The Supercharged Approach

Most teams start with AI the same way. A few people use it on their own, and it stays with them. This is how we get from there to a team, and then a company, that works on shared AI.

It is the frame behind everything I do at Supercharged, from a keynote to a company that runs on agents. First every person, then the team, then the company. What we leave behind keeps working after we step back.

Charge levels

How deep does one person go? We count it in five levels. Each level adds one building block and ends at a limit that makes the next one worth taking. We use them to place people in a short survey before we start, and as the spine of every workshop.

0 Ask. Plain chat. You type a question, it answers, and tomorrow it has forgotten you.

1 Context. It knows who you are and what you know: a note on you and how you write, and one folder of your real material.

2 Connections. It reaches your real work. Mail, calendar and notes, read-only at first.

3 Capabilities. It does the weekly jobs. Two routines, written down, that it runs when you ask.

4 Cadence. Things run without being asked. A fixed weekly slot, a log of what you learned, one improvement a week.

After level 4 the work goes wider. That is the step from one person to a team.

Three layers

How far does it reach? There are three layers, and each one depends on the one before it.

Personal. One person and their own setup. Single player.

Shared. A team works on the same knowledge, the same skills and the same rules. Multiplayer.

Mycel. The company layer. Personal setups and team setups grow into one network, linked by consent and held together by rules. It runs without anyone starting it.

You cannot share what nobody uses. And you cannot run a company on skills that live in one person’s account.

Mycel is what Supercharged builds toward. The name comes from mycelium, which grows from many points at once. Its threads form where something useful flows. A company brain grows the same way.

The five Cs and the rules

What do we build? The same five building blocks at every layer: Context, Content, Connections, Capabilities and Cadence. Around them sits one frame of Rules that says what AI may do. Companies need that frame before anything else.

Put the blocks against the layers and you get the map below. Each cell is something concrete we build. Where a team stands is the set of cells that already work. The next step is one cell to the right or one row down.

The five Cs and the Rules at each of the three layers

	Personalsingle player	Sharedmultiplayer	Mycelthe company, runs on its own
Contextit knows who we are	Notes on who you are and how you write, taken from a short interview.	A team brief, shared project instructions, a glossary.	A model of the company built around its core object, such as the customer or the deal. Tone and policies.
Contentit knows what we know	One folder of real material.	A shared knowledge base. Every entry links to its source. Each area has one owner, each file a review date and only what is true now. One project per customer or deal, so nothing mixes.	Sources of truth stay where they are and are read through one gateway.
Connectionsit reaches real work	Mail, calendar and notes. Read-only first.	Team tools such as the CRM or Notion, with team permissions. One map of where data lives and what AI never does there.	One MCP gateway. Everyone signs in as themselves, the source keeps its permissions, every change is logged.
Capabilitiesit does the weekly jobs	Two routines, written down.	A skills registry. Criteria agreed with the people who use the output, three test cases and one owner per skill.	Agents in the background, in draft mode. A second agent checks each result against the criteria. One approval queue.
Cadencestill alive next month	A weekly slot, a log of what you learned, one improvement a week.	A digest of learnings. A correction counts once. When it repeats and someone says yes, it becomes a rule or a test case.	A fixed path from workflow to agent, a health check, reviews after two and six weeks.
Ruleswhat AI may do	Read-mostly. Nothing leaves without my approval.	AI guidelines in three groups: prohibited, guardrailed, allowed. A traffic light for data. AI output shows its sources.	Four fields per area: alone, do and tell, ask first, never. Reading and acting kept apart. An audit log.

How an engagement runs

Two diamonds, as in design thinking. First we open up to understand, then narrow down to what matters. Then we open up again to bring many people along, and narrow down to what gets built and handed over. After that a loop runs: reviews after two and six weeks, then the next round, one cell further.

Understand

Sponsor and mandate first: who carries this, and what is in it for them. A survey about what people do with AI today, which places everyone on a charge level. Interviews with leadership and team leads. We reuse what exists before we ask anything twice. Compliance and IT are in the room from day one.

Focus

A map of the core lifecycle, whether that is the customer, the deal or the product. Who is involved, what information flows, where work repeats. Every process gets one of three verdicts. The existing tools already do it, we build or buy something, or it stays with people. Then a roadmap: what comes first, the effort, the benefit and who owns it.

Enable

AI guidelines on two pages. Workshops along the charge levels, one room per team. Everyone builds their own setup on the same shared knowledge, so the shared parts show up while people build. At least 70 percent of the time is hands-on, and everyone leaves with something that works. Then 1:1s and office hours, which is where most of the building happens. And we watch who starts building on their own.

Build and hand over

Shared skills in one registry, the team’s tools connected through one gateway, and the first agents. Agents start in draft mode: they prepare, a person approves. Every new workflow takes the same six steps, below. At least two people inside the company own what we built, and then we step back.

Where a client starts differs. A VC firm that already uses AI every day goes straight to Enable. A mid-size company starts with Understand. Some start by building one connection to the system they run on, and get the full picture after. The path stays the same. Only the entry point moves.

From workflow to shared skill

Every workflow that becomes shared takes the same path.

1. Capture it with the person who does it today.

2. Agree what good looks like with the people who use its output. Write the criteria down, with three test cases: a normal one, one where information is missing, and one where it must not act.

3. Write the rules. What it may do alone, what it does and reports, what it asks about first, and what it never does.

4. Run it in draft mode. It prepares, and a second check compares each result with the criteria before a person sees it.

5. Release it once it passes its tests and has proven itself on real work.

6. Move it into the background. It runs on a schedule, and people approve what it prepares.

What we believe

Practice before concept. Show, then do. At most seven minutes of demo before people work themselves.

Leave something that runs. Every format ends with a working tool. Done is a URL.

Context beats model. Bad output is usually a setup problem. The day the AI reads your history, it gets better at everything.

Personal first, never alone. Everyone builds their own setup, in one room, on shared knowledge. What helped one person spreads through a weekly digest. A team system pushed onto people ends like the company wiki.

A skill is a team agreement. Before we build a shared skill, the people who use its output agree what good looks like. The skill holds the agreement. People use what they helped build.

Show the source. Anything AI writes for others links back to where it came from. If you would not defend it in the room, do not share it.

Describe the goal. A skill says what done looks like and leaves the steps to the model. It points to the knowledge base instead of copying it, so a price changes in one place. Better models then make every skill better.

AI prepares, people decide. Anything that reads incoming content can at most write a draft. Sending, paying and signing stay human.

A correction becomes a rule only with a yes. A correction counts once. When it repeats, it becomes a rule or a test case, with its reason, after a person agrees. AI never rewrites its own rules. The rules are the work. The model can be replaced.

Tools that outlive the builder. At least two people inside own each one. We step back on purpose.

Use what is there. The client’s own tools come first. We build only where it gives a real edge.

Own your context, rent the models. Rules, skills and connections belong to the client and work across vendors.

We run it ourselves. Every demo is my own system. If I do not use it daily, I do not teach it.

Two of these ideas came from others. ProductBench put the skill as a team agreement into words, and AI FIRST shaped the test cases, the second check and rules that need a yes.

Working together

If your team is past the demo stage and wants AI to keep working when nobody is watching, start at Supercharged. Or send me an email.