Source: Anthropic Academy — Claude 101 — “Organizing Your Work and Knowledge” section (Projects, Artifacts, Skills) + April 2026 Memory GA release
Time: Read 4 min | Watch/click-through 15 min | Practice 10 min
The Three Surfaces (in increasing explicitness)
Claude has three ways to keep context across work. Use them in this order from “automatic” to “deliberate”:
- Memory (auto, always-on unless disabled) — Claude remembers facts about you and your work across separate chats.
- Projects (explicit, scoped) — a persistent workspace with custom instructions + uploaded knowledge files. You decide what goes in.
- Artifacts (ephemeral, per-chat) — the right-hand pane Claude uses for structured output (tables, docs, code, mockups) that you can iterate on without losing chat context.
The Claude 101 course goes deep on Projects + Artifacts. Memory is newer (April 2026 GA) and the WEO overlay below covers what matters for client work.
Watch / Click-Through First
Claude 101 — “Organizing Your Work and Knowledge” section (~15 min)
Direct: anthropic.skilljar.com/claude-101
What’s in it:
- Introduction to Projects — persistent workspaces with knowledge files and custom instructions.
- Creating with Artifacts — the right-hand side panel where Claude writes code, docs, tables, and visual outputs you can iterate on.
- Working with Skills — a preview (we go deeper on Skills in Module 6).
Anthropic’s course covers the mechanics. The WEO-specific overlay below covers how to use this at WEO — client-per-project pattern, what to load, what NOT to load.
Why It Matters at WEO
The single biggest efficiency gain from Claude for account managers and content strategists is Projects — specifically a project per client. Instead of re-pasting the brand guide, target audience, voice notes, prior campaigns, and “remember we don’t say X” every chat, you load all of that once into a Project and every chat in that project starts with the full context already loaded. This is the difference between Claude saving you 30 min/week and Claude saving you 5 hrs/week.
Memory (the new surface) compounds Projects: Claude remembers your role, your working style, the three clients you ask about most, and the kinds of output you reject. By week two, your average prompt gets better without you trying.
WEO-Specific Overlay — Memory (the new default)
Memory launched as a Max-plan preview in early 2026 and went GA on Free / Pro / Team / Enterprise in April 2026. It’s on by default. Key mental model:
- What Claude remembers: facts you’ve told it (your role, your company, preferences), patterns it’s observed (you always ask for bullets, you hate the word “leverage”), and persistent context you’d otherwise re-type.
- What Claude does NOT remember: full conversations, attached files, client data from Projects (Project knowledge stays scoped to that Project).
- Where to see it: claude.ai → your avatar → Settings → Memory. It’s a plain-text list. Edit or delete any row.
- Where to turn it off: same pane, toggle at top.
Three WEO patterns with Memory:
- Let Memory learn your role on day one. Spend 10 minutes early on telling Claude things like “I’m an SEO strategist at a dental marketing agency. I usually want output in markdown tables. I prefer tight copy — no ‘in today’s fast-paced world.‘” It will remember and apply this.
- Prune weekly. Every Friday, skim Memory. Delete anything stale (old client names you’ve moved off, preferences that have evolved).
- Don’t let Memory swallow client specifics. If Claude tries to remember “Client X’s unannounced campaign is Y,” delete it. Client-specific knowledge belongs in that client’s Project, not in cross-chat Memory that could surface in unrelated work.
Memory governance (what NOT to let Claude remember)
- Client names tied to unreleased strategy. “I’m working on Smile Springs’ rebrand” — fine if already public, delete if still confidential.
- Patient-adjacent info. “Dr. Jones at Practice X has been asking about…” — delete.
- Internal compensation, pricing, or personnel info. Never. Memory surfaces across chats and you won’t always remember what it remembered.
- Anything classified Level 3 or 4 under the WEO data-classification model. See Module 7.
Rule of thumb: Memory is your scratchpad, not the agency’s filing cabinet. Projects are the filing cabinet.
WEO-Specific Overlay — The Client-Per-Project Pattern
For ongoing client work, use one Claude Project per major client. Inside each project, load these knowledge files:
- Brand guide — voice, tone, banned words, brand personality.
- Service offerings — what this client actually sells, in their words.
- Target audience doc — personas, demographics, psychographics.
- Prior-year campaign brief — what ran, what worked, what didn’t.
- Recent blog titles / meta descriptions — so Claude doesn’t repeat topics or voice you’ve already published.
- Logo + brand assets (optional but helpful if you use Claude Design in the same project).
Then in the custom instructions (a.k.a. “system prompt”) for that project, write something like:
You are a senior dental marketing copywriter working on [CLIENT NAME]. Always reference the attached brand guide for voice. Always check the attached prior-campaign brief before proposing new angles. Flag any claim that needs clinical verification by wrapping it in [VERIFY: …]. Never invent statistics — say “data not available” instead.
Now every new chat in this project has all of that context without you re-explaining. You can focus on the actual task.
WEO-Specific Overlay — What NOT to Load
Do not load any of this into a Claude Project, even on the Team plan:
- Real patient records, patient names, or PII.
- Client contracts, payment terms, internal pricing.
- Unreleased strategic decisions (mergers, personnel, layoff planning).
- Anything marked “Confidential” or “Internal Only” that hasn’t been cleared by the AI council.
Module 7 has the full rules. The short version: if you wouldn’t email it to a vendor, don’t put it in Claude.
WEO-Specific Overlay — Artifacts
Artifacts are Claude’s “right-hand side panel.” When you ask Claude to produce something structured — a table, a doc, a piece of code, a landing page mockup — Claude opens an Artifact pane where you can watch it being written, edit it, and iterate without losing the conversation context.
Three WEO patterns that are gold with Artifacts:
- Campaign brief generator. Ask Claude to produce a campaign brief in an Artifact. Edit directly in the Artifact. Export the final version.
- Meta description table. Ask Claude to produce a table of 10 meta descriptions with character counts. Artifact makes it a live table you can copy row-by-row.
- Hero-section mockup. Ask Claude to produce a simple HTML mockup of a hero section. Review in the Artifact preview. Share with design.
Worked Example — Setting Up “Smile Springs Family Dental” as a Project
Concrete walkthrough of the client-per-project pattern with a fictional client. This is what a fully-loaded Project looks like on day one — copy the shape for real clients.
The client (fictional)
- Name: Smile Springs Family Dental
- Location: Columbus, OH
- Services: General, pediatric, Invisalign, implants, emergency, cosmetic
- Brand voice: Warm, plainspoken, trustworthy. Differentiator: Saturday appointments, no-wait booking, family-first.
- Primary audience: Families with kids and adults 35–55 in the Columbus metro.
- WEO engagement: SEO + content + paid search, started Jan 2026.
What goes in the Project
Five knowledge files, loaded once. Every chat in this Project reads from them automatically.
File 1 — brand-guide.md (~400 words)
# Smile Springs Family Dental — Brand Guide
## Voice
Warm, plainspoken, trustworthy. Think "the neighbor who happens to be a
dentist" — not "the brand that markets to you." Confident about expertise,
never clinical or cold in delivery.
## Must-haves
- "Saturday appointments" as a differentiator any time we talk scheduling.
- "No-wait booking" when talking about the in-office experience.
- "Family-first" as a frequent framing when the service applies to families.
- Real human names in testimonials (always fictional or cleared with client).
## Banned phrases
- "State-of-the-art"
- "World-class"
- "Commitment to excellence"
- "In today's fast-paced world"
- "Your smile is our priority"
- Em dashes in marketing copy (en dashes OK in clinical content)
- "Revolutionary" / "game-changing" / "cutting-edge"
## Banned moves
- Never imply clinical outcomes without a clinician-verified claim.
- Never use stock-photo-looking AI imagery of patients.
- Never use "limited time only" or fake urgency.
- Never promise outcomes ("whiter in one visit" → "whiter in many cases").
## Tone calibration
Sentence length: short to medium. Vary. No long corporate paragraphs.
Contractions are fine. Occasional warmth ("let's get you in") is on-brand.
Professional but not stuffy.
File 2 — services.md — the practice’s services in the practice’s own words, lifted from their existing website.
File 3 — audience-personas.md — three personas: “Mom booking for the family” (primary), “Empty-nest parent restoring after decades of neglect” (secondary), “Young professional in 30s getting back on routine” (tertiary). Each with age, common concerns, information-seeking patterns, common objections.
File 4 — prior-year-campaign-brief.md — what WEO ran for Smile Springs in 2025. What the KPIs were. What worked (pediatric content + Saturday-appointment messaging). What didn’t (cosmetic-focused ads for the 35–55 segment).
File 5 — recent-content-inventory.md — titles + brief summaries of the last 6 months of blog posts, so Claude doesn’t propose topics already covered or contradict angles already taken.
The custom instructions (Project-level system prompt)
This is what Claude reads at the start of every chat in this Project, before you even say anything:
You are a senior dental marketing copywriter working on Smile Springs
Family Dental (Columbus, OH). WEO is their content + SEO + paid search
partner since January 2026.
Default behaviors for every response in this project:
1. Always reference attached brand-guide.md for voice. Never use
banned phrases or banned moves listed there.
2. Always check recent-content-inventory.md before proposing new
topics — don't repeat covered ground or contradict published angles.
3. Flag any clinical claim that needs verification by wrapping it
in [VERIFY: specific claim]. Do not include unverified clinical
claims in any deliverable.
4. Never invent statistics. If you want to cite a number, say
"data not available" instead unless I've provided the number in
this chat.
5. If a request would violate brand guide constraints, call it out
and propose an on-brand alternative. Don't silently comply.
6. Default deliverable format: markdown, scannable, short sentences.
7. At the end of every response, include a one-line summary of what
you just produced.
I'll give you specific tasks in each chat. Treat the above as
standing rules.
The first three chats (what this looks like in practice)
Chat 1 — “Q2 blog content plan.”
You open a new chat in the Smile Springs project. You type:
“Produce 12 blog post ideas for Q2. Audience-balanced across our three personas. Include one evergreen SEO-play per month and one timely/seasonal angle per month. For each: working title, target keyword, primary persona, estimated impact (1–5), suggested publish order.”
Claude already has the brand guide, the services, the personas, and the prior-year history. It produces a ranked list of 12 ideas that (a) don’t repeat last year’s content, (b) fit the three personas, (c) reference specific services, and (d) flag any claim as [VERIFY] where clinical accuracy matters.
Chat 2 — “Draft the hero section for the Invisalign service page.”
New chat. Same Project. You type:
“Draft the hero section for the Invisalign service page. Include: headline, 2-sentence subhead, 3 benefit-led bullets, primary CTA text, secondary CTA text. Audience: persona 1 (mom booking for family) with Invisalign for herself.”
Claude writes hero copy in Smile Springs voice (warm, plainspoken, no banned phrases), references “Saturday appointments” and “no-wait booking” as the differentiators (because the brand guide said to), flags the Invisalign timeline claim as [VERIFY: typical treatment duration], and formats as markdown. You didn’t have to re-specify anything.
Chat 3 — “Reply to client email about adding emergency-services page.”
New chat. Same Project. You paste the client’s email verbatim. You type:
“Draft a reply. Confirm we can add an emergency-services page to the April sprint. Note the three things we’d need from them (practice emergency hours, after-hours contact policy, any insurance nuances). Warm, brief, professional.”
Claude writes the reply in a voice that matches the Smile Springs relationship (warm, not corporate), references the specific campaign sprint (because it read the prior-year brief and saw the sprint cadence), and flags the emergency-hours question sensitively (because it’s aware the topic is clinical).
What would have happened WITHOUT a Project
Every one of those three tasks would require you to re-paste the brand guide, the banned phrases, the audience personas, the differentiators, and the prior-year context. That’s 5–10 minutes of setup per chat. Project setup is a one-time 20-minute cost that saves that 5–10 minutes, every chat, forever.
Advanced Project Patterns
Branching chats within a Project
One Project can hold many parallel workstreams. Keep chats task-scoped, not session-scoped:
- One chat per deliverable. The hero-section chat stays focused on hero-section iteration. The meta-description chat stays focused on metas. Don’t mix.
- One chat per strategy conversation. Longer-running “planning” chats are fine — the quarterly content plan gets one chat that evolves as you refine.
- Name chats descriptively. Claude.ai auto-names chats based on opening message; rename them for clarity if you’ll revisit.
A well-organized Smile Springs Project after 3 months looks like 30–50 chats: 4–5 ongoing planning chats, 30+ deliverable chats (most “done” after a few turns), and 5–10 client-communication chats.
Handing off a Project to a teammate
When an account manager transfers a client to another AM, the Project is the handoff asset. The receiving AM:
- Gets added to the Project (Team plan — Project access is per-seat).
- Reads the custom instructions to understand the standing rules.
- Skims the last 5–10 chats to see active workstreams.
- Asks Claude for a summary: “Produce a status brief of this client based on the Project knowledge files + the last 10 chats. What’s in progress, what was recently completed, what’s coming up, any open issues I should know about.”
Claude produces a clean handoff brief. The receiving AM is productive in 30 minutes instead of 3 days of reading old email threads.
Archive vs. delete
- Archive — client is inactive but might come back, or you want the history as a reference. Archive the Project (Claude.ai setting) — knowledge preserved, Project hidden from main view. Can be un-archived later.
- Delete — client offboarded permanently, or the Project had enough messy experimentation that you want a clean slate. Delete removes the Project and its history.
Rule: when in doubt, archive. Storage is cheap; re-creating context is expensive.
The 90-day Project health check
Every quarter, do a 10-minute audit per active Project:
- Brand guide still current? If the client rebranded, voice shifted, or banned-phrases list evolved, update the file in the Project.
- Services still current? Clients add and drop services. The Project’s
services.mdneeds to track that. - Recent-content-inventory still current? This one decays fastest — update monthly if possible.
- Custom instructions still match how you actually work? If you’ve adopted new default behaviors ([VERIFY] tags, ending with summaries), put them in the system prompt.
- Any stale workstream chats? Archive completed-deliverable chats to keep the main view scannable.
- Any accidental cross-client bleed? Review recent chats for anything that referenced other clients by name. If so, surface the governance issue with the AI council.
- Custom instructions not contradicting themselves? 3 months of small edits can accumulate into contradictions. Re-read.
- Any Memory entries referencing this client that should be in the Project instead? Memory is cross-chat and leaks. Client-specific context belongs in the Project.
Memory × Projects — How They Interact
Three rules for resolving which surface to use for which kind of context:
- Client-specific knowledge → always Projects, never Memory. Anything tied to a specific client goes in that client’s Project. Memory is cross-chat and will surface the context in unrelated work.
- Personal working style → always Memory, never Projects. Your preferred output format, banned phrases you always exclude, role context about what you do — all Memory. Putting it in every Project is duplicative.
- Role context → Memory for the one-line summary, Projects for the specifics. “I’m an SEO strategist at a dental marketing agency” belongs in Memory. “I’m the SEO strategist on the Smile Springs account, reporting to the Columbus account director” belongs in the Smile Springs Project’s custom instructions.
What this looks like in practice
You’re an SEO strategist. You’re working on Smile Springs this morning and Riverside Orthodontics this afternoon.
Smile Springs chat — Memory loads: “I’m an SEO strategist at WEO, I want markdown tables, I hate corporate jargon.” Project loads: Smile Springs brand guide, services, audience, custom instructions. You get output that’s both in-your-style AND on-Smile-Springs-brand.
Riverside Orthodontics chat — Memory loads: same as above (unchanged). Project loads: Riverside’s brand guide (different voice), services (different mix), audience (different demographic). Same-you, different-client, same-session — no confusion, no cross-client leak.
Memory gone wrong: you said in a chat “Smile Springs’ Q2 campaign focuses on pediatric” and Claude filed that in Memory. Next week, in the Riverside Orthodontics chat, Claude references Smile Springs’ pediatric campaign. Embarrassing and a governance issue. Fix: open Memory (Settings → Memory), delete the Smile Springs entry. Never let Memory hold client-specific context.
Troubleshooting Projects + Memory
Common issues and their fixes:
- Claude ignores the brand guide. Usually the file wasn’t attached to the Project, or the custom instructions don’t reference it. Verify: (a) file is in Project Knowledge, (b) custom instructions say “Always reference attached brand-guide.md.”
- Claude ignores custom instructions on short prompts. Short prompts get short responses, and Claude sometimes skips the preamble where rules live. Fix: in the prompt itself, reiterate the one rule that matters: “Per brand guide, no banned phrases.”
- Project feels stale / outputs feel generic. Check if all 5 knowledge files are still attached, not just an old copy. Projects don’t auto-refresh when you edit a file elsewhere — re-upload.
- Memory surfaces wrong client’s info in a chat. Two things: (a) delete the offending Memory entry, (b) audit other Memory entries for client-specific content and delete those too.
- Custom instructions longer than ~1500 words. Claude follows the first half more closely than the second. Trim. Move details into knowledge files; keep custom instructions as the ruleset only.
- Need to share a Project with a freelancer or partner. Don’t. Team plan Projects are per-seat. Either: (a) add them to WEO Team, or (b) export a “handoff brief” from the Project to share as a doc.
Key Takeaways
- Three context surfaces, in increasing explicitness: Memory (auto) → Projects (scoped) → Artifacts (per-chat).
- Memory is on by default as of April 2026. Prune it weekly. Don’t let it swallow client confidential info.
- The WEO pattern: one Project per major client, loaded with 5 standard knowledge files + custom instructions.
- A fully-loaded Project costs 20 minutes to set up and saves 5–10 minutes per chat forever.
- Custom instructions = the Project’s standing rules. Make them short, actionable, and prioritize the ruleset.
- Client-specific context → Projects. Personal working style → Memory. Never mix.
- Run a 90-day Project health check — 8 items, 10 minutes, prevents silent drift.
- Artifacts are for structured output — tables, docs, mockups — not for chat.
- Read Module 7 before you load anything into a client Project or let Memory remember client-sensitive context.
Related
- Next: Module 5 — Claude Design
- Module 7 — WEO Rules (mandatory before loading client data)
- Module 6 — Skills, Tools, Connectors (Skills go deeper here)
- Anthropic Academy — Claude 101
- Claude AI reference topic (advanced usage after the course)
Try It (10 min, hands-on)
- Complete “Organizing Your Work and Knowledge” on Claude 101. ~15 min.
- Check your Memory pane. Avatar → Settings → Memory. See what’s already there. If empty (new account), leave it — it’ll fill as you work. If full of stale or irrelevant entries, prune.
- Seed Memory with your working style. New chat, send: “For future reference: I’m an [ROLE] at WEO Marketly, a dental marketing agency. I usually want output in markdown bullet points, tight copy, no corporate jargon. I hate ‘in today’s fast-paced world’ and ‘it’s important to note.‘” Claude will add this to Memory.
- Create a Project in claude.ai called “Sandbox — [Your Role]“. This is a safe practice project with no client data.
- Upload one SAFE file — e.g., a publicly-available dental-industry article, or a made-up brand guide for a fictional practice. Do NOT use real client files yet.
- Add a custom instruction to the project (one sentence): “Always end responses with a one-line summary of what you just did.”
- Start a chat in the Project. Ask a question about the content of the file you uploaded. Notice Claude reads from it automatically.
- Ask Claude to produce a table of 5 related blog-post ideas. Watch the Artifact pane open on the right. Edit one row directly.
Done? Move on to Module 5.