The Context Audit: Five Steps to Find What You Still Know

By Ryan Vanshur

Picture a renewal brief that reads beautifully, names the competitor, quotes the current price and leads with a differentiator the team is proud of. The rep opens with it on the call, and the customer asks whether anyone has been paying attention, because the competitor was bought last spring, the price moved in the summer, and the differentiator got dropped before the quarter even began. The model did what it was asked. It worked from a folder of documents that had stopped being true, and it had no way to know.

Most teams read that moment as a model problem and go shopping for a better model, and usually that's the wrong fix. What a team needs is a context audit, which is an inventory of what it knows, a check of what is still true, and a folder of text files the AI can read. We walk through the five steps here one at a time. The running example is PlotIQ, a fictitious deathcare software company, so every step has a concrete subject to test against.

Why the Audit Comes Before the Build

Many teams start by building a knowledge graph or wiring a tool into the CRM. Both are reasonable, but both inherit whatever sits underneath them. A graph built on a stale pile just repeats the stale parts faster, so we audit first. The audit needs no engineering, only an afternoon, a shared folder, and one rule for what counts as true.

The frustration of a model that keeps getting the product name wrong is useful, though, because it's telling you something about the pile underneath it, and listening costs nothing. The pile is almost always the problem, and it's far cheaper to fix before anything gets built on top of it than after.

Step One: Run a Content Audit

We start by collecting what each team actually has, and we don't filter yet, so product docs, enablement material, call transcripts, decks, and the one-pager the sales engineer keeps on the desktop that nobody else has seen all go in the pile. The goal is the full inventory, and the full inventory is always bigger than anyone expects.

For PlotIQ, that meant pulling material from sales, customer success, product, and marketing into one place before anyone decided what to keep.

Step Two: Burn What Is No Longer True

Now go through the inventory and ask one question of each item, which is whether it's true today, not when it was written and not after the next release. Then mark what passes and let the rest go. This is the step most teams skip, and it decides whether anything built on top of the material is worth reading at all.

Check each item against four common failures:

• Products that have been renamed
• Pricing that has moved
• Competitors you no longer meet
• Processes from before the last reorganization

In the audit behind the essay, roughly three quarters of the pile failed that check, and your ratio will differ, but expect a large share to go. Stale material is worse than no material, because people trust it and don't go check.

The instinct is to keep everything, because throwing work away feels like waste and a big pile feels like an asset. That instinct runs backwards. Anything built on top of the material inherits its accuracy, so a stale note never sits harmlessly in a folder, and instead it gets retrieved, cited, and turned into a confident sentence in a renewal brief. If a note's status is unclear, treat the doubt as the answer and move it to the burn pile until someone confirms it.

Step Three: Move What Survives Into Markdown Folders

Put what survives into plain text files, organized into folders the way the team thinks about the business. Accounts, competitors, plays, personas, and pricing each get a home. Every model can read a text file, every person can edit one, and no platform decision stands between the team and its first session.

When the folders are ready to connect to a tool, the Context OS Setup skill covers the setup. Steps one, two, and four need no tooling at all, which is the point of doing them first.

Step Four: Extract the Tribal Knowledge

The richest material isn't in any document, because it lives with the people who have been in the job longest, and it has never been written down anywhere. We sit down with them and ask three questions:

• Why do we win the deals we win?
• Which objection actually kills us?
• What does the buyer say in the second meeting that tells them it's real?

Write each answer down as a short node in a folder, with the person's own words where possible, and note who said it and when, because the date tells you when to check the answer again. Until it's captured, nobody who talks to the AI gets the benefit of it.

Step Five: Set the Maintenance Cadence

Ingest new material as it arrives, run a short health pass every week, and do a deeper review once a month, and then write the cadence down and put one named person on it. A cadence with no owner is a plan that lasts only until the next reorganization, and by then the folder has aged a season.

Steps one through four are a project with an end date, but step five is a habit, and it's the one most teams underestimate when they plan the work.

Some of the maintenance can be automated, so new material gets ingested at the end of a working session without anyone touching it. The rest is a person on a Friday afternoon cleaning up nodes and running a longer pass once a month. The folder will never be perfectly accurate, and perfect is the wrong target. It only needs to be accurate enough that the people who check it find it right, because that's the bar credibility gets built on.

Data has a shelf life, and the failure doesn't show up in week one but in month six, when the tool everyone loves is quietly running on a snapshot of the business from March. The cadence keeps the folder on the right side of that line, and the first pass is quick, but keeping it true is the longer job, so plan for both.

What the Audit Buys

Three things change once the folder is clean, and they tend to show up in this order.

1. The repeated prompts go away. A grounded request returns a usable first draft often enough that people stop iterating and start checking.
2. New hires and veterans read from the same ground truth. The rep in her first month no longer guesses which objection matters, because the person who has been there longest already wrote it down.
3. The result shows up in the numbers a board cares about. Watch cycle time, win rate, and deal size, since an adoption count on its own tells you very little about whether the audit paid off.

Account intelligence is the easiest place to check whether the audit worked. Ask a grounded system who the team is up against at a specific account and what it wins on, and the answer can be checked in about thirty seconds against what the team already knows. When it holds up over and over, people stop double-checking, and that's the moment the tool starts to earn its place on the team.

Go Deeper

• The Context Audit, the essay this guide grew out of.
• Context OS Setup skill, the tooling for step three.
• Context Gap skill, a check for what already exists before you build.
• Your AI Doesn't Remember Last Quarter, a related read on the same problem.
• GTM Knowledge Base Maintenance, the step-five companion guide, publishing alongside this one.

On Monday, the Brief Reads Right

Picture PlotIQ on a Monday, three months into the audit, and a sales engineer opening the account folder before a renewal call. The competitor note carries a date from this quarter, the pricing file says it was checked on Friday, and the objection list has a line from the longest-tenured rep about the second meeting, in that rep's own words. The sales engineer asks the model for the renewal position, and the draft names the current competitor, quotes the current tier, and raises the objection the buyer raised last time. She checks each line against the folder in about two minutes and sends it.

The model didn't get smarter that week, and nobody bought a new one. Someone simply checked the folder on Friday, and that small habit is the whole difference between the two briefs.

1. Join the Guild and get the weekly operator brief. Subscribe free.
2. Run step one this week. Block an afternoon and start with one team's folder.