Skip to content

Claude: Intermediate

Intermediate

You can get a long way with single chats, but the real shift is giving Claude lasting context so you stop re-explaining yourself. Three features do most of the work: Projects, custom instructions, and Memory. This guide wires them into a workflow.

A Project bundles a knowledge base (uploaded files, text, code) with its own instructions, shared across every chat inside it. Make a Project per client or topic, load its reference material once, and every chat in it starts already grounded.

A clean pattern that mirrors a reliable workflow:

  1. Create a Project and add its knowledge base: briefs, style guides, past work, the documents you keep re-pasting.
  2. Use one chat to plan: research, outline, agree the approach.
  3. Use a separate chat to execute: draft, build, produce the output.

Keeping planning and doing in different chats stops a single thread from sprawling. There’s no formal plan mode in the chat app, so you create the separation yourself.

Two layers of persistent instructions, with no file to manage:

  • Profile preferences (account-wide). Open Settings (your initials, lower-left) > Profile > “What preferences should Claude consider in responses?” These load into every new chat on every plan. Put your role, domain, default language (Swiss German vs English), and preferred output format here.
  • Project instructions (scoped). Layered on top of your profile, applying only inside that Project.

Skills add a third, on-demand instruction layer you invoke per message. Anthropic is folding the old custom Styles into Skills; the migration is still rolling out as of mid-2026, so you may see both surfaces for a while, and a migrated style is called with a /slash command like any other Skill. There’s more on building and installing Skills at the Advanced level.

LayerScopeWhen it loadsBest for
Profile preferencesAccount-wide, every chatAutomatically, at the start of every new chatYour role, default language, preferred output formats
Project instructionsOne Project, layered on top of your profileAutomatically, in every chat inside that ProjectClient context, the house style and rules of one engagement
SkillsOne task or messageOn demand: you invoke it, or Claude picks it when it fitsRepeatable procedures and formats: report templates, checklists, a house tone

Worked example. A consultant at a Zurich agency puts “I’m a marketing consultant; answer in English; keep outputs under 300 words unless asked” in her Profile preferences. Her Muster AG Project adds “Muster AG communicates in formal Swiss High German (Sie-Form); never promise delivery dates”. When the monthly report is due, she calls a /monthly-report Skill that carries the exact section structure. Each layer stays small because it only holds what the broader layers can’t.

Claude has a genuine auto-memory. It synthesises key points from your chat history (your role, projects, and working preferences) and reuses them in later chats.

  • It’s available on all plans and on by default, refreshed automatically in the background.
  • Each Project has its own separate memory, isolated from other Projects and from general chat. This surprises people: context does not cross Project boundaries by design.
  • Manage it at Settings > Capabilities (“Manage memory”): toggle it on or off, Pause it (keep what exists, stop adding), or Reset it (permanently deletes everything, including project memories).
  • You can tell Claude what to remember mid-chat, and the edit applies immediately.
  • On paid plans, Chat search retrieves detail from your past conversations on demand, handy when you know you discussed something but not in which chat.
  • Incognito chats are excluded from memory and history: use them for throwaway or sensitive one-offs.

There is no /clear or /compact in the chat app. Your levers for a clean context window are:

  • A new chat per task: the single biggest lever.
  • Separate Projects to wall off unrelated knowledge bases.
  • Incognito for one-offs you don’t want remembered.

When a chat nears its window limit, Claude auto-summarises earlier messages so it can keep going, but that silently drops detail. A chat that “continues forever” is not preserving everything. For accuracy-critical work, start fresh and re-attach the source rather than correcting a tired thread. As a rule of thumb: if two corrections in a row don’t land, start over.

  • ✅ You’ve set up a Project with a knowledge base and instructions
  • ✅ Your Profile preferences carry your role and defaults into every chat
  • ✅ You can find, pause, and reset Memory, and you know it’s per-project
  • ✅ You start a new chat instead of nursing a long one

Ready to connect your own tools and data, build Skills, and lock down your data controls?