Run Your Business on Claude.ai

Most people pay for Claude and then use one percent of it. They open a chat box, ask a question, copy the answer, and close the tab.

That is a search engine with better grammar. It is not a business tool.

The features that actually move revenue are the boring ones sitting one menu away: Projects, Connectors, Skills, Cowork, scheduled tasks, Artifacts, and memory. Wired together, they stop being a chatbot and start being the layer where your operations actually run.

Here is how I think about each one, what it replaces, and the setup that makes it stick.

1. Projects: give the model a company, not a question

A Project is a workspace with its own files, instructions, and memory. Everything you chat about inside it inherits that context.

This is the single highest-leverage feature and almost nobody sets it up properly. Without it, you re-explain your business at the top of every conversation. With it, you explain it once and every future answer is already calibrated to your pricing, your positioning, and your customers.

I do not make one big "work" Project. I make one per function, because the instructions are genuinely different:

• Sales: ICP, objection handling, pricing rules, past proposals that closed.
• Content: brand voice, banned phrases, past posts that performed, distribution channels.
• Finance and admin: invoice templates, vendor list, tax context, currency, fiscal year.
• Delivery: your process docs, SOPs, scoping templates, handover checklists.

The instructions box is where the value is. Most people leave it empty or write one line. Write it like you are onboarding a contractor who is competent but knows nothing about you.

Project instructions I actually use

You are working inside the Sales workspace of a small AI consulting studio.

Context:
- We build AI products and automation for founders and agencies.
- Typical engagement: 3 to 12 weeks, USD 4k to 25k.
- We do not do staff augmentation, hourly work, or pure design projects.

Rules for every response:
- Use the pricing and scope rules in the uploaded files. Never invent a price.
- Write in plain, direct language. No corporate filler. No "leverage synergies".
- When drafting outbound or proposals, lead with the client's problem, not our tools.
- If a request needs information we do not have, ask one specific question instead of guessing.
- Default output format: something I can send, not a summary of what I could send.

2. Connectors: stop being the copy-paste layer

Connectors plug Claude into the tools your business already runs on. Google Workspace, Microsoft 365, Drive, Gmail, Calendar, Airtable, Apollo, Ahrefs, Amplitude, and a long tail of enterprise systems. They run on MCP, which is the same protocol I use for custom integrations in code.

The mental shift: a connector is not just search. Several of them write. The Microsoft 365 connector can draft and send email, manage calendar events, and create files in OneDrive and SharePoint. Once write access is on, the model is not advising you about the task, it is doing the task.

What this replaces in practice:

• Exporting a CSV to ask a question about it.
• Screenshotting a dashboard into a chat.
• Reading forty emails to write one status update.
• Hunting for "that one doc from March".

Prompt: the Monday morning sweep

Using my email and calendar from the last 7 days:

1. List every client or prospect thread where the ball is in my court.
2. For each one, give me the single next action and a draft reply.
3. Flag anything that has been waiting more than 4 days.
4. List meetings this week that have no agenda attached, and draft one for each.

Output as a single table plus the drafts underneath. Do not send anything.

"Do not send anything" is deliberate. Give it write access, then keep the trigger. You want the work done and the decision yours.

3. Skills: teach your process once

A Skill is a packaged set of instructions plus files that Claude loads when the task matches. Think of it as the difference between telling someone how you write proposals every single time, and handing them the playbook.

Prompts are throwaway. Skills are assets. The moment you catch yourself pasting the same three-paragraph instruction for the third time, that is a Skill.

The ones that earn their keep in a small business:

• Proposal writer: your structure, your pricing tiers, your terms, your case studies.
• Brand voice: with real examples of your writing and an explicit list of what you never say.
• Client onboarding: the kickoff doc, the questions you always ask, the access checklist.
• Weekly report: the exact sections your clients expect, in the exact order.

You do not have to write a Skill by hand. Ask Claude to build it from the artifacts you already have.

I am going to paste three proposals I have sent. Two closed, one did not.

Build me a reusable Skill for writing proposals in my style.

It should capture:
- the section order I actually use
- how I describe scope, timeline, and pricing
- the tone and sentence rhythm of my writing
- what I include for credibility
- what I deliberately leave out

Then write the Skill instructions so that a future request like "write a proposal for X"
produces a finished document, not an outline. Ask me for anything missing before you write it.

On Team and Enterprise plans this scales up: Skills can be managed across the whole organisation, and plugins bundle Skills, connectors, and sub-agents into one installable package. That is how you stop having "the one person who is good at prompting" and start having a company capability.

4. Cowork: hand over the task, not the question

Cowork is the piece most people have not tried yet, and it is the one that feels like hiring.

Instead of a conversation, you hand Claude a multi-step job. It works inside the folders and tools you point it at, shows you each step it takes, keeps going when you close your laptop, and hands back a finished deliverable. It runs on desktop, web, and mobile, and it is built for knowledge work rather than coding: research, analysis, document production, data wrangling, competitive teardowns.

Things I have handed to it that would otherwise eat an afternoon:

• Take a folder of 30 call transcripts and produce a themes report with quotes.
• Audit every page on a site against a positioning doc and list the contradictions.
• Turn a messy spreadsheet of leads into a segmented, enriched, deduped sheet.
• Read a stack of vendor contracts and build a comparison table with renewal dates.

How to brief it so it does not wander

Task: build a competitor teardown for the three companies listed in /research/competitors.md

For each competitor:
- pricing model and actual numbers, with the source URL
- positioning in one sentence, in their words
- who they clearly target and who they clearly ignore
- their three loudest claims and whether they back them up

Constraints:
- Only use sources you can cite. If you cannot verify a number, write "unverified".
- Save the output as /research/teardown.md
- Keep a running log of sources in /research/sources.md

When done, tell me the three things that surprised you.

The last line is not decoration. Asking for surprises pulls out the judgement that a plain summary buries.

5. Scheduled tasks: the part that makes it a system

A tool you have to remember to use is not infrastructure. Scheduling is what turns Claude from something you visit into something that runs.

Cowork supports recurring and on-demand scheduled tasks. The obvious candidates are the reports nobody enjoys writing:

• Monday 8am: pipeline status from CRM plus email, with stalled deals flagged.
• Friday 5pm: what shipped this week, drafted as a client update.
• Daily 7am: overnight support tickets grouped by root cause, not by timestamp.
• Month end: expenses categorised from the finance connector, exceptions listed.
• Weekly: competitor pricing and homepage diff against last week's snapshot.

That last one is quietly the best. Nobody has time to check competitor pages weekly, so nobody does, so you find out about a repositioning six months late.

6. Artifacts: ship deliverables, not chat logs

Artifacts turn a response into a real document, dashboard, calculator, or page that lives in its own pane, can be edited in place, and can be shared with someone who was never in your chat.

For a business this is the difference between "here is my analysis" and "here is the tool". Some of the highest-return things I have built as Artifacts in under ten minutes:

• A pricing calculator for a service with too many variables to quote from memory.
• A one-page ROI model a prospect can play with during a call.
• An internal dashboard summarising a data pull, refreshed by a scheduled task.
• A client-facing project status page instead of a fourth follow-up email.

Build this as an interactive artifact, not a description.

A pricing calculator for our consulting services.

Inputs: project type, number of integrations, timeline in weeks, whether ongoing support is included.
Outputs: total price, weekly rate, and a plain-English scope summary I can paste into a proposal.

Use the pricing rules from the files in this project. Do not invent multipliers.
Make it work on mobile. Keep the visual style simple and confident.

7. Memory: the compounding part

Memory means the useful facts survive the conversation. It has moved from vague daily summaries to individual categorised entries that Claude reads and updates, and on the higher plans you can point it at past chats for finer-grained continuity.

The business version of this is simple: the model stops asking you what your business does. It knows a client prefers Tuesday calls, that one vendor is on 60-day terms, that a certain phrase is banned in your copy.

Two habits make memory work for you instead of drifting:

1. Tell it explicitly what to remember. "Remember that we never quote below 4k" beats hoping it picked that up.
2. Audit it monthly. Stale facts are worse than no facts, because they are confidently wrong.

The gap: background agents that write code

Here is the honest limit. Cowork is built for knowledge work, not for shipping code. It will happily research, analyse, and produce documents in the background. It is not the thing you point at a repo and ask to fix a bug, open a PR, and keep working while you are asleep.

For that side of the business, Claude Code is the tool, but it lives in your terminal and it wants your attention. The moment you want coding work happening in the background on a schedule, across branches, without you babysitting a session, you need something built for that shape of problem.

That is exactly the gap tellnova.com is going after: background coding agents. Same idea as handing off a research task to Cowork, except the deliverable is a pull request instead of a document. If your business ships software, this is the half of the automation story that Claude's own background agents deliberately do not cover.

The one-week rollout

Do not try to adopt all of this at once. You will end up with four half-configured Projects and no habit. This is the order that works:

Day 1  Create one Project for the function that eats the most of your week.
       Write real instructions. Upload your five most-referenced documents.

Day 2  Turn on the connectors for the tools that Project touches. Read-only first.

Day 3  Work inside that Project all day. Every time you paste the same
       instruction twice, write it down. That list is your Skills backlog.

Day 4  Turn the top item on that list into a Skill.

Day 5  Hand one genuinely multi-step task to Cowork. Brief it like the
       teardown example above. Judge it on the deliverable, not the vibes.

Day 6  Take the report you write every week and schedule it.

Day 7  Enable write access on one connector. Keep "do not send" in your prompts
       until you trust the drafts.

What this actually replaces

Not people. That framing is lazy and mostly wrong for a small business, because the work being absorbed is usually work nobody was doing in the first place.

What it replaces is the gap between knowing what should get done and having the hour to do it. The competitor research you keep postponing. The follow-up that goes out four days late. The report you write from memory because pulling the real numbers takes twenty minutes. The proposal that is worse than your best proposal because you were tired.

A business does not usually die from lacking a strategy. It leaks from the hundred small things that are obvious and never get done. That is the exact shape of work this stack is good at.

The short version

Projects give it your context. Connectors give it your data. Skills give it your process. Cowork gives it the whole task. Schedules give it a heartbeat. Artifacts give you something to hand a client. Memory makes all of it compound.

Set up one Project properly this week. Not seven. One. The rest only makes sense once the model actually knows who you are.