How to Customize Claude to You
Claude has instructions, projects, connectors, and skills, with no obvious way to tell which one you need. I ran the same request five times to find out.
I ran a lunch and learn with a client’s team, going deep on how they could get more out of Claude. Most were already using it, some for a while, and they wanted to know: what’s the difference between instructions, a project, connectors, and a skill, and which one do you actually need? The four sit there in a menu with no way to tell from the names alone.
I answer it now by running the same request five times, adding one of the four to the desk each time.
“Write the Monday client update for Acme.” Word for word, no edits.
Five different emails came back.
Same seven words every time. What changed was everything sitting around them.
Five Runs on the Desk
I built this as a desk, the actual picture on WorkHoodie’s explainer page: a desk that starts bare and picks up one of Claude’s four features with each run, instructions, then a project, then connectors, then a skill. Same request sent every time. Four of these five are something you could add to your own desk, not just mine.
Start with nothing on the desk. No instructions, no project, no connectors, no skill, just Claude and the request. I send it anyway.
What comes back is upbeat and polished, full of emoji, with specific numbers for signups and cost per customer that I never gave it anywhere. It sounds finished. None of it is real. Give Claude nothing to go on and it doesn’t come back and say “I don’t know.” It picks a voice and invents the rest. Confident, upbeat and unusable is the honest read.
None of that comes from Claude guessing better. It comes from setting a base expectation, once, for how you want it to work.
Now let’s add instructions. Instructions are rules you write once that Claude follows on every request after, not just this one. On the desk, that’s three lines written plainly on a mat, always in view: WRITE PLAINLY, NO EMOJI, NEVER INVENT A NUMBER. I send the exact same request.
The tone changes completely. No hype, no made-up figures. It also tells me flatly that it doesn’t have this week’s numbers and asks me to send them over. That’s honest, and it’s also useless on a Monday deadline, because the rules made it honest, not informed. It still knows nothing about Acme. Instructions are for the tone and the ground rules you want followed every time, not for facts about a particular week.
Being honest about the gap doesn’t close it. What’s missing is the material itself, the account’s own rules and files.
Now let’s add a project. A project is Claude’s word for a dedicated space that holds its own instructions and its own files, separate from the general rules on the mat. On the desk that’s a folder, the Acme client folder, open next to the mat. It comes with three rules of its own: “Client voice, not ours.” “Always cite the sheet.” “No promises on dates.” Three files land inside it: the campaign brief (scope, dates, channel mix), the Q3 data spreadsheet (spend, CAC, retention), and a one-pager on how Acme writes. Same request again.
Now the update sounds like Acme, and the numbers are real. Q3 is measured on retention, not signups, so it leads there: repeat revenue up 11 percent, spend flat, CAC steady. Every number is real and every number has a date on it, and none of the dates are today. Everything in the folder is as of last Friday. It reads like a good update for last Friday, because it is one. A project gives an account its own rules and its own paper trail. It doesn’t know about anything that changed an hour ago.
That’s a fair account of last Friday, not this morning. What’s missing is today.
Now let’s add connectors. Connectors are live links to the tools the team actually works in, not a document someone has to remember to update. On the desk that’s two apps, live, right next to the folder: Drive and Slack. Same request.
Claude checks before it writes and comes back with two things that happened since Friday: the launch slipped to the 14th, surfaced in the team’s Slack channel on Tuesday, and the offer in the brief got cut from 20 percent to 15 percent, edited yesterday at 4:12pm. Two screenshots back it up, timestamped 9:41 and 9:43. That’s the only run with facts that exist nowhere else on the desk. Current, but it still comes back in the wrong shape, not something I’d send as-is. Connectors close the gap between what’s written down and what’s true right now. They don’t touch how the update reads.
Being current doesn’t fix how it reads. What’s missing is the format the team already agreed on.
Now let’s add a skill. A skill is a saved page Claude reads before it answers, the same way a new hire reads the handbook before writing their first update. On the desk that’s one more page landing: the “Weekly client update” page from the team’s own handbook, three wins, one risk, one ask, under 200 words. Same request, one more time.
What comes back needs nothing. The wins are the lifecycle rebuild shipping, repeat revenue up 11 percent, and two landing tests live with spend still flat. The risk is the launch moving to the 14th, which compresses the promo window. The ask is confirming the 15 percent offer by Thursday so creative can lock it in. Same facts as the connectors run, shaped the way the team actually writes these updates. Nothing left to edit, that’s what the last piece adds. A skill is for the shape of the finished thing, the format the team already agreed on, not for facts or tone on their own.
What Actually Changed
Line up the five runs and the order tells its own story: the instructions made the writing honest, the project made it informed, the connectors made it current, and the skill made it finished. The words “write the Monday client update” never changed once across five tries. What changed was how much of the actual business was sitting on the desk when Claude sat down to write.
That’s the part most people miss when they try AI once, get something close to a bare desk, and decide it isn’t ready for real work. It might be exactly ready. It just doesn’t know anything about your business yet, and it never will unless someone decides what belongs in front of it, one piece at a time.
The five runs are how this demo isolates each feature, not an order you have to follow. If your problem is stale numbers, add connectors without ever writing an instruction sheet first. If your problem is only tone, the first one might be all you need.
If you’re only adding one thing to your own desk, start with instructions. They’re the only one of the four that costs nothing to write and the only one that applies to every request you’ll ever send, not just this one.
AI made competent free. The judgment to use it well didn’t get cheaper. Deciding which three rules earn a spot on the mat, which files are worth pulling out of the folder, which apps get a live connection, which page of the handbook actually applies here, that’s still a job. It’s a different job than most people think they signed up for when they opened a chat window and typed a question.
Customize Your Own Desk
Reading about instructions, projects, connectors, and skills is one thing. Watching Claude get it wrong with none of them turned on, then get it right once all four are on the desk, is another.
Go to /how-it-works/claude-chat and run it yourself. Same prompt, sent five times, one of the four features added to the desk each time. It runs in a few minutes and it’s worth watching start to finish, especially the moment the desk goes from a bare table to something that actually knows your business.