On the 26th of September I asked Claude to show me what it was carrying before I typed a word. One hundred plugins. Eighty-one of them I had installed in one click back in June, and the list included legal research, Kubernetes and financial-modelling tools my pelvic health clinic will never use. Every one of them had been eating Claude's attention on every chat, all summer. Turning off 77 of them took one chat. That was the fifth point on Nick Saraev's list, and it was the biggest fix on it.
Nick Saraev has spent more than $30,000 and about a thousand hours on Claude Code. His video, I've spent over $30,000 on Claude Code, is the fifteen things he wishes he had known at the start. The ideas here are his. The plain words and the mentoring frame are mine, because most of the people I coach do not write code and still need every one of these. I have kept his order, simplest to most technical, and numbered the moves 1 to 50 myself, so you can say "I'm on move 23" and we both know where you are.
You do not need to be technical. I am a pelvic health physiotherapist who learned this by paying for it, one chat at a time. You need a Claude you can open, one real task you do every week, and the patience to run a prompt more than once.
A prompt you can trust, because you tested it more than once and kept the version that scored higher.
A leaner Claude that starts faster and thinks clearer, because you turned off what it was carrying for no reason.
A handoff habit, so a long job never turns to mush. You close the chat, fix the summary, and start clean.
Two doors, in plain words
Every move here works in one of two places, and some only in one. I say which under each principle.
| Door | What it is | What it has |
|---|---|---|
| The Claude desktop app (or claude.ai) | The chat window most people use. Connectors to Gmail, your calendar and your files live in Settings. | Every prompt in this guide. None of the slash commands. |
| Claude Code | Claude running in a terminal, inside a folder on your computer. It reads a file called CLAUDE.md at the start of every chat and can run several helpers at once. | Everything, including /context, /compact, /btw and /insights. To install it, ask the desktop app: "Walk me through installing Claude Code." |
Before you start
- Claude, signed in. The desktop app covers sittings one and two. Claude Code opens the rest.
- A preferences file, or the willingness to make one. In Claude Code it is
CLAUDE.mdin your folder. In the desktop app it is the instructions box of a Project, or your preferences in Settings. If you have neither, principle 13 makes one. - One real task you do every week to practise on. Mine is the weekly review of my clinic and shop numbers. Pick something you will do anyway.
1 of 15Never trust the first answer
Claude is not a calculator. The same words in can give a different answer out, so one great result proves nothing.
- Same words in, different answer out. Run a prompt twice and you get two answers, sometimes close, sometimes nothing alike. Never turn a prompt into a team rule because it worked once.
- Run it ten times and count. Seven out of ten is your real score, not the lucky run you remember.
- Change one thing, run again, compare the counts. Keep the higher number. Nick calls these test runs "evals". You are being a scientist for ten minutes.
- Shrink the gap between the best run and the worst. You have to live with the worst run too, so steady beats brilliant.
Every prompt I hand a mentoring client now gets three runs on the same example before it leaves my hands. Both doors.
2 of 15Give Claude a way to check its own work
A first draft is rough, for people and for Claude. Quality comes from the check-and-fix loop.
- Think in drafts. One attempt is a first draft, so expect first-draft quality.
- Build in a loop. Draft, check, fix, check again. Quality climbs every time round.
- Give it something concrete to compare against. An example you love, a checklist, a screenshot of the target, a test the result has to pass.
- A rough checklist with a fix loop beats a perfect prompt with no loop. Something to compare to is what changes the result.
Every email I draft with AI runs through a small checker of my own banned words before I read it. It catches what I have stopped seeing. Both doors.
3 of 15Let the thing itself carry the instructions
Separate notes drift. Put the rules inside the file they are about.
- Separate notes drift. The plan says version 1 while the real thing is on version 4, and Claude trusts the wrong one.
- Put the instructions inside the file they describe. Rules in the template, the checklist in the document, comments in the code. Then Claude has to read them every time it works there.
- One exception: your preferences file.
CLAUDE.mdholds how you like to work, your lessons, your rules. It does not describe what the thing does. - Feed the source. Paste the transcript, the email, the brief. Your summary of it loses the parts you did not know mattered.
4 of 15Let Claude write the prompt
- Stop writing the perfect prompt. Say "I want X for Y audience. Help me write the prompt." Claude is now better at prompting Claude than you are.
- Answer its questions. Who it is for, what great looks like, whether you have an example. Then use the prompt it writes.
- Skip this for quick daily asks. Use it for anything large, repeated or expensive.
Both doors. Run it in a fresh chat, and when the prompt comes back, run that in another fresh chat (move 19).
I want [what you want, for example a weekly update email to my team] for [who it is for, for example my five-person clinic team]. Help me write the prompt. Ask me what you need to know first. 1. Ask me, all at once: what a great result looks like, whether I have an example I love, and anything else you need to know. Wait for my answers. 2. Write the prompt I should use, addressed to you, with my answers built in and any fill-ins left in [square brackets]. 3. Tell me in one line to paste it into a fresh chat, and anything I need open first (a file, a folder, a connector). Then stop. Next: run the prompt in a fresh chat, and bring the result to Prompt 2.
5 of 15Watch what eats Claude's attention
About a third of Claude's attention is used up before you say a word. This is the principle that found my hundred plugins.
- Type
/context. It shows what is loaded, tools, connectors, memories and skills, and how much room is left for you. - Turn off connectors and plugins you are not using today. Every one you load costs attention and slows the start of every chat.
- Longer chats cost more and get dumber, both at once. Nick calls this context rot. It cannot be avoided, only managed.
- Copy the good prompt into a fresh chat. When Claude has spent ten turns helping you shape a brief, do not run it in the same tired chat.
- Compact early.
/compactfolds the chat so far into a short summary and keeps going. Nick does it long before Claude offers to.
Claude Code has /context and /compact. To switch things off, type /plugin for plugins and /mcp for connectors.
The desktop app and claude.ai do not have these commands. The nearest thing to /context is opening Settings, then Connectors, and counting what is switched on. The nearest thing to /compact is a new chat with the good prompt pasted in.
6 of 15Loosen the leash
- Treat Claude like a contractor. A contractor has their own good process from hundreds of jobs. Tell them the outcome and skip the steps.
- Say what done looks like. "You are done when this is true, that is true and that is true."
- Drop the long "never do this" lists. Those were for older models. Today's Claude infers most of it.
- One line beats ten rules. "If you are not sure, ask me. This matters." Put it in your preferences file once.
Every repeatable job I give Claude now opens with a Done when line. Both doors.
7 of 15Diagnose before you fix
- Say "list every problem you see. Do not fix anything yet." Claude has its own idea of what counts as a problem, and it is not always yours.
- Cross out the ones you disagree with. "Too casual" might be exactly the voice you wanted.
- Hand back only the problems you want solved. Half the tokens, half the rework, and nothing you liked gets fixed.
Both doors. Paste a draft, or point Claude at the file.
List every problem you see with this. Number them. Do not fix anything yet. I will tell you which ones to fix. [paste the draft here, or name the file if I am working in a folder] 1. One line per problem, numbered, most serious first. Say what is wrong, not how you would fix it. 2. Wait for my list of numbers. 3. Fix only the ones I name, with the smallest change that solves each. Leave everything else exactly as it was, including anything you flagged that I did not pick. Report what you changed, by number. Then stop.
8 of 15Connectors first, then slim down
A connector is the one-click link to Shopify, Gmail or your calendar. A skill is a short written recipe Claude follows, in your words.
- Prototype with a connector. One click, sign in, test. Prove the idea works before you build anything.
- Once it works and matters, turn it into a lean skill. Only the steps you need. A connector carries every tool it has whether you need it or not, and slows every start-up.
- Prune the connector list. Nick saves seconds on every launch and a lot of attention in every chat.
Connectors live in both doors. A skill is a file, so it needs Claude Code or a desktop app chat opened in a folder.
9 of 15Run separate jobs at the same time
- Batch your asks up front. "Fix the hero, the form and the logout button" in one message, instead of three messages waiting on each other.
- Say "do these in parallel as mutually exclusive scoped tasks." Those exact words tell Claude to split the work so no two helpers touch the same thing.
- Add a merge step at the end. One pass that brings the pieces together.
- Accept a little more error for a lot more speed. For most work that is a good trade.
Claude Code runs the helpers side by side. In the desktop app the same words still work, but the tasks run one after another, so you save your own waiting time and nothing more.
Do these in parallel as mutually exclusive scoped tasks, then merge: 1. [task, naming the file or piece it may touch] 2. [task, naming a different file or piece] 3. [task, naming a different file or piece] Merge means: [what you want at the end, for example one page called merged.md that puts the three together in that order, or "check the three agree and fix any seam"]. Rules: no two tasks touch the same file. If a task finds something outside its own file that now contradicts it, leave that file alone and put it in the report. When all three are done, do the merge as one pass. Report the three results and what the merge changed, in five lines or fewer. Then stop.
10 of 15Start fresh with a handoff note
- Long chats collect contradictions. You said "not X" on turn three and "a bit like X" on turn twelve. Claude splits the difference, and the result is mush.
- Ask for the handoff. What is done, what decisions you need to make, what is next, what problems are open.
- Fix the summary yourself before you hand it over. This is your one chance to correct the record.
- Paste it into a new chat. Claude recognizes a handoff, asks about anything missing, and works from a short clean brief. Cheaper and sharper.
I turned this into a one-word command the day I watched the video. Both doors. If the new chat cannot open files, paste the whole summary in.
Summarize where we are: what is done, what decisions I still need to make, what is next, and what problems are open. Write it so a fresh chat can pick up without reading this one. 1. Go back to the start of this chat (or the start of the summary I pasted), not the last few turns. Use these five headings: Done (with file names or links), Decisions (two lists, Made and Open), Next (the first step concrete enough to start on with no question), Open problems (say "I assumed" wherever you guessed), Rules I stated in this chat (my corrections, and anything I said to leave alone). Where I changed my mind, keep only the final version. 2. Leave out every client or patient name and any health detail about a person. 3. Show me the five sections and ask me to correct anything wrong or missing. Wait. 4. After my corrections, save it as handoff-[today's date, written YYYY-MM-DD].md if you can write files, and print a short opener I can paste into a new chat, under 400 words, that begins "Picking up from a handoff". If you cannot write files, say so, and make the opener carry all five sections instead, under 600 words. Then stop. This chat is done.
11 of 15Ask side questions with /btw
- While Claude works, type
/btwand ask what it is doing. It answers without stopping the job, and without your question polluting the main task.
Claude Code only. In the desktop app, wait for the job to finish, or ask your side question in a second chat.
12 of 15Fan out, then fan in
- Cheap helpers scan widely, the strong model decides. Research is expensive when the top model does it. Let small, cheap helpers read the field and report back.
- Ask for it by name. More coverage, faster, cheaper.
Claude Code, where the helpers exist. The desktop app covers the angles one at a time.
Use a fan-out and fan-in strategy. Delegate to cheap subagents to find as much as possible about [the question, for example ways a small clinic can cut no-shows], including approaches most people don't use. Then combine it all into one brief and make the recommendation. 1. Split the field into at least four angles and send one helper per angle. Each helper reports a list of approaches, one line each, with a source or example where it can find one. Mark any source it could not check as unverified. 2. Combine everything into one brief: every approach found, grouped, with the three least common marked. 3. Make the call: one recommendation, why, and what it would take to try it within a week. Save the brief as [filename] and give me the recommendation in three lines. Then stop.
13 of 15Keep your CLAUDE.md honest
Your preferences file is read at the start of every chat. In Claude Code it is CLAUDE.md in the folder. In the desktop app it is a Project's instructions. Whatever it is called, it is the one file Claude believes without checking.
- Old facts make bad decisions. Nick left months-old revenue numbers in his and got advice for the smaller business he used to run.
- Audit it now. Every line either earns its place or goes. A character Claude does not need is attention wasted.
- The shape that works: what is where, your preferences, what it may do without asking, who you are, lessons learned.
- Re-read it every time the model updates. Yesterday's rules were written for yesterday's model.
Both doors. Mine gets a dated line at the top of every block of numbers now, so Claude can see how old they are.
14 of 15Keep a backup tool ready
- Have a second tool set up in the same folder. When one service has an outage, whole companies stop. Yours does not have to.
- Share the rules. A file called
AGENTS.mdthat mirrors yourCLAUDE.mdmeans the backup knows everything Claude knew.
For Claude Code users working in a folder. In the desktop app, keep the same preferences in another tool's project too.
15 of 15Learn from mistakes, in the positive
- Write the solution, not the ban. "Batch the file reads" teaches. "Don't retry the old API" only forbids.
- After any build, ask "how could you have done that faster with fewer tokens?" Then put the answer in your preferences file. Nick says this saves him three or four hours a week.
- Use
/insightsto see the patterns you would never notice yourself. Small conveniences compound. "Open the file when you are done" is a rule that pays every day.
/insights is Claude Code. In the desktop app, end the week by asking what you made Claude redo and what one rule would have prevented it. Prompt 6 works in both doors.
How could you have done that faster with fewer tokens? Give me one lesson, written as what to do next time, and add it to my preferences file. 1. Look back over [the whole build in this chat, or name the build] with evidence: a file read twice, a wrong path tried, a check skipped, rework after I corrected you. List up to five, one line each. The lesson may be for you, or for how I asked. 2. Pick the one most likely to come up again. Write it as a single instruction in the positive, under 30 words. "Batch the file reads", never "Don't retry". 3. Add that one line under a heading called Lessons in my preferences file ([CLAUDE.md, or the file where I keep how I like to work]). Change nothing else in that file. If you cannot write files, print the line and tell me to paste it into my Project instructions. Report the lesson and where you put it, in two lines. Then stop.
Three mistakes I made so you don't have to
- I trusted prompts after one good run. Three runs on the same example is my floor now, and a prompt that scores two out of three does not go to a client.
- I let the tool list grow without looking. One hundred plugins, 81 of them from a single click in June, and every chat all summer paid for them. Move 16 takes ten seconds.
- I pushed long chats instead of handing off. A handoff note and a fresh chat take two minutes. I spent months not knowing that.
If you want me beside you while you put these to work, start with the free audit at The Bottom Line or write to me at kristen@blueberrytherapy.ca.