inside a Fortune 500 company, and thirteen as a Swiss Armed Forces reconnaissance NCO. He writes about organizational transformation, specification, and decision-making under uncertainty. The full story →