“A house assistant who remembers everything is useful, until it starts bringing up your salary in front of guests.”
1. Memory: scope, exclusions, and the "forget this" command
Where: Settings → Capabilities → Memory. What it does: memory stores what Claude judges worth keeping across chats. Three controls most people miss: per-project scope, an exclusion list, and an on-the-fly forget command.
Why it matters: default memory drifts. After a few weeks it fills with one-off corrections ("I prefer tabs in Python," said once, about one file), project facts that leak into unrelated chats, and stale role context. Output quality slips because Claude is optimising for the wrong "you."
- Turn on project-scoped memory so memories formed inside a Project stay in that Project; this alone fixes most drift.
- Add anything you never want surfacing elsewhere to the exclusion list (medical, legal, salary, client names); these persist regardless of context unless excluded.
- Use the inline command in any chat ("forget what you remembered about [topic]") and Claude confirms what it removed. No menu needed.
2. Extended Thinking: the per-chat toggle
Where: the model selector in the chat composer (Off / Light / Full). What it does: adds a reasoning pass before the answer; the per-chat toggle overrides the global default.
Why it matters: thinking helps on math, debugging, and multi-step planning. It's wasted on summarising, translating, formatting, and quick lookups, where it adds latency and tokens for the same answer. The article reports dropping Opus token spend ~18–25% in week one by defaulting thinking to Light.
- Set the default to Light (thinks only when the model judges it useful).
- Switch to Full explicitly for genuinely hard work.
3. Custom Styles: output contracts, not just "voice"
Where: the style selector in the composer → Create New Style. What it does: a style is more than a tone toggle. It's an instruction file (a couple hundred to ~1,500 words) applied before every response in that style.
Why it matters: the real use isn't "make it shorter." It's enforcing structure across every response without re-pasting it: citation format, banned words, required sections, code-fence languages, length caps, whether to ask follow-ups.
- Create one style per recurring workflow (e.g. "Draft for X," "Code review," "Summarise PDF").
- A handful of these in rotation replaces most of your saved prompts.
4. Projects: the "instructions" field most people leave empty
Where: any Project → ⋯ → Edit project instructions. What it does: the instructions field is a system-prompt equivalent injected into every chat in that Project. Most people remember the knowledge-upload feature on the same page and leave instructions blank.
Why it matters: without it, every chat in the Project starts cold and you re-establish context each time. With it, you set role, default skepticism, formatting, and hard "never do this" rules once.
- Treat it like a CLAUDE.md for Claude: role, default skepticism level, formatting rules, what to never do.
- Keep it under ~400 words, and prune it about once a month.
5. Search past chats (Pro+)
Where: Settings → Profile → Search past chats. What it does: lets Claude search your conversation history when relevant. Why it matters: it's off by default for new accounts, and, per the article, it does keyword matching, not semantic search. "That thing we discussed last week" returns nothing; the literal nouns from the old chat do.
- Turn it on.
- Query with content nouns, not meta-words: "Polymarket Iran" works; "that thing we talked about" doesn't.
6. Web search: per-conversation toggle and citation behaviour
Where: the + menu in the composer toggles web search per conversation; Settings → Capabilities → Web search controls citation style (Inline / Footnotes / Hidden). What it does: the citation setting changes how sources appear in the answer.
Why it matters: the default Inline citations break copy-paste. Paste the answer elsewhere and you carry markers that point nowhere. Footnotes keeps the body clean and lists sources at the end.
- Switch to Footnotes if you ever copy web-search answers into another doc, email, or message.
- Keep Inline only if you read everything inside Claude.
7. Connectors: the "trust this folder" trap
Where: Settings → Connectors → trusted folders (for the local-folder access feature the article calls "Cowork"). What it does: once a folder is trusted, Claude reads from it every session without prompting.
- Open the trusted-folders list and remove anything that isn't a current, active project.
8. Incognito mode and what it actually skips
Where: sidebar → New incognito chat (Cmd/Ctrl + Shift + N). What it does: an incognito chat is deleted on close. Why it matters: people assume it just hides the chat from the sidebar. Per the article it actually skips four systems: memory write, chat history, past-chats search index, and training opt-in (if enabled).
- Use it deliberately for anything sensitive: salary, medical, family, legal drafts, client names. Three keystrokes, no thinking.
In one line each
- Memory: scope it per-project, exclude sensitive topics, and use the inline forget command to stop drift.
- Set Extended Thinking to Light by default and Full only for hard work; styles and project instructions enforce structure without re-pasting.
- Turn on past-chat search and query with content nouns; switch web citations to Footnotes if you copy answers out.
- Audit trusted folders, and use incognito for sensitive data: it skips memory, history, search index, and training.
Where to go next