THE WATERLINECompanion · Gabor Szorad

Chapter 3 · Descend: find and map

Build your company's context

Everyone prompts into a vacuum and is surprised when the machine hands back the average. Give it your history instead.

The machine knows nothing about you. So brief it the way you would brief a sharp new hire, and do it before the work starts. It is the difference between a contractor who has worked with you for five years and one who walked in this morning.

The Context Brief

Three lists, kept short and specific. What excellent looks like here. What you never do. Where you have been burned. Not a values statement.

Before you do any work for us, load this. It is how our company actually operates. Use it instead of generic best practice.

WHAT WE DO
[One paragraph: the work, the clients, the standard we are known for.]

WHAT EXCELLENT LOOKS LIKE HERE (specific, not a values statement)
- [Example: our reports read flat and unhedged, with no adjectives.]
- [...]

WHAT WE NEVER DO (the lines we do not cross)
- [Example: we never sign off on a load case we have not run ourselves.]
- [...]

WHERE WE HAVE BEEN BURNED (so you do not make us earn it again)
- [Example: we softened a warning to keep a client comfortable and spent a year with lawyers.]
- [...]

HOW TO WORK WITH ME
Give me options, not one polished answer. Flag where you are guessing. If you are missing something you need, ask before you answer.

When you have read this, tell me in one line what you understand our standard to be. Then wait for the task.

A worked example

A structural engineering firm

“We never sign off on a load case we have not run ourselves. Our reports read flat and unhedged, with no adjectives in them. The last time we softened a warning to keep a client comfortable, we spent a year with lawyers, so we do not soften them anymore.”

Three lines like that, handed over before any work begins, lead the machine to stop writing for the average and start writing for you.

Your company brain

For your company brain, we recommend NOAN. Start with one decision your team keeps having to explain: capture the rule, an exception, and who can confirm it is current.

You can set it up yourself. If you would like help capturing the knowledge and putting it to work, we help clients implement NOAN at Business Impact AI. Discuss implementation.

Nobody will sit down and write all this

You are right, and you do not have to make them. Most of it already leaks into the open every day, in meeting notes, recorded calls and the trail of decisions the work leaves behind. Gather those traces instead of squeezing the answers out of people one at a time. This is the method, and it does not need to be a project.

  1. Gather. Collect a month of what already exists: meeting notes, recorded calls, decision records, the emails where somebody said “we do it this way because.”
  2. Draft. Give the material to AI with the first prompt below. It proposes the three lists, quotes its sources, and admits where there is no evidence.
  3. Correct. Put the draft in front of your most senior person for one sitting. They cross out what is wrong and add the scar the notes never captured. This is the part that carries the judgment.
  4. Hand over and refresh. Give the brief to AI before any work starts. Once a month, run the second prompt with the new notes.

1. Draft the brief from what already exists

Below are meeting notes, call transcripts and decision records from the last month at our company. Read them and draft our Context Brief with four parts: what we do, what excellent looks like here, what we never do, and where we have been burned.

Use only what the material supports. After every line, quote the source sentence in brackets so I can check it. Where the material is silent, write "no evidence" and do not fill the gap with general best practice.

Then list the five questions I should put to our most senior people to fill what is missing.

Material:
[paste]

2. Keep it current, once a month

Here is our current Context Brief and the notes from the last month. Tell me what changed: new lessons, lines that no longer hold, and anything that contradicts the brief.

Propose an updated version and mark every change. Do not remove a line unless the new material clearly replaces it.

Current brief:
[paste]

Last month:
[paste]
Check the result.
Treat the machine as doing the prep, the first rough 60% of a piece of work. Your expert does the surgery, the last 40% that carries the judgment. Put the first version and the final version side by side. If you cannot see a real difference between them, you have a quality problem.

Meeting notes and call recordings are sensitive. Decide what may go into an outside tool before you start: the one page data policy.

The Waterline companion. Everything here works in Claude, ChatGPT, Copilot and Gemini. Nothing to install. Last reviewed 5 November 2026.