For collaborators — the course

The exercise

Work with public or made-up material — nothing from your organization.

1. On your machine, create a folder named claude-practice. Put exactly three plain-text files in it: a public press release you copy from any organization’s website, a made-up meeting agenda you write yourself, and a shopping list.
2. Give Claude access to that folder — and only that folder — using whatever access surface your setup provides (a project folder, a connected directory, or pasting the three files into one chat if no folder surface is available).
3. Ask: “List the files you can see, and summarize each in one line.”
4. Then ask: “What is in my Documents folder?” — and read the answer carefully. If Claude asks permission to reach anything beyond your folder, say no — that prompt is the boundary working, and declining it is part of the exercise.

Expected outcome

Claude names the three files and summarizes them. In this exercise’s setup, when asked about Documents, Claude worked from what you gave it — it either said it could not see that folder, answered strictly from the three files, or asked to go wider (which you declined). The practice this teaches is the habit, not a product guarantee: choose a narrow folder every time (Use a dedicated working folder, not broad desktop access), and treat any widen-access prompt as a decision point, never a formality.

An example failure

A learner grants access to their whole home directory “to save time”, then asks a broad question — and the reply quotes from a file they forgot was there. Nothing malfunctioned: the tool worked exactly at the width it was given. The failure was the width. Access is a decision you make before the first prompt, not a thing to tidy up afterwards.

Facilitator check

Ask the learner to answer, in their own words: which folder did Claude have, what was in it, and what would change if the folder also held something confidential? A pass is a specific answer (“these three files, and nothing else”) — not “I think it only saw the folder.”

Next step

Credentials are one of the things that must never even enter the conversation — continue to the Secrets module and its exercise (Exercise — spot the secret, name its home).

Source & currency

• Source: this curriculum’s own practice, written for the collaborator course (Norrsken C2); the underlying access findings carry their own provenance on the Folder & Desktop Access module.
• Status: a synthetic-data exercise — it makes no product claim of its own.

The exercise

All six strings below are fictional — invented for this drill, valid nowhere. Do steps 1 and 2 on paper or out loud — none of the six strings needs to enter a chat; only step 3’s checklist drafting uses Claude, and it carries no credential-shaped string at all.

1. Sort these into never goes in a chat and fine to discuss: a string beginning sk-fake-… labelled “API key”; the sentence “our newsletter goes out on Tuesdays”; a password written on a sticky note in the vignette; a “database connection string” with a password inside it; the name of your password manager; a one-time login code from a text message.
2. For each item in your never pile, practice the replacement sentence out loud: “The key lives in our password manager under service-x” — the name of the place, not the value.
3. Ask Claude to help you draft a checklist titled “what to do if a credential ends up somewhere it shouldn’t” — using no real credential anywhere in the conversation.

Expected outcome

Four items land in the never pile (the fake API key, the sticky-note password, the connection string, the one-time code); two are fine (the newsletter fact, the manager’s name). The drafted checklist ends with the owner rotating the credential — because once a value has been exposed, the remedy is a new value, not deletion of the message (When a key slips into a chat by accident: destroy it, then store it properly).

An example failure

A learner pastes a real-looking key “just to test whether it works.” The test proves nothing a fictional string would not have proved — and if the key was real, it now exists in a conversation whose retention they do not control. The drill’s rule (Never paste credentials into a chat, a shared folder, or an artifact) has no just-this-once tier.

Facilitator check

Ask the learner to say the replacement sentence for one never item without being prompted. A pass names a location (“in the keychain”, “in our manager under this entry”) and never speaks the value. If they hesitate on the one-time code, point out that short-lived is not the same as harmless.

Next step

What happens to everything you do type? Continue to Retention and its exercise (Exercise — find your own privacy toggle).

Source & currency

• Source: this curriculum’s own practice, written for the collaborator course (Norrsken C2); the credential rules carry their provenance on the API Keys & Secrets module.
• Status: a synthetic-data exercise — every string in it is fictional by construction.

The exercise

Nothing sensitive is typed in this exercise; you are reading settings, not sending content.

1. In Claude, open Settings → Privacy and find the toggle labelled “Help Improve our AI models” (label re-checked 2026-09-17 — receipt R90).
2. Read its current state on your account. Do not change it during the workshop — the point is to know where it is and what it says. After the workshop, the setting is yours: choosing either state on your own account is legitimate, now that you know what each state means.
3. Say out loud what the ON state means: chats can be retained in de-identified form, for an extended period, for model improvement (receipt R89).
4. Now find the thumbs-up / thumbs-down controls on any reply — and before rating anything, state the rule: submitting feedback or a bug report is an explicit opt-in that can permit use of the related conversation, with its own long retention (receipt R88). Feedback is one such route among others; the toggle is not the whole story.

Expected outcome

The learner can navigate to the toggle unaided, reports its actual state (either state is a legitimate finding — this is a reading exercise, not a compliance instruction), and states the feedback rule unprompted before touching a rating control.

An example failure

A learner switches the toggle OFF and concludes that everything they previously typed is now gone. The OFF state is a commitment about future training use — it is not an automatic deletion of anything (Consumer plans: the training toggle governs training use — up to 5 years if on; no automatic window published if off); deletion is its own act with its own stated window (receipt R89). The setting reaches chats you already sent — for future training use — but training already done is not undone, and nothing is deleted.

Facilitator check

Ask: “What does the thumbs-up button do besides express an opinion?” A pass mentions that it can open the related conversation to further use and retention (The feedback button hands over the whole conversation for 5 years — and it, or a bug report, permits training). If the learner says the toggle covers it, walk them back through step 4.

Next step

The strictest boundary of all — patient data. Continue to HIPAA and its exercise (Exercise — the boundary drill).

Source & currency

• Source: this curriculum’s own practice, written for the collaborator course (Norrsken C2); the underlying retention findings are receipted on the Data Retention & What Goes Back module (fresh fetches 2026-09-17: receipts R88, R89, R90).
• Status: a settings-reading exercise — no content is submitted; the product facts it points at carry their own chips.

The exercise

Every scenario below is fictional — written for this drill, describing no real person. The drill is spoken or on paper; the point is that most cards never reach a chat at all.

State the ground rule first: the workshop runs on consumer accounts, and most ways of using Claude are never covered by a BAA (Most ways of using Claude are never covered by Anthropic's BAA — and on two cloud platforms, coverage is someone else's call) — the workshop account is one of the never-covered ways. Then sort the cards:

1. “Summarize this (fictional) clinic letter about a named patient’s seizure history.”
2. “Draft a template letter a clinic could adapt — no patient in it, invented placeholders only.”
3. “Here’s a spreadsheet of (fictional) members with diagnoses attached — make a chart.”
4. “Explain what a variant of uncertain significance means, in plain language.”
5. “I removed the (fictional) patient’s name from the case description — now can I paste it?”

Expected outcome

Cards 1 and 3 stay out — patient-identifying content, real or realistic, does not enter a consumer chat. Cards 2 and 4 may proceed — no individual is in them. Card 5 also stays out, and it is the card the drill exists for: removing a name is not de-identification, and the judgment of “identifying” is not improvised at a keyboard — see the pre-flight checks (A de-identification pre-flight — the checks to run before anything is pasted) and route the real-world version to your organization, its counsel, and its Anthropic contact.

An example failure

A learner reasons “it’s de-identified enough” about card 5 and pastes it. The failure is not the paste — the card was fictional — it is the reasoning, which in real work substitutes an individual’s on-the-spot judgment for an organizational decision that was never theirs to make. The drill counts this as a miss even though nothing sensitive existed.

Facilitator check

Hold up card 5 and ask why it stays out even with the name gone. A pass reaches “because removing a name is not de-identification, and this decision belongs to the organization, not to me at a keyboard.” Also confirm the learner can say, unprompted, whether today’s workshop account is covered for patient data (it is not).

Next step

With the boundary set, build something repeatable — continue to Skills and its exercise (Exercise — describe a source-checking Skill).

Source & currency

• Source: this curriculum’s own practice, written for the collaborator course (Norrsken C2); the coverage findings carry their receipts on the HIPAA & Patient Privacy module, and the vendor’s own caution about sensitive input is recorded at receipt R87 (fetched 2026-09-17).
• Status: a fictional-scenario drill — it asserts no coverage fact of its own beyond what the linked findings establish.

The exercise

Use only public web pages as material.

1. In an ordinary chat, write instruction text for a repeatable task, in plain sentences (a Skill is instructions, not programming — A Skill is a folder built around a single SKILL.md): “When I give you a claim and a URL I have opened, quote the exact sentence from that page that supports the claim. If no sentence on the page supports it, say so plainly instead of approximating.”
2. Test it once: give Claude a claim from this curriculum together with text you copy from the cited public page, and ask it to apply your instructions.
3. Now stress it: give a claim the page does not support, and see whether your instructions produce “not supported” — or a helpful-sounding stretch.
4. Tighten one sentence of your instructions based on what happened, and run the stress test again.

Expected outcome

Two runs: one where the instructions work, and one where the first draft bends — typically the model “finds” support that is really a paraphrase. The tightened version (e.g., adding “quote verbatim or refuse”) behaves better, and the learner has experienced the whole craft loop: describe → test → watch it fail → tighten (Building a simple Skill needs no code — but the interactive builder is not built in).

An example failure

A learner writes “check whether this is true” as the instruction. That delegates the standard of evidence to the model — and the reply reads confident either way. Instructions that name the evidence (“quote the exact sentence”) fail visibly; instructions that name a verdict (“is it true?”) fail invisibly. The invisible failure is the one that reaches a family.

Facilitator check

Ask the learner what their step-3 failure looked like and which single sentence they changed. A pass describes a concrete observed failure, not “it worked fine” — an exercise where nothing failed means step 3 was skipped, and the loop was not experienced.

Next step

The habit under all of this — checking claims at the moment of use — is the course’s last stop (Exercise — trace one claim to its receipt).

Source & currency

• Source: this curriculum’s own practice, written for the collaborator course (Norrsken C2); Skill product facts carry their receipts on the Skills module.
• Status: a public-material exercise; the Skill described is a teaching vehicle, not a shipped artifact.

The exercise

This exercise uses the site you are reading and public vendor pages — nothing else.

1. On any module page, pick one finding whose chip shows ✓ (checked against an Anthropic page on a stated date).
2. Follow its receipt link into the verification log, and read the row: the URL, the date of the fetch, and the short verbatim quote.
3. Open that URL yourself, in your own browser, and find the quoted words on the live page.
4. Report one of two outcomes: “found — the page still says it”, or “moved — the page no longer says that.” If it moved, you have just done the most valuable thing a reader of this site can do: flag it to the group with the URL and what you saw.

Expected outcome

The learner completes the chain claim → chip → receipt → live page with their own eyes — not the site’s. Most runs end in “found.” A “moved” run is not a failure of the exercise; it is the exercise working (The three checks: available, eligible, checked — and The Fable/Mythos episode: model availability is verified, never assumed is this site’s own worked example of exactly that discovery).

An example failure

A learner reads the ✓ chip and stops — “it’s verified.” The chip records that someone else checked on a past date. The habit this course exists to leave behind is checking at the moment of use; a verification with a date on it is a fact about that date.

Facilitator check

Ask which claim they traced and what the vendor page said in its own words. A pass quotes or closely paraphrases the live page — not the site’s summary of it. If time allows, ask what they would do if the page had moved (a pass names flagging it with the URL, not silently distrusting the site).

Next step

Course complete. Bring your traced claim to the workshop’s recap — and if you are writing a Discovery Lab thesis in the closing block, this habit is the “evidence of success” column.

Source & currency

• Source: this curriculum’s own practice, written for the collaborator course (Norrsken C2); the receipts mechanism it exercises is the site’s own verification log.
• Status: a self-referential exercise by design — the site is the material.