I asked an agent the same question three times last month: what's our current position on pricing? Not because the answer kept changing. Because I couldn't remember where I'd put it.
I'd written it down, a good version of it, with the logic and the exceptions and the two accounts where we'd broken the rule on purpose. It existed somewhere in a repo, or a doc, or a Slack thread I'd have to go dig for, and it was faster to re-explain the whole thing to the machine from scratch than to go find what I already knew.
That's the low, dumb frustration most operators are living with right now. You believe context matters, and you've read that the model is only as good as what you feed it. And you're still typing the same paragraphs into a chat window every session, re-briefing an AI on your ICP, your pricing logic, your last three decisions, like it's the first time every time.
The context isn't missing so much as scattered and single-use, and you're the only integration layer holding it together.
But you are not starting from zero. The thing you keep re-typing is proof the material already exists. You don't need to author a context layer from nothing. You need to structure the vital slice of the work you're already producing into a set of canonical sources of truth, the documents you and your agents both run against, and structure them so that you can actually read, check, and stand behind what's in them.
Pick the handful of docs that carry the load
Most people hear "structure your context" and picture documenting their entire operation, so they never begin. Don't do that. Most of your artifacts answer almost none of the questions you actually get asked.
A small fraction of them carry almost all the weight: the current ICP, the top five objections and their rebuttals, the pricing logic, the handful of positioning claims you repeat in every deal. Those are your canonical sources of truth, the documents you re-run against, the ones an agent should draft from instead of guessing. Pick those first. Leave the long tail rough.
The concrete first move takes ten minutes. Write down the five things you re-explain to the AI, or to a new hire, most often. That list is your first set of canonical docs, and you already know what's load-bearing, because you say it constantly. The raw stuff underneath, the transcripts and the old decks, can stay messy while the quality effort goes into the few synthesis pieces on top.
Where the material already lives
The context you keep re-explaining is sitting in six systems that don't talk to each other, or to your AI.
- Your CRM. HubSpot, Salesforce: the accounts, the deal history, the closed-won and closed-lost notes, the ICP as it actually behaves rather than as your deck describes it.
- Notion, Drive, the wiki. The positioning docs, the battlecards, the strategy memos you already wrote and haven't opened since.
- Sales decks and one-pagers. Your single best articulation of why you win, trapped in slides no agent can read and no rep can find at 4pm on a Thursday.
- Call recordings. Gong, Fireflies, the Zoom cloud. The highest-signal record of customer language and objections you own, and almost none of it is structured into anything.
- Your own website and published content. Your live positioning, already public, already yours.
- Email and Slack. Where the actual decisions got made and then never got written down anywhere durable.
Your team shouldn't have to read 47 call transcripts to find the pain points. They should read the one synthesis doc that already pulled the pattern out of all 47.
The raw material is the 47 transcripts, and the work is the synthesis. You've been skipping the synthesis because it's nobody's job and it never feels urgent.
Some of it you have. Some you send an agent to get.
There's a second source most people miss, and it's newer. Not everything in your context layer has to be something you already produced. Some of it is context you can now send an agent to go get.
Point an agent at your three main competitors: their site, their pricing page, their positioning, the way they talk about the category. Have it read all of it and write up where they're strong, where they're soft, and where you actually differentiate. That competitive read used to be a research project you'd do once and lose.
Now the output is a durable battlecard your team reads and your agent pulls from the next time it drafts an objection-handling email. You did it once, and it pays out every time after.
The second intake mode is pointing an agent at a transcript or call recording, having it draft the synthesis, and keeping the output instead of letting it die in a thread. You get the piece you never had time to write, and the agent has a canonical source the next time it needs to draft from that content.
What a canonical source of truth actually looks like
Here's one. This is an abstracted version of a real ICP doc, the kind you'd keep in a foundation folder and run your account scoring against.
---
description: Who we sell to, the scoring criteria, and the disqualifiers
type: synthesis
status: active
owner: Priya (RevOps)
updated: 2026-07-01
last_reviewed: 2026-07-01
tags: [icp, targeting, foundation]
related_docs: [[win-loss-analysis]], [[persona-map]]
---
# Ideal Customer Profile
## The Answer (read this first)
| Segment | Fit | Why |
|---------|-----|-----|
| Mid-market B2B SaaS, 200–800 employees | STRONG | 7 of our last 10 wins [VERIFIED: closed-won report, 2026-Q2] |
| Series B–C, post-first-RevOps-hire | STRONG | They have the pain and the budget [VERIFIED] |
| Enterprise (2,000+) | WEAK | Long cycles, we lose to incumbents [VERIFIED: 4 of 5 losses] |
| Pre-seed / seed | DISQUALIFY | No budget, churns in 90 days [VERIFIED] |
## The Evidence (check it here)
- **Champion is a practitioner, not an exec** — 7 of 10 wins [VERIFIED: win-loss, [[win-loss-analysis]]]
- **They already tried a spreadsheet and it broke** — recurring in discovery [SINGLE-SOURCE: 3 recent calls]
- **Renewal risk when the champion leaves** — [HYPOTHESIS: 2 observed, needs more data]
None of this is clever, and it isn't supposed to be. Walk the parts.
The frontmatter is the accountability layer. owner names the human on the hook for keeping it true, Priya, not "the team" and not "the agent." updated and last_reviewed tell you when it last actually moved versus when someone last looked at it and confirmed it still holds. status: active means it's live and things run against it. A doc with no owner is a doc nobody maintains, and in operations, no one is above owning the data.
The answer sits on top, the evidence sits underneath. A person opening this reads the table, gets the current position in fifteen seconds, and closes it. That's the job most of the time. The evidence is there for the moment you need to check a claim or an agent produces something that smells wrong and you want to trace it. Synthesis first, detail on demand. You don't read the whole thing to use it.
Every claim carries where it came from. [VERIFIED: closed-won report] is a fact you can trace. [SINGLE-SOURCE] is one data point, believe it lightly. [HYPOTHESIS] is a guess you're tracking, not a truth you're asserting. This is how you check a source of truth without trusting it blind. You can see, per line, how much weight it holds and where to go if you doubt it. A claim with no tag is a claim nobody sourced.
The wiki-links connect it to what backs it. [[win-loss-analysis]] is the detail doc the strong-fit call rests on. Follow it when you need the receipts. The related_docs in the frontmatter do the same job for navigation. A human walks these links to audit; an agent walks the same links to ground its output in the same evidence.
Now notice what you can do with this doc that you can't do with a black box. You can open it and say yes, that's still what we believe, or no, enterprise moved to PARITY last quarter and this is stale. You can find the one claim an agent got wrong and see it was a [HYPOTHESIS] you never should have let it treat as fact.
You built it for yourself to read. That's what lets you own it. The agent runs against it fine either way.
You still own the truth, so you have to be able to read it
That last point is the whole reason the structure looks the way it does. The point of a canonical source of truth is that your agents stop guessing and start running against a settled answer. But you are still the one who settled it, and you are still the one on the hook when it's wrong. An agent that drafts outbound off your ICP doc is only as trustworthy as your ability to open that doc, read it, and say yes, that's still what we believe.
So the constraint that actually matters isn't "make it machine-readable." Agents are forgiving readers. They'll parse almost anything. The constraint is that a human has to be able to comprehend, check, and stand behind every canonical doc, because a human keeps final oversight of what counts as true. The moment your source of truth becomes a black box, a thing the agent uses and you can't quite explain, you've lost the one thing that made it worth trusting.
That happens more easily than you'd think. You let an agent maintain a synthesis, it accretes, and six months later there's a "current ICP" doc that drives your scoring and your outbound and your new-hire onboarding, and nobody in the building can tell you why it says what it says or when it last changed. It works until it's confidently wrong and you have no way to see it, because you built a thing you were never able to read.
A source of truth you can scan in a minute, whose claims you can trace to where they came from, whose last real change you can find, is a source of truth you can actually own. One you can only feed to a model and hope about is a risk you haven't priced yet.
You can run a test on your own layer Monday. Take one real question your org runs on, "what's our position on pricing," "what's our rebuttal to build-vs-buy," and open the canonical doc that answers it. Can you, a human, read it top to bottom in a minute and say yes, I stand behind this? Can you see where each claim came from and when it last moved?
If the honest answer is "the agent keeps this current and I mostly trust it," then what you have is a black box you've decided to believe, and you should stop calling it a source of truth.
Run your agents against it
Once the doc exists, the agent stops guessing and runs against the settled answer. It pulls the ICP, produces something slightly off, and because you can trace the output back to the doc it ran against, you catch it and sharpen the source. A layer you can actually read is a layer that gets used, which means it gets corrected. Someone opens the objection doc, sees it's stale, and fixes it.
Every use is a chance for a correction, and the corrections accumulate. That's the compounding. Each round, the layer gets truer.
When I watched our own system take hold, what struck me was not the token savings but that the org got smarter without anyone working harder, because the system captured what used to live in people's heads, structured it, and redistributed it automatically. This is infrastructure, not automation. Most orgs I've watched reset every quarter. This is how one stops.
The concrete proof is boring and specific. New-rep onboarding that used to take 90 days of tribal knowledge transfer takes 30, because the actual customer quotes and the decisions are sitting in a canonical doc a person can read, not locked in someone's head. The same doc the new hire reads to get ramped is the doc the agent runs against, and because a human curated it and can vouch for it, both of them are working from something true.
I'll give you the honest caveat, because I don't fully know the answer. I don't know if this advantage is sustainable long-term. Maybe the tools commoditize and everyone builds these layers and it becomes table stakes. But right now, the operators structuring what they already have are pulling away from the ones still re-typing it into a chat window every morning.
The moat is yours to build
So the moat was never the folder structure. A competitor can clone your repo, copy your directory tree, even lift your captured files, in an afternoon. What they can't clone is a living layer that both your humans and your agents run on and correct every day. It holds the accumulated judgment of everything you've decided is true.
You already have the raw material. It's in the CRM, the transcripts, the decks, the threads. The task is picking the vital handful and building each one so a person can read it and stand behind it, while letting capture ride on the work you're already doing.
Start with the five things you re-explain most. That's your first set of canonical docs. Open each one, ask whether you can actually read it and stand behind it, and fix the ones you can't.
The material is already yours. You've never made it usable once, in a form anything can run against without you in the loop. Until you do, you're the integration layer, holding it all together in your head. So start Monday, with one doc you can read.
Related reading: Once your layer exists, Three Mechanisms That Keep a Context Layer Self-Healing covers how to keep it current without a separate maintenance job. And The Four Ways Your AI's Content Goes Stale maps the decay vectors so you can spot them before they cost you.
Victor Sowers builds AI-native GTM systems at STEEPWORKS. 15 years scaling B2B SaaS, two exits, and 2.5 years of production AI-in-GTM.
