Why doesn't one person's AI fix reach the rest of your team?

Three reps wrote three versions of the same meeting-prep prompt, and the one that works lives on one laptop. Sales bought one AI tool, marketing another, and each tells the buyer something different. Nothing between those setups is shared, so a fix stays where it was made. A shared harness is the one setup your whole team works AI through, designed from the buyer back. Fix it once and everyone has it at their next session, and the buyer hears one answer.

The form of your harness decides how fast it gets better

Ask four teams what a harness is and you get four answers. Whatever form yours takes, it’s your team’s interface with AI. Each form gets better a different way.

FormWhat it isHow it gets better
A chat assistantA chat app with the company’s connectors switched on. Most teams start here.The vendor ships a feature, or an admin turns on a connector.
A coding agentAn editor where AI meets the work, with rules files in the repo. The engineering version.Someone edits a rules file, and it changes like code.
A workspace with skillsA workspace set up before the employee arrives, with skills colleagues share.A colleague shares a skill, and everyone has it.
The full stackSkills, connections and checks a revenue team built for itself, versioned in GitHub.A pull request to the repo.

In the first form your team waits on a vendor’s next release. In the last, the person doing the task makes the next run better that afternoon, and everyone gets the change. Aim for the last form.

A longitudinal cross-section of a pinecone, with one half depicting naturalistic woody scales and the other half represented as a geometric node-and-link network.

AI-native teams design the harness from the buyer back

Most GTM teams run SaaS GTM design. Every role exists because a buyer needed something from you, and each role now works on its own software with its own AI. Make each role faster and the handoffs stay where they were.

SaaS GTM design

Five roles on five stacks. The buyer climbs over every handoff and gets a different answer at each door.

What the drawing shows

Five roles, each on its own SaaS, on one shared context. Demand gen: campaigns, event invites, webinars, paid media, website. Content: thought leadership, case studies. BDR: first outreach, event registration. Sales: prospecting, discovery calls, demos, proposals, negotiation, renewals. CSM: onboarding, adoption, support. The customer experience line steps over every handoff.

  • Roles defined by human capability constraints
  • Process designed around human specialization
  • Software designed to be sold by the seat

AI GTM design

One team on one harness. The buyer gets one answer, at the first touch.

What the drawing shows

One GTM team, human in the loop, with every task the five roles had, on one shared harness. The customer experience is one flat line.

  • What does the customer need?
  • What can AI do?
  • When do we bring people in to give the customer the experience we want?

If you designed for the buyer, you wouldn’t build five doors into your company that each give a different answer. So don’t look inside the role. Ask why it exists. Why do BDRs alone carry event invites, when everyone on the team has a network to invite? Why should only solutions consultants show a demo or answer a technical question? With the right context in the harness, more of your team can answer the buyer the first time.

Mark Roberge
Mark RobergeCo-founder and managing partner, Stage 2 Capital; founding CRO, HubSpotStage 2 Capital

Roberge built HubSpot’s sales team in the age of specialization: the right hire for every stage of the funnel. Now he says AI runs that logic backward. “With specialization, you can hire the perfect person. But now you have these handoffs that are very inefficient for the customer… AI can empower a single individual to handle all those roles.” In the next breath: “It doesn’t mean you’re getting rid of those people.”

The Science of Scaling, September 7, 2026, at 44:17.

Fewer handoffs for the buyer, not fewer people. One team, as wide as the harness under it. Build the harness for that team, not for the five you have.

A harness starts as a repo and gets better as a loop

The simplest shared harness is a repo on GitHub: a CLAUDE.md file that says how the team works, and a folder of skills. Everyone downloads it, and everyone’s Claude works the way the team works. Say a rep runs the sales-deck skill and every deck opens on your company overview. She changes one line so it opens on the buyer’s problem, and commits it. Everyone’s Claude picks up the change at their next pull or session.

Illustration: one change to a shared sales-deck skill.

The repo is only how the skills reach each laptop. What makes the harness get better is the loop around it: use it, improve it, and everyone has the better one. When an output comes out wrong, rewriting it fixes one page. Fixing the skill, the check or the context behind it fixes every page after it.

The loop, in four steps:

  1. Someone runs a skill, and the output comes out wrong, slow or off-brand.
  2. They tell a builder the same day.
  3. The builder fixes the skill, the check or the context.
  4. Everyone has the fix at their next session, because the harness updates itself.

As of · our brand image service

On October 1, 2026, the image model behind Studio, our brand image service, stopped making transparent backgrounds. Pushed, it painted a checkerboard into the picture. The next day the fix went into Studio: ask the model for white, then remove the white. Every teammate got clean images on their next request, including one who never opens Studio, because her Claude calls it.

A fix can also be a check. Since September 19, 2026, when an agent tries to redraw another company’s logo, our harness refuses the edit and says where to get the company’s own file.

Measure the loop by cycle time: the days from “this is wrong” to “fixed for everyone.” When the fix waits on a vendor, you count it in vendor releases. When the harness is yours, you count it in days. A short cycle is what separates an AI-native GTM team from one that adopted tools.

Anyone can build fast. Building right takes a product team.

Anyone can build now. A seller mocks up the screen she needs in an afternoon; a marketer turns a question into a working page. That’s build fast, and you want all of it. Knowing which skills the team needs, and keeping the ones it runs on working, is build right.

Scroll, and the loop builds.

The whole team works AI through one harness. SaaS, skills and MCPs, standing on shared context and security.

The harness grows past skills. Middleware, the workflows between the systems your team already works in. AI plugins, tools packaged so anyone installs them in one step.

Somebody has to build and run all of that. AI builders: GTM engineers and RevOps. They bring the team AI enablement and take its requirements back. Most of what the team asks for is faster horses: today’s job, with better tools. Worth having.

Two loops, two jobs. The builders grow into an AI GTM product team. On the left, build fast: what the team asks for today, shipped this week. On the right, build right: the processes leaders change, run as systems of action, where routine steps run without a person.

Every box is a place to ask: what’s the bottleneck, and who fixes it?

The whole team works AI through one harness. SaaS, skills and MCPs, standing on shared context and security.

The harness grows past skills. Middleware, the workflows between the systems your team already works in. AI plugins, tools packaged so anyone installs them in one step.

Somebody has to build and run all of that. AI builders: GTM engineers and RevOps. They bring the team AI enablement and take its requirements back. Most of what the team asks for is faster horses: today’s job, with better tools. Worth having.

Two loops, two jobs. The builders grow into an AI GTM product team. On the left, build fast: what the team asks for today, shipped this week. On the right, build right: the processes leaders change, run as systems of action, where routine steps run without a person.

The GTM team improves its own skills: build fast. The AI GTM product team, GTM engineers and RevOps with an executive sponsor, changes the systems of action: build right. Routine steps in a system of action can run without a person, and a named person still owns what reaches a customer.

Every box is a place to ask: what’s the bottleneck, and who fixes it?

Most teams that get this far stay on the left loop. It pays, and nobody’s job changes. The right loop is where the buyer’s experience changes. Take the demo handoff above: redesign discovery to demo so the AE answers routine technical questions from approved evidence and brings in a solutions consultant only past that. Judge it by how long the buyer waits for an answer, and whether deals move. Work like that changes people’s jobs, which is why it needs an executive behind it. Go after it anyway. That’s where the return is.

Most of a harness is bought. The skills that route it are yours.

What makes it yours is that Claude knows which system to go to, and for what. Here is ours.

As of

SystemBuilt or boughtWhat Claude goes there for
HubSpotBoughtThe relationship record: contacts, deals, every outside conversation
SybillBoughtRecords our sales calls and sends each one on to Octave
OctaveBoughtOur packaging: personas, offers and proof points, and the call transcripts
NotionBoughtHow we operate, so Claude reads our procedures where we keep them
SupabaseBoughtThe database our own apps share
StudioBuiltBrand images, asked for in the flow of work

A connection is half the handshake. Octave’s MCP offers our Claude more than 160 actions. With nothing else, an agent reads the list, picks one, tries another, and may not take the same path twice. Our skill is the other half: go to Octave when you have this question, and ask it this way. One question asked one way gives you consistency, and then speed. The skill is where your team’s judgment lives, which is why buying a tool never finishes a harness.

Shared ContextWho should keep the context your GTM team's AI reads?Go deeper

At two hundred people, a release needs a route everyone can see

We’re three people, and a change gets agreed in one message. This is the route we’d expect at two hundred; we haven’t run it at that size.

StepAt two hundred peopleAt Fresh Context, October 2026
IntakeOne place anyone reports a wrong output, with the output attachedA message to Sam
ApprovalA named owner for each skill, check and store; a change to a customer-facing claim also needs that layer’s keeperSam, in one reply
ChecksThe existing checks, plus a fixed set of test tasks run on the old and new versions and comparedStudio’s tests and the checks that refuse a bad write
RolloutA pilot group for a week, then everyone, with the old version one step awayEveryone at once
UseRuns of the new version counted by team after two weeks; a team still on its own copies is a findingWe see it; we don’t count it
ReversalRevert the one change, rerun the checks, tell the people who ran itOne small change and a redeploy
The two-hundred-person route is an expectation. The Fresh Context route is how we run it today.

Naming those owners is the job of the harness’s product owner, the seat GTM product thinking argues for. The checks matter more as you grow, because a wrong fix travels as far as a right one.

Your buyer never meets your org chart. They meet your harness.

This week, follow one buyer across your team, from first touch to signature. Write down every handoff and every system they pass through. Take the worst handoff and ask the three questions: what does the customer need, what can AI do, and when do we bring people in? Then name who fixes the harness when its output is wrong, and count the days the last fix took to reach everyone.

What leaders ask about a shared harness

Our engineers already run a harness. Can GTM use theirs?
Give GTM its own. Engineering's harness in a marketer's session is baggage: the agent sifts through tools the marketer never uses. Nothing says you can have only one. Several can live in one repo, as long as someone decides which one each person loads.
Can people still build their own skills?
Yes, for their own work. Once a second person needs a skill, it moves into the shared harness under an owner. The thing to stop is a private copy of a shared skill, because the copy stops getting fixes.
What about access and security?
Every read and write runs on the person’s own permissions, so the harness never shows or changes anything they couldn’t already. A tool that can’t enforce that stays out of shared sessions until IT approves a scoped option. IT reviews each new connection and any change to permissions or to where data goes. The owner reviews changes to skills and checks.