Level 2 · Connect — the lesson
Connect (MCP)
Claude reaches your real tools itself — one tool, usually one action at a time, inside fences you draw.
The line One tool and (usually) one action at a time, with you approving each write — not whole multi-step jobs run unattended. That's Level 3.
Read the full referencethe deep, complete version of this levelPlay the models below — each one you try climbs a floor of your tower.
Grant a tool — flip a scope, watch what it can do
Claude scanned your inbox — but open each flag before you trust it
Confirming the Phase 1 boundary before we start on Monday. Need your sign-off by Friday.
Genuine — a real scope question waiting on you. Act on it.
Invoice #4471 attached. Amount due: $12,400. ("Scope" appears only in the SOW footer.)
Mislabeled — it's an invoice. Claude matched the word "scope" in a footer. Asking it "are you sure?" wouldn't have caught this. Opening it did.
Claude flagged both as "scope decisions." Open each — the model can't grade its own homework.
Do it now — 4 small steps
One step = one paste into Claude (the mobile app is perfect). Your progress saves on this device.
-
Step 1 of 4 · Connect one tool
One connector is enough to feel the difference.
In the Claude app: Settings → Connectors → connect your Calendar (or Gmail). Grant access and review the scopes it asks for.
You should see: A “connected” tick next to the tool.
-
Step 2 of 4 · Ask about your real week
Now it answers from your world, not general knowledge.
What does my week look like? Anything I should prepare for?
You should see: Your actual meetings, with prep suggestions.
-
Step 3 of 4 · Ask something honest
Connected Claude can judge, not just fetch.
Which of my meetings this week could I skip or shorten? Be honest.
You should see: An opinion grounded in your calendar.
-
Step 4 of 4 · Make the artifact
From fetch → to finished briefing.
Write my Monday briefing: top 3 priorities from my calendar, anything I seem to owe people, and one thing I should decline.
You should see: A briefing you'd actually read on Monday.
The ideas behind the steps — tap a takeaway to unpack it.
What Connect is
Takeaway You stop being the copy-paste bridge — within the scope you grant.
At Chat, you were the bridge — read the calendar, copy the bit that matters, paste it in. A connector removes the bridge: a scoped link to one real tool (Calendar, Gmail, Drive, or a custom MCP server). Now Claude fetches your real data itself and can propose actions.
The new risk: a wrong action on a live system
Takeaway "It ran" isn't "it's right" — it proposes, you commit.
A wrong answer used to cost a worse paragraph. Now it's a draft to the wrong client. The risk lives in writes — so keep the irreversible verb yourself: it drafts, you send; it proposes an event, you confirm. Undo a grant anytime in Settings → Connectors.
Verify against the source, never the model
Takeaway Read before you trust, re-read after you act — check the tool, not the model.
The model can't grade its own homework — "are you sure it's the right file?" just gets a confident "yes" and the same wrong record re-cited. The catch: it flags a "Phase 2 — scope & SOW" thread that's actually an invoice with "scope" in the footer. Only opening it caught that. Re-open the real draft/event after every action.
Is this the right level for your task?
You want to handle something using your real calendar, mail, or files. Which fits?
Pick the closest match — you'll get a verdict.
The capstone — the whole level in one prompt
Do a low-stakes, **read-only** slice of your real world (~15 min) — your actual week is ideal. Earn the read before any write. Connect your calendar/mail/Drive at read-only scope, then paste this in.
I'm practising connecting you to my real tools at READ-ONLY scope. Be my tutor and hold me to the discipline. My objective: tell me what this week actually looks like and what I need to prep — especially Tuesday — read from my real calendar, inbox, and Drive. Don't guess. Rules for this run: 1. READ-ONLY. Do not draft, label, create, or change anything. If something needs an action, propose it and stop for me. 2. For every claim, cite the specific event / file name / sender it came from, so I can open the real record myself. Don't infer beyond what's in the tools. 3. List my meetings with times; for each, flag anything I still need to prep and say which file or email tells you that; surface any unanswered emails tied to those meetings. Then, before I act: pick the single most load-bearing item and tell me exactly which record to open to verify it, and what would tell me you got the WRONG record (wrong version, same-titled file, a thread whose subject misleads). Don't tell me you're sure — point me at the source. Finally, remind me to review my connector scopes and revoke anything I won't need again.
Self-check
✓ You've got this level. On to Level 3: Delegate →
You've leveled up when…
You've mastered Connect when, unprompted, you grant exactly the scope a task needs, gate every write with propose-then-confirm, and verify the effect against the real record. The tell you've outgrown it: you keep wishing Claude could run the whole multi-step job end to end instead of you approving one fetch or action at a time.
Level 3 — Delegate (Cowork): hand over a whole job, not one action. Give Claude a standing space and review the run instead of approving each step.
Next · Level 3: Delegate →