EgD-SAPF-001 r2 · Start here

Make the AI work accumulate

You already use Claude, Copilot or Perplexity on your projects. This is how to point them at a GitHub repository so the work lands somewhere, the decisions are recorded, and the spend shows a return.

Written by a functional consultant, for functional consultants · 2026-09-18

Start the tutorial Set a usage standard

The operator on one side of the repository, the objective on the other. The triangle in between binds every token of the model's processing to the axis between them — EVE Enterprise Harness at entry, canon at exit, observation registry logging every attempt to leave that axis.

The three edges

1.EVE Enterprise Harness

How the AI surface behaves before it does anything. The read-first rule. The one edge every session starts on.

2.Canon

How the artifact leaves the repository. The output shape. The edge the work travels along on the way out.

3.Observation registry

How the truth of what happened is preserved. The record of what was done wrong so it can be done right. The edge that closes the triangle.

Take the EVE Enterprise Harness away and the surface drifts. Take the canon away and the artifact leaks. Take the observation registry away and the other two become preferences instead of rules.

Sovereign Start

The problem, in one paragraph

You ask Claude to work through a config decision. You get a good answer. You close the tab. Next week someone asks why the decision went that way, and the reasoning is gone — it was in a chat window nobody kept. So you ask again, and pay again, and get an answer that is slightly different from the first one. Multiply that across a project team and the tools have cost real money without leaving anything behind.

Developers solved this a long time ago. They put the work in a repository. The repository is the control unit — it holds the decision, who made it, when, and what it was based on. That is all this site is suggesting: that the functional side of the project use the same control unit the technical side already uses, and point the AI tools at it.

Inner — the record

Your repository. The decisions, the analysis, the working papers. Yours, readable, in plain files.

Mid — you

Where you and your AI tools work. Reads from the system, writes to the repository. Never writes back into the system.

Outer — the system

SAP, or whatever the client runs. Untouched. This layer sits on top and changes nothing underneath.

Set up the repo

One repository per workstream, with a shape the whole team recognises. Twenty minutes.

Add the harness

One file of instructions your tools read automatically, so every session starts warm.

Set a usage standard

Agree what the team spends on, and what a result looks like. Otherwise the bill grows without the output.

What a harness is

A harness is a small set of written instructions that your AI tools read before they do anything — the same instructions, every time, whichever tool you are using.

Without one, every session starts cold. The tool does not know your project, your client's naming, what was already decided, or where to put the output. So it guesses, or it searches the internet for something already sitting in your repository. That is where most of the wasted spend goes — not on hard questions, but on re-deriving things you already had.

With one, the tool opens the repository first, reads what has already been agreed, and writes its output back as a file rather than a message in a window. The kit is small — one file you edit, and the tool-specific files are generated from it so Claude, Copilot and Perplexity cannot end up following different rules.

See how it is set up

The point is the decisions

On a programme, the expensive thing is rarely the document. It is the decision the document was supposed to support — and the fact that six months later nobody can reconstruct why it went the way it did.

So the thing worth capturing is not the transcript. It is the decision: what was chosen, what the alternatives were, what it rested on, and who agreed. That is a short file. It costs almost nothing to write and it is the only part anyone asks for later.

If the session were deleted right now, could someone rebuild the reasoning from the repository alone? If not, the work is not finished.

Who this is for

Functional and business consultants. Process leads, analysts, PMO, solution architects on the functional side — people who are accountable for decisions rather than for code, and who are now using AI tools daily without a shared way of organising the output.

You do not need to be a developer. The tutorial assumes you have never used GitHub, and the parts that are genuinely technical are done for you by a script.

What this is not

Not a methodology. If your firm has one, follow it. This is a place to put work, not a way of doing it.
Not a product. Nothing to buy, nothing to sign, no partnership tier. GitHub's free plan is enough.
Not a replacement for anything. It sits on top of what the project already runs and changes nothing underneath.
Not client material. Nothing from a client engagement is published here, ever.

Join in, or just ask

This is one person's learning, shared so other people can do the same thing and improve on it. There is no list to join and no account to create beyond the GitHub one you will need anyway.