The Eleven Comms Docs Your AI Tools Need to Actually Work
Your AI stack is fine. The main problem is what you're feeding it.
I have a weird firm. One half is the standard type of strategic communications advisory work you’d expect from someone with two decades in the field. The other half is what I call context engineering for communications: building bespoke AI systems, scaffolding, and tooling to help comms teams do more and focus on stuff that matters.1
One thing I’ve observed across all my context engineering clients: everyone wants to go right to building the tools. Tools are fun. Tools are satisfying. No kid at Home Depot gravitates to the drywall aisle. They go straight for the power drills.
You know what doesn’t give that same rush? Document creation and curation. At best, it’s a clerical process that takes some hunting around cloud drives, pinging of colleagues, and version checking. At worst, it’s a painful exercise akin to going under your house with a flashlight and finding several load-bearing supports are rotting out.2
And like those supports, a weak document foundation for your AI systems eventually brings down everything you build on top of it. That’s why I always start here with clients, and why I’m going to give you a bit of a peek into my process.
Four parts of one core
Every piece of communication your team produces answers four questions:
What are we saying?
Who is this for?
How do we sound?
Who has to clear it?
These four questions form the basis for the four types of core documents I have clients develop: Story, Audience, Voice, and Gates.
I call the assembled set the reference core: the durable material both your team and your tools read from. Which layer a document belongs to determines how you write it, who owns it, and what happens when it contradicts another one.
Across the engagements I’ve run, most teams have partial coverage of one layer and nothing in the other three. The institutional knowledge that would fill them exists, but it’s distributed across six people’s heads and a Google Drive that looks like a bomb went off inside it.
Story: what we say
The corporate narrative. In prose, not a deck or bullets. It should have your positioning, three or four key messages with proof points for each, the category you’re claiming, what actually differentiates you, and what this quarter is about. Written in paragraphs so the connective logic between ideas is more obvious. A deck of bullets gives you the framework and throws away the substance.
The fact sheet. Every approved fact, superlative, or claim you use (headcount, funding, valuation, growth, volume, customer count, etc.) carrying three attachments: a disclosure status of public, confidential, or embargoed until a specific date; a link to the evidence; and the date it was last verified. This one doc powers drafting, fact-checking, exec prep, and legal review simultaneously.
Product and service one-pagers. One per offering. What it is, who it’s for, the proof, and the approved language for describing it. The half that carries the weight is the approved language, written as the exact string you want used rather than a description of the vibe. Boring to write, critical to consistency.
Audience: who it’s for
Audience personas. A profile for each external audience that matters. A real profile carries the depth of a full buyer profile: where they exist, what they care about, what language lands and what makes them tune out, where they actually consume information, their media diet, who in their lives they trust, the mental model they arrive with, and the objections they bring. Make it both demographic and psychographic as much as possible.
Employee persona(s). Your most consequential audience and the one most likely to have no document at all. What reassures them, what spooks them, how they read an org-change announcement differently than the press does. If you have radically different employee constituencies within your organization like corporate and store-based retail, create multiple personas. Add your board as its own profile when board communication is a real part of the job.
Optional: profiles for your priority reporters, treated as an audience. Their beat, what they’re actually after, what they’ve already covered. Keep this separate from your media CRM, which tracks live state and belongs outside the core.
Voice: how we sound
The brand voice guide. The house voice, which belongs to the company rather than to any individual, and shouldn’t be built by averaging your executives together.
The editorial spec. House style, AP plus your exceptions, product naming and capitalization, terms you use and terms you’ve retired, format conventions. If there’s a mandatory legal disclaimer living in your style guide right now, pull it out. That’s not style. That’s a gate, and it belongs in a different layer for reasons I’ll get to.
A profile for each principal. Two halves in one document, and most teams only have the first. That half is how they sound: register and cadence across different audiences and media, the turns of phrase they consistently return to, the words they use and the words they’d never use. The second half is all about receiving information: what format they want it in, how much lead time they need, what they always hate seeing in a doc. Lastly, the part everyone skips, which is where what they say they like diverges from what they actually approve.
Gates: who can stop it
A gate is a constraint you clear past. Miss on voice and the work is weaker; miss on a gate and it’s unshippable. Voice is owned by comms and measured on a spectrum. A gate is owned by someone holding actual veto (legal, IR, your head of comms) and it’s measured pass or fail.
Red lines. What you cannot say, legally or strategically. Banned phrasings, topics that route through legal before they go anywhere, embargo and MNPI rules, your policy on naming or not naming competitors. The medical device company that can’t say “cures.” The private company that can’t/won’t confirm a valuation. Every organization has a version of this list, and most have never written it down, which means it exists as the institutional memory of whoever has been there longest.
The legal and compliance gate. The predictable flags, plus any mandatory language (e.g., the disclaimers you evicted from the style guide). This is what I call a “preflight” gate, not a substitute for a proper legal review. It makes legal’s job easier and faster by catching the obvious things before they reach their queue. Build this together with that team.
The editor-in-chief gate. Your head of comms’ own quality control bar. This is the checkpoint everything clears before it goes up the chain. Right now it lives entirely inside one person’s judgment and their calendar. Which works fine right up until they’re on a plane or at a silent meditation retreat.
If you’re public or on an IPO track, add the IR and disclosure gate. Forward-looking statements, Reg FD, quiet-period discipline, which metrics can be disclosed when. Not every company needs this one, but the ones that do can’t treat it as optional.
“We already have our AI platform hooked up to all our docs in [Google Drive/OneDrive/Notion].”
Fantastic. But that’s not a reference core. That’s just a pile of stuff, occasionally useful depending on the job. These 11 docs I’ve described might be somewhere in there scattered about, or exist in the aggregate in fragmented forms. But now you’re asking an LLM to perform forensic analysis on the totality of your team’s work corpus in a search for authority and meaning.3
A reference core must be unified and centralized in whatever doc storage system you use, just as findable by you as it is for your tools and agents. Make someone responsible for protecting and maintaining it. Otherwise, it’ll descend into the same kind of mess that every company’s shared drive inevitably becomes absent someone curating it.
When docs disagree
They will disagree at some point, either in substance or somewhere in the messy probabilistic reasoning of an LLM. Your voice guide will argue for a punchier claim than your fact sheet has cleared, or a sterling client example from 18 months ago will contain a phrase your legal team has since banned.
Decide the pecking order once, in advance, and write it down. I’d suggest something like:
Gates → fact sheet → corporate narrative → voice → gold examples → live intel.
Gates sit at the top here because they hold veto. A voice guide shouldn’t promote a number legal hasn’t cleared. A great-but-old example can’t be allowed to override a phrase that’s now prohibited. If you don’t decide and codify this all now, you’ll end up doing it when something’s already gone wrong.
Where to start
Don’t try to one-shot all 11 docs at once. Start with the corporate narrative and the fact sheet, because everything else draws on that substance and the fact sheet stops errors that get people fired for carelessness. From there, work on the core gates (red lines and legal, plus IR if you’re public), because those are the ones where being wrong is expensive rather than merely embarrassing. Then voice and your audience profiles.
A side benefit of this exercise that’s arguably the main benefit: the assembly and existence of these documents makes your human team better whether or not some agentic tool ever sees them. It’s the housekeeping every comms leader wishes they had time for. The tools are just giving the excuse to finally do it.
We’re all using the same underlying LLMs. Nobody’s AI is inherently better than anyone else’s. Without a reference core, it’s never been easier to be average.
I occasionally reduce this down to “comms exec with a GitHub”.
Can you tell I’m in a home improvement mindset this week?
This is the same fallacy people run into with their executive comms agents, but that’s a newsletter for another day.
Great toolkit Jim, well said.
Helpful!