You just got Claude — what do you set up first to get real work out of it?

Last updated 2026-08-17 · Overview — no new claims; each statement links to the entry that carries the primary sources.

This is the order Alejo actually walks people through on coaching calls — what gets set up first, where people stall, and what waits for a later session.

1. Pick a real task before touching settings. Alejo opens every call the same way: name the work that matters this week, then set Claude up on that — a tool session with no task behind it produces nothing you’ll return to. That’s practice advice from the calls, not a product fact.

2. Install the desktop app and start in Cowork. Chat is identical everywhere; the desktop app adds what matters — local files, dictation, quick entry (Desktop vs browser). Cowork is the managed agent: it works only in folders you connect, inside a sandbox that can’t wreck your machine (Cowork sandbox, Agent isolation). Pick a folder you can see in Finder, or your output stays invisible.

3. Use the best model and pay for enough. Defaulting to the top model sidesteps quality differences you can’t verify. If you’re rationing prompts, upgrade the subscription — overage credits bill at full API rates while plans discount steeply (Plan economics); truly high-volume bounded jobs route to cheap API tiers at ~1/10th–1/30th the price (Cheap-model routing).

4. Dictate, don’t type. Every session starts smarter when you talk for minutes instead of typing three lines: people give more context out loud, and context is what the model needs (Wispr Flow). Alejo’s frame on calls: each session is a newborn — smart, but it knows only what you show it.

5. Connect your data — then verify every connection. A connector sees the whole account, and instructions don’t narrow that (Connector scope). This is where people stall first: agents report success on fetches that failed, so demand a concrete test — a verbatim quote, the last email (Web fetch limits). Drive search is a live ~10-result keyword query; download the working set locally or build a navigation map once (Drive search, Doc freshness).

6. Loosen permissions deliberately. Alejo’s line in workshops: “lesson number one: always say yes.” Auto-approve mode has a second model review actions, and connector writes plus org rules keep prompting on every rung (Permission modes). What protects you is prompts, scope, and version history — not the agent’s promises (Agent safeguards).

7. Keep context in files, not memory. Automatic memory is a lossy synthesis; instruction files you write load every session and never surprise you. The inclusion test is staleness, not importance (Memory vs files). Work in markdown, feed spreadsheets as CSV (Markdown vs Office, Feeding spreadsheets).

8. Restart and fan out. When a session degrades or the task switches, leave it — degradation starts well below the context limit (Context rot). For many-source pulls, dispatch subagents, one per source (Subagents in chat).

9. Encode what repeats — the later-session material. A procedure you’ve pasted three times becomes a skill (Skills vs prompts); a daily routine becomes a scheduled run, local if your machine stays on (Local scheduled runs), hourly-at-best in the cloud without code (Event triggers). Claude Code is the same engine with your own file system and configuration — the graduation step, not the entry point (Cowork vs Code).

The rest of the wiki answers the single questions behind each step: full index.