GrowthAssistant Got 3,000 Hours a Year Back Without Automating Anyone Out

Claude is probably already in your building. Someone found it useful, started running it for something specific, and now a version of their process is stored as a prompt in Slack or in a document somewhere that only makes sense to the person who wrote it. Nobody else can find it or review it, and there’s no way to tell which of those habits are worth spreading and which ones are just luck. That’s most companies with Claude seats today. The pattern behind this applies the same way in any company where people are already building their own fixes, not just talent placement.

GrowthAssistant had already gone further than that. When I started working with them, they had Claude company-wide, and people were building and sharing skills using the platform’s own features. What was missing was governance, distribution, and shared learning. I’m their outside AI consultant, the person who asks what a recruiter or an account manager is doing by hand that a machine could take off their plate. This kind of work can go much further than most companies ever take it, and GrowthAssistant already has.

1. Manual

   1. Using Claude by hand

      A chat window, one task at a time.

   2. Running skills on your own computer

      Built once and shared. A person still starts every run.

2. Automated

   3. Running on a scheduleNow

      Nobody has to remember to start it.

   4. Running off your own computer

      Handed over with a Slack tag, where the work already sits.

3. Autonomous

   5. Running under its own identity

      People work with it like a teammate, not a tool.

   6. Running as one system

      One system across agents and people.

Where this started

GrowthAssistant brought me in instead of living with that gap the way most companies do. I’m in their Slack for anyone with a question, and early on I ran a lunch and learn for the team, walking through what Claude can actually do once you stop treating it like a chat window. Neither matters much alone. Together they set up a way for what one person learns to reach everyone else. The thing that carries it is a plugin.

What a plugin is

A plugin holds skills, agents, and commands together, and it can be set as a requirement for everyone on a company’s Claude team, not just the people who happen to know it exists. That’s what turns governance and distribution into the same fix: review something once, and it reaches everyone at once, so every improvement compounds instead of sitting on one desk.

AI app

Skills

Agents

Commands

Skills

Agents

Commands

The pattern

The mechanism repeats the same way every time, and it’s simpler than it looks in sequence. Governance runs through me. Distribution runs through the plugin I just described.

The person closest to the problem writes the first version, and the version everyone actually uses still gets built to the same standard as the rest of the folder. Every skill comes through me for review, so I’m the only one who sees all of them. When the same problem shows up in more than one person’s work, I teach whoever’s hitting it, or fix it once and let the plugin carry it the rest of the way.

1. In the work

   Someone hits a problem in their real work.

2. They build it

   The person closest to the problem writes the skill themselves.

3. Review and coach

   I review it, and coach that person on what I'm seeing across the team.

4. Everyone gets it

   The reviewed skill lands in the plugin, on every desktop.

and the next problem starts it again

What moves a company up

The cycle you just saw produces the plugin, and the plugin gets a company most of the way there. What comes after is the agent, work running under its own identity instead of a person’s, and then the app, one system tying it all together.

What it produced

In the first ninety days, I projected 2,750 hours a year of returned time, work people kept choosing on their own rather than a folder of skills written and forgotten, and that pace held once the window closed: the current trend has it on track for 3,102 hours a year, 2.2 times the weekly rate of the prior sixty days. One of the skills that went into the plugin helps their recruitment team scale: someone on the GrowthAssistant team built the first version because they knew the problem better than anyone, I brought it up to the same standard as everything else in the folder, and the plugin carried it to everyone rather than leaving it on one desk. The review habit and the plugin kept all of it running without me in the room. A contractor who ships a batch of skills and leaves doesn’t produce a number that keeps growing after the invoices stop.

When the tool doesn’t reach where you work

Claude can’t always reach the systems a business already runs on, Slack included, so someone copies data by hand instead. A connector fixes that, letting Claude reach those systems directly, on a standard called MCP.

The connector you need doesn’t always exist, and when a version of it does, it’s often built for a shape of company you don’t have. Claude’s native Slack connector is built to bind to a single workspace. GrowthAssistant runs two, and there’s no setting that points one connector at both. The limit is in the tool, not in anything GrowthAssistant did.

AI app

Slack workspace A

Slack workspace B

I built two custom connectors instead, one for each workspace, rather than asking GrowthAssistant to merge their two workspaces into one. A handful of people are connected today, and dozens more already have Claude in both workspaces and could join tomorrow without anyone provisioning anything new.

Each one is a custom MCP I built, not Claude’s native connector, which puts the permission rules in my hands instead of the platform’s. Read-only, draft-only, a channel allow list, a block list: those are the toggles I can set and adjust as the work changes, not fixed defaults inherited from Slack or Anthropic.

AI app

Slack workspace A

Slack workspace B

This closes a gap, and what they already had runs stronger for it. Their work still runs on a schedule, on their own machines, not off them. What changed is that the plumbing for what comes next exists now, because I built it, not because GrowthAssistant asked for it by name.

Where they actually are

The diagram above runs on one axis: where the plugin runs, and who sets it going. The three phases separate on who’s still in the loop. Manual means a person starts every run. Automated means nothing gets redone by hand, but a person still approves. Autonomous means it acts under its own identity, with nobody watching.

GrowthAssistant sits squarely in the automated phase: the plugin runs on a schedule on people’s own machines, every skill reviewed by hand before it ships, with nothing yet running under its own identity. That’s the honest state of the work right now.

We didn't have anyone here who could own AI, and we assumed fixing that meant a full-time hire. Now we have Sean. When we're deciding which tools to buy or what to build next, he brings a clear recommendation, and I trust his judgment on it. I think we can get to all of our goals that way.

Most companies that ask me for agents are really asking to skip the step where a person still reads every draft before it goes out. GrowthAssistant hasn’t skipped it. Every skill and both connectors are still routed through someone deciding whether the output is good enough to send. That’s slower than the pitch. It’s also the only way autonomous work, the kind that runs without anyone watching, ends up trustworthy enough to turn on.