I’ve been using Claude daily since late 2025 — building sites, running newsletters, managing projects.
A friend just started using it and asked for advice. I realised the useful stuff isn’t “write better prompts.” It’s the structural habits that took me months to figure out.
Here’s what I’d tell myself on day one.
1. Talk to Your Wrench
I had coffee with a friend a while back who needed a better way to track his calls and to-do items. He had access to Microsoft Copilot but was still doing everything with spreadsheets and memory.
My advice wasn’t about prompting. It was: describe your problem to the tool. Tell it what the best outcome looks like. Tell it what software you already have. Then don’t trust the first answer blindly — challenge it, verify it, repeat.
He looked at me like I’d suggested talking to a wrench. Which I suppose I had. The wrench just talks back now.
We’ve never been at a point where you could ask a tool what it does, where it struggles, and what it needs from you to work well. But you can now. Try it: “What are the things you’re most likely to get wrong? What should I always double-check?” You’ll get a more honest answer than most software documentation.
That back-and-forth — between you and the tool, about the tool itself — is completely new. I skipped it for months.
2. You Have to Tell It What You’re Working With
Before you ask Claude to help with anything, take stock — what apps, platforms, and accounts do you actually use day to day?
Claude can connect your tools in ways you wouldn’t think of — but only if it knows they exist. I didn’t mention that I had a Kit.com account for newsletters, an Obsidian vault for notes, and a website until weeks in. Once I did, it started connecting them. “You have a newsletter platform and a vault full of content — want me to help you build a pipeline between them?”
Write a quick inventory. Doesn’t need to be exhaustive — just your tools, your accounts, what you’re already paying for. Drop it into a Project or paste it at the start of a conversation. The connections Claude makes from that inventory are worth the ten minutes it takes to write it.
3. Plan Before You Build
The instinct is to jump straight to “build the thing.” Fight it.
Walk Claude through your expected outcome first. What should this look like when it’s done? What are the steps? What could go wrong? Have it challenge your assumptions — “Do you actually need a spreadsheet for this, or would a simple list work?”
I wasted entire conversations rebuilding things that a fifteen-minute planning chat would have caught. Claude is a better thought partner than executor. Use that.
The pattern that works: describe the outcome, let Claude propose approaches with trade-offs, pick one together, then break it into steps. Each step gets done and checked before moving to the next. It takes longer to start. It takes far less time overall.
4. One Hour In, It’s Already Losing the Thread
Long conversations drift. I watched it happen repeatedly before I figured out what was causing it.
In a fresh conversation, Claude is sharp. It follows your instructions, stays on task, does what you asked. An hour in, with a sprawling context full of tangents and corrections and half-finished ideas, it starts losing the thread. It’ll wander off your question, pick up something tangential, and explore that instead. Your original ask never gets answered.
The fix: shorter conversations with clear goals. One task per conversation when you can manage it — start, do the thing, save what you learned, move on. Next task gets a fresh start with fresh context.
This also protects you from a subtler problem — context fatigue makes Claude more likely to fabricate, more likely to drift, and less likely to follow your instructions. A clean start costs you thirty seconds. A drifted conversation costs you an hour of unpicking things that shouldn’t have gone wrong.
5. Make It Run Post-Mortems on Itself
This is the single biggest thing I wish I’d started earlier.
When something goes wrong — Claude fabricates a detail, drifts off your task, gives you confidently wrong advice — don’t just correct it and move on. Have Claude analyse its own behaviour. What happened? Why? What instruction would have prevented it?
Then save that instruction. Not a vague principle like “be more careful.” Something specific: “When I ask you to send something, show me exactly what you’re about to send and wait for my go-ahead.”
Every post-mortem makes the next conversation better. I keep a running document of these rules now, and every line traces to a specific failure. The document isn’t a prompt — it’s scar tissue. It grows from use, not planning.
6. The Conversation Vanishes. The File Doesn’t.
Claude’s memory within a conversation is temporary. When the chat ends, everything in it vanishes. If you let Claude manage everything inside the conversation, you lose it all.
Get in the habit early: important decisions, research findings, plans — all of it goes to a document. Not because Claude can’t hold them — because you need them next Tuesday when you’re in a different conversation with zero context.
I use a simple pattern: anything I’d want to reference later, I ask Claude to write up in a file I can save. Project decisions, research summaries, step-by-step plans. Next conversation, I drop that file in and Claude picks up where the last one left off — but only because the breadcrumbs exist outside the conversation.
This ties back to shorter conversations too. If every conversation produces a saved summary, the next one can start clean without losing anything. The documents become the continuity that Claude’s memory can’t provide.
7. It Will Confidently Make Things Up
This catches everyone. Claude doesn’t hedge when it’s wrong. It doesn’t say “I think your URL might be…” It says “your profile at linkedin.com/jimchristian” with complete confidence — except that’s not my handle.
URLs, dates, stats, job titles, how a specific app works. If Claude didn’t look it up in this conversation, treat it as a guess. Especially when it sounds right — plausible is the failure mode’s natural camouflage.
The fix: tell Claude to verify claims before stating them. “Don’t guess URLs — ask me or search for it.” “Don’t cite numbers from memory — check the source.” You can’t prevent the tendency, but you can add a speed bump that catches most of it.
8. Build Guardrails to Protect Yourself From Yourself
A lot of the mistakes aren’t Claude’s fault. They’re yours, amplified.
You’ll approve an email without reading what it actually says, say “looks good” to a draft without checking the facts, and let Claude do something that can’t be undone because you didn’t ask what the default behaviour was.
I asked Claude to send a newsletter digest once. The default was “all subscribers.” The digest went to my entire mailing list instead of the intended audience. You can’t unsend an email.
The solution is habits that slow you down at the right moments. Tell Claude: “For anything that can’t be undone — sending emails, publishing content, deleting files — show me exactly what you’re about to do, explain it, and wait for my okay.” Build a review step before anything goes out.
The speed of AI makes it easy to skip the thinking part. Build friction into the moments that matter. Future you will be grateful.
9. Don’t Let It Review Its Own Work
When Claude writes something and you ask it to review what it just wrote, it will approve it. Every time. The same conversation that produced the problems will rubber-stamp right past them.
This is the same reason you don’t proofread your own novel — but with AI the temptation is stronger because the review feels instant and thorough. It’s neither.
The fix: start a new conversation for reviews. Fresh context spots what the writing conversation can’t see. Or just review it yourself — print it out, read it cold. But don’t ask the writer to be the editor. It doesn’t work for humans and it doesn’t work for AI.
10. The System Compounds. Invest in It.
The first conversation with Claude feels like using a smart assistant. The fiftieth — if you’ve been building the system around it — feels like working with someone who knows how you think.
Not because Claude got smarter. Because you taught it. Your saved instructions have guardrails from post-mortems, your reference documents have your tools and preferences, and your plans mean new conversations start with context instead of from scratch. The wrench remembers every conversation you’ve had about how to hold it.
Each thing you save — a post-mortem rule, a tool inventory, a planning template, a review checklist — makes every future conversation faster and safer. The first week is slow. You’re building the foundation, not the output. But by week three, things that used to take an hour happen in minutes.
That’s the part I wish someone had told me. Claude isn’t a tool you use — it’s a tool you teach. The individual outputs are useful, but the system that produces them is everything.
Start with a short document of instructions. Grow it from there.
Keep Going
If this was useful, here’s where to find more:
Signal Over Noise — my weekly newsletter on AI integration. Not hype, not news roundups — tested workflows and honest assessments of what actually works. The “Talk to Your Wrench” concept came from Issue V2-06.
Second Brain Chronicles — a companion newsletter where I document the system itself. The post-mortems, the failures, the setup that makes all of this work. If you liked section 5 and 10, this is where that thinking lives week to week.
LinkedIn — where I share shorter takes on AI practice and integration.
I’m genuinely curious what your first month with Claude looked like — find me on LinkedIn and let me know.