You’re Using 10% of Claude
The nine bottlenecks stopping you and the seven-step setup that clears them. I’ve run my whole business on this for three years.Open Claude. Blank box. Type. Read. Copy it somewhere else. Close the tab.That’s the 10%. It’s what almost every business owner does with it, every day.The other 90%…
The nine bottlenecks stopping you and the seven-step setup that clears them. I’ve run my whole business on this for three years.Open Claude. Blank box. Type. Read. Copy it somewhere else. Close the tab.That’s the 10%. It’s what almost every business owner does with it, every day.The other 90% consists of instructions that run before every conversation, containers that hold your business context permanently, memory that carries across sessions, packaged workflows that fire on their own, live connections to the tools you already pay for, and jobs that run while you sleep.None of it is hidden.How to Use Claude Better Than 99% of PeopleYou are sitting on a Ferrari and driving it like a bicycle.By the end of this issue, you’ll have the Automation Audit worksheet and my full Custom Instructions block, ready to use tonight.Most business owners spend between two and four hours a day on admin. Email, approvals, chasing updates, pulling reports, re-entering data that already exists somewhere else. Two hours a day across a working year is roughly 500 hours.You already know this. You’ve probably already tried to fix it with AI, got something half-working, and quietly stopped using it around week three.I did too. Last year I spent about six weeks building automations around a content process that didn’t work when I ran it by hand. Every automation I built made the mess arrive faster.The Ladder Almost Nobody ClimbsThere’s a ladder from that blank chat box to a system that runs without you. Seven rungs. Level 0 is where 99% of people live.You type, it answers, you leave. No memory of you, no context about your business, nothing reusable.Every conversation starts from zero, which means you pay the explanation tax every single time you open it. Levels 1 through 7 are where Claude stops being a chatbot and becomes infrastructure. You set instructions that run before every conversation.You give each function its own room. You let it remember you. You package your best prompts so they fire automatically. You plug it into the tools your work already lives in. Then you stop asking and start delegating.Every rung takes about twenty minutes and pays back over months rather than instantly. That’s the entire reason almost nobody climbs it.Why most of this dies before rung twoThe research on this is unusually blunt, and it’s worth knowing before you start. Deloitte’s Tech Trends work this year found that most organisations automate their existing processes rather than redesign them.Only around 11% have moved to genuinely redesigned, agent-first workflows. Everyone else made a broken process run faster.McKinsey’s agentic AI research found that among pilots that stall, roughly 61% of failures trace to governance and direction rather than model capability. Deloitte’s State of AI work points to skills, not compute.RIP McKinseyThe Forbes analysis in July was direct about it: the businesses that succeed are the ones automating repetitive, well-documented processes.Well-documented is doing all the work in that sentence. Claude is not your bottleneck. It hasn’t been for a while.The nine bottlenecks and the rung that clears each oneRead this first. Find yours. Then read only what applies.Bottleneck 1 has no rung, because it happens before Claude opens. Here’s the fix, and it takes twenty minutes. Pick the task that annoys you most.Run it manually once, and narrate every decision into Claude as you go.Not the clean version. The real version, including the part where you check the client’s last invoice before deciding the tone of the email.Then paste this:I just described how I actually do this task, includingthe messy parts. Write it back to me as a numberedprocess document.For each step, mark it as one of:- MECHANICAL (same every time, no judgment needed)- JUDGMENT (I decide based on context)- RELATIONSHIP (a person needs to feel something)Then tell me which MECHANICAL steps could run withoutme at all, and flag any step where my description wasvague enough that you'd have guessed.“Flag any step where my description was vague“ is the line that saves you six weeks. The vague steps are exactly the ones your automation gets wrong, and you won’t notice until it’s been wrong for a month.Automate the mechanical. Assist the judgment. Never automate the relationship.If this is useful, forward it to one business owner who’s already tried and given up. Takes 10 seconds.Anthropic publishes a full guide on how their models want to be instructed. Most people have never opened it.→ docs.claude.com/en/docs/build-with-claude/prompt-engineering/overviewI’ve pulled the parts that change business output into Step 4. Skip the rest unless you enjoy that sort of thing.Old way:Open chat → re-explain your business → get a generic answer → rewrite most of it → close the tab → repeat tomorrow.New way:Claude already knows your business → already has your files → already has your rules → produces something you send instead of something you fix.Ninety minutes of setup separates those two.You Now Know The Ladder Exists. Below The Line Is How You Climb It.All seven steps in full: the Custom Instructions block, the Projects, the skill build, the connectors, the scheduled tasks that run while you sleep. Plus the Automation Audit worksheet and the sorting template.Free readers stop here with a good idea. Paid readers close this tab with a system that runs without them.Step 1: Write your Custom Instructions onceClears bottleneck 2. Removes roughly 5 to 10 minutes of re-explaining per working day.Custom Instructions are a block of text Claude reads before every conversation you have, permanently. Not per chat. Every chat.Highest-leverage twenty minutes on this entire list, sitting in a settings menu almost nobody opens.Quick setup:Profile icon, bottom leftSettings, then PersonalizationFind the instructions fieldPaste your block, saveOpen a new chat and test itWritten for a business owner:I run a [type] business with [number] people. My clientsare [who]. I personally handle [the functions you own].How I want you to respond:- Lead with the answer. Context after, only if I need it.- Give me the finished thing, not a description of how you'd approach it. If I ask for a draft email, write the email.- If my request is vague, ask one clarifying question before starting. Not three.- If you think my plan is weak, say so directly and tell me why. I'd rather be corrected than agreed with.- When you give me options, rank them and tell me which one you'd pick and why.- Assume I'm short on time. No preamble.If you think my plan is weak, say so directly is the line doing the heavy lifting. Without it Claude defaults to agreeing with you, and an advisor who agrees with everything is worth nothing to a business owner.Test it properly. Ask something you’d normally ask. If the answer doesn’t feel different, your instructions are too vague. Add specifics until it does.Step 2: Give every part of the business its own ProjectClears bottleneck 2, the structural half. Kills the re-upload and re-explain loop entirely.A Project is a container holding its own instructions, its own files, and its own context, separate from everything else you do.Without Projects, everything bleeds. Your client tone leaks into supplier emails. Last month’s numbers contaminate this month’s analysis.Quick setup:Projects in the sidebar, create oneName it after a function, not a taskWrite instructions specific to that function onlyUpload the files it should always have access toNever work in that stream outside the Project againThe split that works for a small business is by function:Clients - service descriptions, standard terms, proposals that closedFinance - invoice template, payment terms, last four quartersHiring - job descriptions, your interview questions, culture docMarketing - voice rules, past campaigns, what performedMy Banner Generator Project holds my format reference, past issues that performed, and a rules file. When I open it, Claude already knows the format. I haven’t re-explained it in over a year.“Name it after a function, not a task. “Clients” survives for years. “Q3 client emails” is dead in ten weeks, and you’ll build forty dead Projects before you notice.”Step 3: Turn Memory on, then prune itClears bottleneck 3. Memory carries what Claude learns about you across separate conversations. You stop reintroducing yourself. Turn it on. Then read the honest part.Quick setup:Settings, then PersonalizationEnable memoryWork normally for a weekOpen memory settings and read what it savedDelete anything no longer trueStale memory actively degrades output, and it’s the failure mode nobody warns you about.I had a note sitting in mine for months about a product I’d abandoned in week two. Claude kept angling suggestions toward it. It took me around six weeks to work out why the recommendations felt slightly off. The fix took ninety seconds once I found it.For a business, this matters more than it does for an individual. Old pricing. A supplier you stopped using. A service you no longer offer. Every one of those quietly poisons every answer until you delete it.Read your memory monthly. Delete aggressively. Three accurate things beat thirty where four are wrong.Step 4: Structure the prompt, and say whyClears bottlenecks 4 and 8. This is where review overhead dies.Anthropic’s framing is to treat Claude like a brilliant new employee with amnesia. Capable, fast, no context on how you work.Four things belong in a real prompt. Role, task, context, constraints. Then one more that almost everyone skips.Lazy prompt:Write a follow-up email to a client who hasn't paid.Structured prompt:You're writing as me. I run a six-person design studio.The client is a mid-size retailer we've worked with fortwo years and want to keep. Task: write a follow-up on invoice 2231, 18 daysoverdue, $4,200. Context: they've always paid, usually late, usuallyafter one nudge. Their finance contact changed in Juneand I think the invoice is sitting with someone whodoesn't know us. We have another project starting withthem in three weeks. Constraints:- Under 120 words- Warm, not apologetic, not stern- Give them an easy out that isn't an excuse- One clear action with a date on itWhy this matters: I need the money and the relationship.Push too hard and I damage a project worth more than theinvoice. Too soft and this sits another month. Optimisefor getting paid without either of us having to mentionit again.“Why this matters“ is the phrase most people leave out. Explaining the reason behind a constraint lets Claude make sensible calls in the gaps you forgot to specify, instead of following your rule literally into a situation you didn’t anticipate. This one line is the difference between output you send and output you rewrite.One more thing worth knowing. These models follow instructions precisely. Ask Claude to “suggest some changes,” and you get suggestions. If you wanted the changes made, say “make these changes.” That distinction has cost me more wasted turns than anything else on this list.Before you go further, slow down here.You’ve read four steps. The instinct is to set all seven up this weekend, feel productive, and abandon five of them by Wednesday. I’ve watched people do exactly that with every system I’ve ever published.Do Step 1 tonight. Only Step 1. Live with it for a week.The owners who actually get their two hours back aren’t the ones who set up the most. They’re the ones who set up the least and then never stop using it.Step 5: Turn your repeated prompts into SkillsClears bottleneck 5. A Skill is a set of instructions Claude loads automatically when the context matches. Write it once; it fires forever; you never paste anything again.That invoice prompt in Step 4? Don’t retype it every month. Put it in a skill and say “chase the Henderson invoice.”Quick setup:Find a prompt you’ve used more than five timesCreate a folder named after the job it doesWrite a SKILL.md inside itThe description decides when it triggers, so write it for a machine judging relevance, not a human reading a menuUpload, then test with the vaguest request you can think ofMinimum viable version:---name: invoice-chaserdescription: Use whenever the user needs to chase an unpaid or overdue invoice, including when they just name a client and say the invoice is late.---# Invoice Chaser## Always ask first if not given- Client name, invoice number, amount, days overdue- Is there active or upcoming work with them## Tone rules- Warm, not apologetic, not stern- Never imply they're difficult- Give an easy out that isn't an excuse## Structure1. One line of genuine context, not small talk2. The specifics: number, amount, date sent3. One clear action with a date4. Close that assumes good faith## Never- Under 30 days overdue: no mention of late fees- No "just following up"- No more than 120 words“The description field is the whole skill. Write it too narrowly and it never triggers. I’ve built genuinely good skills that sat unused for weeks because the description was wrong.”Now the honest part, and it’s the most important thing in this issue.Do not start here. I built skills before I had a repeatable process and burned about six weeks on it. You cannot package a workflow you haven’t run by hand enough times to know its shape.If you skipped the twenty-minute documentation exercise near the top, go back and do it. A skill built on an undocumented process just makes bottleneck 1 permanent, and now it’s permanent and automated.Step 6: Connect Claude to where the work already livesClears bottleneck 6.Cowork goes further, reading your local files directly and working on them as an agent while you do something else.Quick setup:Open the mode selector, switch to CoworkPoint it at one folder, not your whole driveInstall connectors for the tools you use dailyGive it a task with a defined end stateWalk away, come backThe biggest mistake is handing over everything. “Organise my files” produces chaos. “Rename every invoice in this folder to YYYY-MM-vendor-amount and sort them into subfolders by quarter” produces exactly that.Describe the end state. Fill in what Claude can’t guess.Where this compounds is in combinations. If your meeting notes live in Notion and your call transcripts land in Drive, Claude can check one against the other and surface every commitment that never made it into the notes. One connector is convenient. Four talking to each other is a different category of tool.This is also the honest answer to tool sprawl. You probably don’t need another subscription. You need the ones you already pay for to talk to each other.Step 7: Stop prompting. Start delegating.Clears bottlenecks 7 and 8. Scheduled tasks let you describe a recurring job once and have it run on its own. Friday file cleanup. Monday pipeline pull. Inbox triage before you’re awake.Quick setup:Inside a Cowork Project, create a scheduled taskDescribe the job as if briefing someone who can’t ask you questionsSet the cadenceCheck the first three runs before you trust itNow bottleneck 7, because it’s the real one.Around a quarter of business owners say they don’t delegate because by the time they’ve explained it, they’d have finished it themselves. That instinct is correct for a one-off and completely wrong for anything recurring.The maths is simple. A twenty-minute brief for a weekly task pays back in two weeks and then keeps paying for years. The same brief for something you do twice a year never pays back at all.Delegate by frequency, not by difficulty. The boring weekly thing beats the hard annual thing every time.The honest limitation: scheduled tasks need your machine awake and Claude Desktop open. Close your laptop at night, and the overnight job doesn’t run. I’ve watched a lot of people set these up and quietly wonder why nothing happened.What the other 90% looks like in a real weekInvoicing and chasingThe skill knows your terms and your tone. Name a client, get a send-ready email. Three to five hours a month back for most owners, and invoices with automated reminders get paid noticeably faster.Client onboardingA Project holding your standard terms, service descriptions, and past proposals. Enquiry in, tailored proposal out, in the time it used to take to find last month’s version.The Monday reportA scheduled task pulls the numbers, formats them your way, and it’s waiting before you open your laptop.Inbox triageNot replying for you. Sorting, flagging what actually needs you, and drafting the four replies that are always the same four replies.A second opinion before anything shipsYour Custom Instructions carry the “tell me if this is weak” rule into every chat. Every proposal gets an honest read without you remembering to ask for one.These are the ones that pay. Skip everything else until these run without you thinking about them.Bottleneck 9: nobody owns itThis is the one that kills it after everything else works.You’ll set this up, it’ll run for three weeks, then a busy fortnight arrives, a scheduled task fails silently, and nobody notices for a month.The research is consistent here. Stalled AI projects fail on ownership and direction far more often than on capability.So before you close this email, write one line:One process. One owner. One review date.If it’s just you, you’re the owner, and the review date goes in your calendar, thirty days out. That’s it. That single line is the difference between the businesses where this sticks and the ones where it becomes another thing they tried.In 60 seconds:Document the process by hand before automating anything.Write Custom Instructions once. They run before every chat.Give every business function its own Project.Turn Memory on, then prune it monthly.Structure prompts, and always explain why.Package repeated prompts into Skills.Connect Claude to the tools you already pay for.Schedule the recurring jobs and check the first three runs.Save this. Screenshot it. Come back to it.The Automation Audit worksheet, the MECHANICAL / JUDGMENT / RELATIONSHIP sorting template, and my full Custom Instructions block, in one file.Click hereRun the audit on one task this week. Not five. One.We tested it in public. Now go use it in private.If this issue gave you something real - a framework, a starting point, a moment where something clicked - share it.PS: Which of the nine is yours?Hit reply with the number; I read every response.Hamza 💙This story is published under the Generative AI publication. Connect with us on LinkedIn and follow Zeniteq to stay in the loop with the latest AI stories. Let’s shape the future of AI together!You’re Using 10% of Claude was originally published in Generative AI on Medium, where people are continuing the conversation by highlighting and responding to this story.Source: Generative AI Pub — Published — Category: Image AI