Same Claude. Same plan. Different system. This is the setup guide that turns a chatbot into a second employee.
Most people use Claude the same way they use Google: type a question, get a generic answer, rephrase, try again. That's not a system. That's a search bar with better grammar. What follows is the difference between Claude-as-chatbot and Claude-as-colleague.
Same prompt, different system
"Hi Sarah, Wanted to send a quick update on the project. We've made solid progress on the key deliverables this week and things are moving along on schedule. I'll circle back later this week with more specifics once a few items are finalized. Let me know if anything comes up in the meantime. Thanks, James"
"Hey Sarah — quick update on the website refresh. Marcus wrapped the content audit yesterday and we're moving into wireframes this week. Two things still open from Friday: the services-page architecture and whether we're rebuilding the blog or keeping it as-is. I'll have both mocked up by Thursday. Let me know if the timeline still works on your end."
Same prompt: "write a client update email." The difference is the system around it — context files that know your business, your voice, and this specific client.
You don't need to be technical. You don't need to understand prompting. If you can describe how your business works, you can set this up.
The teams getting the most from Claude treat it like a new hire who needs onboarding, not a search engine that needs better questions.
The always-on layer. Your name, how you want Claude to communicate, what kind of output you prefer. Two minutes of configuration that changes every conversation going forward.
Open Claude, go to Settings → General → Personal Preferences. Set the basics:
This is lightweight by design. Think of it as the baseline Claude reads before every single conversation, regardless of context.
The business layer. Who you are, what you do, how you work, what matters. This is the single most impactful thing you can configure. It's why Claude stops sounding generic.
Open Settings → Cowork → Global Instructions and write a paragraph that covers the essentials. Start with something like this:
The more specific you are here, the less you correct later. "We're a marketing agency" isn't enough. "We're a 6-person B2B content agency in Denver that writes long-form thought leadership for fintech founders" is.
The per-engagement layer. Create a project for each client, initiative, or area of work. Inside the project, the right-side panel holds three things that travel with every conversation in that project: Instructions, Scheduled tasks, and Context (folder access and memory).
Open a project, click the pencil next to Instructions, and write the specifics: who the client is, what the deliverable is, what's already been decided, and the one thing you can't get wrong. Three to five sentences is plenty.
You are Claude, working with Sarah on the Q2 website refresh. Tone: direct, no jargon. Don't assume the blog rebuild is approved — it's still open. Sign off as me, never as a team.
The Scheduled block is where this gets powerful. Attach a recurring task to the project — a daily log, a weekly status draft, a Monday briefing — and it runs on its own. Project instructions plus scheduled tasks turns the project from a folder of conversations into a system that does work while you're not in it.
The personality layer. Take that piece of writing you brought (the one from the prereqs). Paste it into a Claude conversation and ask:
"Extract my writing patterns: sentence length, vocabulary, tone, formality level, what I seem to avoid."
Copy the analysis. Paste it into your Global Instructions from Step 02. That's it. Every conversation now reads with your voice as the baseline.
This is what makes output sound like you, not like AI. The difference is immediate and obvious.
Not a test. An actual task you'd normally spend 30 minutes on: a client email, a proposal introduction, a meeting prep document. Run it through the configured system and compare the output to what you'd normally write.
This is where the "oh, this is different" moment happens. The output won't be perfect. But it will be recognizably yours in a way that unconfigured Claude never is. That's the proof that the setup works.
Steps 01–05 set up the personal layer. This part is where Claude stops being a generalist assistant and starts being shaped to how you actually work. Two moves: organize your work into projects, then add capabilities that match what your day looks like.
Everything inside a project stays in scope: the instructions you wrote in Step 03, the scheduled tasks that run on their own, the context folder Claude can read from, and every conversation you've had in that project. Open a project and you drop back into the right headspace in one click.
Create one per client, per initiative, or per area of responsibility. Don't over-engineer it. Most people need three to five projects to cover 80% of their work.
Out of the box, Claude can think and write. To let it act in your tools and run repeatable work, you add capabilities. Three types, stacked on top of each other.
Repeatable workflows Claude runs on demand. Each skill enforces its own output format, style, and completion criteria — so the same task produces consistent work every time.
Bridges to the tools your work already lives in. Connectors let Claude read from and act in Google Drive, Gmail, Calendar, ClickUp, Notion, Canva, Figma, Slack, Linear — and many, many more.
Packaged bundles of skills and connectors, typically organized by role. Install one plugin and you get the whole stack — the right skills plus the connectors those skills need.
Click Customize in the Cowork sidebar. Three tabs: Skills, Connectors, Plugins. Install what matches your work. Resist the urge to install everything — unused capabilities are noise.
Start opinionated: install the Productivity plugin. It's useful for everyone. Then add the plugin that matches your work — Sales if you run a pipeline, Operations if you manage projects, Brand Voice if you write for clients.
Plugins give you pre-built capabilities. But the real unlock is building your own. Here's a skill you can create in two minutes that'll change how every session starts.
daily-log
--- name: daily-log description: > End-of-day project log that captures what was accomplished, what was decided, what's unresolved, and what to pick up tomorrow. Designed to run as a scheduled task at end of day so every new session starts with full continuity. version: 0.1.0 --- # Daily Log You are the team's daily continuity system. Your job is to review everything that happened in today's session — tasks completed, decisions made, files created or changed, open threads — and write a structured log entry that lets tomorrow's session start exactly where today left off. ## Critical Rules - **Format is non-negotiable.** Follow the entry structure exactly. No prose paragraphs. No filler. Every line carries information. - **Write for discovery.** Use consistent section headers, ISO dates, and short declarative lines so Claude can grep this file and find relevant context fast. - **Latest entry first.** New entries go at the top of the file, below the header. - **Rolling window.** Keep the last 5 business days in daily-log.md. Move older entries to daily-log-archive.md. - **One entry per day.** If invoked multiple times, update the existing entry — don't create duplicates. - **No fabrication.** Only log what actually happened. ## Entry Format ``` ## YYYY-MM-DD — [Day of week] ### Done - [Specific deliverable — name the thing, not the activity] ### Decided - [What was decided] — why: [one-line rationale] ### Changed - `[filename]` — [what changed and why] ### Open - **[Thread]** — waiting on: [who/what] · next step: [action] ### Pick up tomorrow - [First priority] - [Second priority] - [Third priority] ``` ### Format rules **Done:** Name deliverables, not activities. "Fee proposal for Greenleaf — $3,500 flat, sent to David" not "Worked on the fee proposal." **Decided:** Every decision gets a **why**. Three months from now, someone will grep this file asking "why did we go with flat fee?" and the answer needs to be here. **Open:** Bold the thread name. Always include what it's waiting on and next step. **Pick up tomorrow:** Ordered by priority. Max 3 items. This is what Claude reads first in the next session to orient.
Once you see it work, schedule it. Cowork → Scheduled → set it to run at end of day. Tomorrow morning, every project has yesterday's full context waiting. That's the compound effect — work keeps moving even when you're not at the keyboard.
The setup takes twenty minutes. Keeping it useful takes awareness. Here are the four ways most configurations stop working.
"We're a marketing agency" isn't specific enough. The system works exactly as well as the context you give it. If your description could apply to any competitor, Claude's output will read the same way.
Nobody updates it. Three months in, Claude is working from stale information: old client names, outdated processes, a voice profile from before you changed direction.
Thousands of skills exist. None of them know your business until you configure them. Twenty installed-but-unconfigured skills is worse than two that are dialed in.
The team uses the system for a week, then reverts to asking Claude questions the old way. The habit matters more than the setup — the routine is the product.
If your context files could describe any business in your industry, they're not specific enough yet. Go back to step 02 and add the details only you know: your actual clients, your real process, the phrases you'd never use, the thing that makes your work different from every competitor's.
This guide gets you to Stage 2 — that's already further than 90% of Claude users. Stage 3 is where it gets interesting: skills and automations working together so the system does real work without you driving every interaction. Stage 4 is where the system starts doing work your business couldn't afford before. I'm building guides for both. Click any stage below to see what's there.
A weekly dispatch from the field. What's working, what isn't, what I'm building next.
friday field notes →