You installed Claude. You set up a workspace. Now: how do you actually work with it day-to-day without the conversation getting messy, the answers getting generic, or the corrections evaporating?
What this episode covers
Most people use Claude wrong because they treat every conversation the same way. The pros switch modes deliberately, prune context aggressively, and make every correction stick.
Think in Chat. Operate in Cowork. Build in Code.
Inside the Claude desktop app, three tabs are three different mental modes. The difference between a frustrating conversation and a great one is often just picking the right one.
Conversations, brainstorms, drafts. No file access. Fast.
Use when: You want to talk something through, sharpen an idea, or get a one-off draft.
Background agent. Runs longer tasks while you do other things. Can schedule.
Use when: "Research these 5 competitors and produce a brief." Anything 10+ min that you'd happily walk away from.
Folder access, persistent memory, skills, your CLAUDE.md. The workspace mode.
Use when: The work needs context about you, your brand, your projects.
The clever-phrasing era is over. Context engineering wins.
Modern Claude doesn't need tricks. It needs you to be specific about what you want, what "done" looks like, and what context to pull from. Six patterns that experienced users run constantly:
Asking "what should I write about X?" produces generic LLM output. Pasting a rough draft and asking Claude to react, sharpen, or list missing angles produces specific output. Claude gets much better when it has something to push against.
The single most reliable way to get the format you want: show two or three examples of exactly what you mean. "Like this, this, and this. Now do four more in the same style."
"You are a senior brand strategist reviewing copy for a non-technical founder. Be direct. Flag tone drift before grammar." This single sentence reframes the whole response register.
Before a multi-step ask, hand the steering wheel to Claude:
I want to plan our Q2 content calendar. First, read my brand voice file and my last 10 posts. Then ask me 4-5 questions before drafting anything. We'll refine together before you produce the calendar.
This costs nothing. It saves you from generic output.
Use @filename to attach a file rather than explaining what's in it. Drag images directly into the prompt. Paste URLs. Claude reads them before responding. Describing the contents is just adding noise.
Long conversations get worse. Here's why, and what to do about it.
Context rot is real. Even when Claude says it has 200,000 or 1,000,000 tokens of memory, the practical degradation starts much earlier. Power users report quality drops around 40-50% utilization, and clear problems by 60-65%. Translation: rotate your conversation at 60% capacity, not when Claude tells you it's full.
Click the model indicator to see your usage. The bigger insight isn't your billing. It's an inverse signal of conversation health. A long-running session sitting at 60%+ has already started losing quality.
/clear/compact/compact preserve the brand voice rules and the 3 decisions we made). Use at phase boundaries, not at the cliff.Before clearing a long session you might want to revisit, ask Claude to write a handoff:
Before I /clear, write a one-page handoff in .memory/Inbox/ called session-handoff-[date].md covering: what we accomplished, what worked, key decisions, what's left to do.
Next session: read .memory/Inbox/session-handoff-{date}.md and pick up where we left off. You get a fresh, fast Claude that knows everything important.
Decide how much rope to give Claude. Different work calls for different leashes.
Claude Code has four permission modes. Click the mode indicator at the bottom of the input bar to switch:
| Mode | What happens | When to use |
|---|---|---|
| Ask permissions | Claude asks before every tool/file/command action | The default. Safe and visible. |
| Accept edits | Auto-approves edits to existing files; still asks for new ones | When you trust the change but want to keep an eye on new files appearing |
| Plan mode | Claude reads, thinks, drafts a plan, waits for approval before doing anything | Multi-step work, anything risky, anything you'd want to course-correct |
| Auto mode | Runs continuously, makes reasonable assumptions, only stops on destructive things | Long-running, low-stakes work where you want to walk away |
For non-developers, plan mode is the single most useful safety net. Before Claude does anything substantive, it shows you the plan. You read it, edit if needed, then approve.
I'll draft the Q2 campaign in 3 steps:
When Claude is done planning, it shows the proposal with three options: Revise (push back on the plan), Accept (run with manual confirmation per step), or Accept and auto mode (run autonomously). For most non-trivial work, Accept is the right move. You stay in the loop without watching every keystroke.
Use it for:
Skip it when the change is one sentence: a typo fix, a quick reword, a single email draft.
A correction that evaporates is a correction you'll repeat next week. Here's how to make them permanent.
Claude has automatic memory. When you correct it, it usually remembers across sessions. Usually. The difference between a correction that sticks and one that evaporates is mostly in how you phrase it.
Wrap corrections in language that signals this is a rule, not a one-off:
Tell Claude where the correction belongs in your memory:
That's a brand voice rule — please add it to my principles file.
That's a factual correction about my company — update identity.md.
Naming the file makes Claude actually go write to it instead of just nodding.
Auto-memory isn't perfect. After meaningful corrections, verify:
What did you remember from yesterday about how I want LinkedIn posts written?
If the answer is wrong or thin, ask Claude to fix the relevant file directly.
The most underrated feature for marketers and operators.
TodoWrite is a built-in checklist tool. You'd think it's for coders, but the best use is non-coding multi-step work. It eliminates the blank-page problem and makes Claude's progress visible.
A real example: ask Claude to plan a launch using TodoWrite. You get a structured checklist with tracks and deliverables, instead of a wall of paragraphs.
Plan a 4-post LinkedIn series on the launch. Use TodoWrite — one todo per post. Each todo: hook, thesis, CTA, and what asset I'll need to attach.
We're shipping the Q2 campaign. Use TodoWrite to break it into work tracks (copy, design, email, social, PR). For each track, list the deliverables and what "done" means.
Read this week's Inbox/ entries and any modified project files. Use TodoWrite to list everything I owe a decision on, with the context for each.
In all three cases, the magic is the same: Claude turns "where do I start?" into a numbered list with progress visible. You watch checkboxes tick off as it works.
Five minutes a week keeps the system sharp.
/clear when switching./rename or use the Code sidebar)..memory//clearOnce a week, run this prompt:
Review my CLAUDE.md and all .memory/ files. Flag: - Entries older than 14 days that may be stale - Duplicates or near-duplicates - Anything contradicting more recent entries - Anything in Inbox/ that should be moved or deleted List your findings, then ask before changing anything.
Five minutes. Compounds enormously. Most people skip this and wonder why their Claude gets dumber over time.
.memory/ like a working knowledge base: prune ruthlessly, structure deliberately, and date-stamp anything decay-prone.
Real patterns from people who use Claude as their daily operator.
The Routines panel (Code tab → Routines) is where scheduled work lives. Set it up once, it runs forever: every morning, every Monday, every weekday at 8pm. Local routines run when your computer is awake; cloud routines run regardless.
Cowork visits 5 competitor sites + your relevant subscriptions, produces a one-pager: changes since last week, pricing moves, messaging shifts. Save 8-10 hours/month.
Every Monday at 7am, visit [competitor1.com], [competitor2.com], [etc.]. Compare to the snapshots in /competitors/baseline/. Output a Markdown brief in /reports/YYYY-MM-DD-competitive.md covering: changes detected, pricing updates, messaging shifts, and 2-3 implications for our positioning.
Pulls your analytics, GSC, and email metrics. Produces a one-page trend report with 3 prioritized actions.
Check this draft against my brand voice file. Flag deviations by severity (high/medium/low). Don't rewrite — just diagnose.
The diagnosis-first version is more useful than auto-rewriting because you stay in control of voice.
Take this blog post and produce: 1 LinkedIn post in my voice, 2 email teaser snippets, 3 IG-caption variations, and a Slack summary for the team. Run all of it through my brand voice file.
I have a meeting with [name] from [company] tomorrow. Read their last 10 LinkedIn posts and their company's about page. Suggest: 3 talking points specific to them, 2 questions I should ask, and one risk I should be ready to address.
The common mistakes that experienced users have already paid for.
Mixing unrelated tasks in one conversation. Quality decays fast. Fix: /clear between unrelated work.
After two failed corrections, the session is poisoned. Fix: /clear and write a better initial prompt that incorporates what you learned.
Bloated instructions cause Claude to ignore them. Fix: Ask of every line: "would removing this cause a mistake?" If not, cut.
"Click X, then Y, then Z." Fix: Tell Claude what done looks like. Let it figure out the steps.
Cowork can drive your browser. It's the most fragile part of the stack. Fix: Always verify outputs from screen-pulled data.
Resets all the learning. Fix: Use memory, lessons files, and handoff docs.
Hallucinations rise; instructions get buried. Fix: /compact at 60%, /clear at 70%.
Causes generic output regardless of how clever your prompt is. Fix: Build it once, reference it always.
New practical guides for European SMEs adopting AI, delivered when they drop. No spam, unsubscribe anytime.
Want this done for you, not just explained? Vectimo runs the AI adoption end to end.
Work with Vectimo →