AI·BI
AI·BI Training

Claude Cowork: Getting Started

AI·BI16 June 2026Training material
AI·BIClaude Cowork: Getting Started16 June 2026
Section 1 · What Cowork is

What Cowork is

Capabilities, craft, and the rules that keep you safe. The shift from Chat to Cowork is the shift from asking to doing.

Cowork is Claude in the desktop app, with the keys to your files, your tools, and your browser. Where Chat answers a question, Cowork takes a task — and gives you a deliverable back, not just a transcript. You attach a folder, name a goal in plain English, and Claude reads what is there, drafts what is needed, runs the calculations, navigates the websites, and writes the output back into your files. You stay in oversight; consequential steps wait for your nod.

Cowork runs in the Claude desktop app for macOS and Windows. It is available on every paid plan — Pro, Max, Team, and Enterprise — and a getting-started guide covers the install.

The shift from Chat to Cowork is the shift from asking to doing.Cowork, in one line
Where it runs — the Claude desktop app for macOS and Windows, on every paid plan (Pro, Max, Team, Enterprise). The getting-started guide covers the install; the product page has the overview.
Section 2 · What it can do for you

What it can do for you

Six surfaces are worth learning. Each is reachable in plain English from any Cowork session. The examples below are illustrative — substitute the data and systems your own work runs on.

Attach a folder; Cowork reads it, drafts what is needed, saves finished work back.

“Use last quarter’s board pack as a template. Draft this month’s commentary into the same folder.”

Permissions-scoped access to Microsoft 365, Salesforce, Slack, Atlassian — the systems you already run.

“Find the policy document from last Tuesday’s exec meeting and summarise the changes against the prior version.”

Sandboxed code

Python, shell, and SQL run inside an isolated VM — crunches CSVs and builds charts without opening Excel.

“Compare suppliers_2024.csv against our current vendor export and produce a shortlist of duplicates.”

Interactive output in your sidebar — a dashboard, tracker, or report — refreshable when the data changes.

“Build a board KPI tracker as a live artifact; refresh it when I drop new data into the BoardKPIs folder.”

Define a cadence once; Cowork runs it — a daily snapshot, a weekly digest, a monthly close-readiness check.

“Every Monday at 8am, pull revenue MoM by business unit, flag anything 10% off forecast, put it in my inbox.”

Connect Claude to Chrome to read and act on web pages — gather pricing, scrape a report, pull a data point.

“Find published cost-of-capital ranges for Australian mid-cap industrials over the last year and summarise the spread.”

Each surface links to its official documentation (↗). Examples are illustrative — substitute your own data and systems.
Section 3 · Plugins, skills, connectors

Plugins, skills, connectors

Three layers of customisation extend Cowork. You see them as a sidebar menu; underneath, they stack like this. Connectors wire Claude into your tools — permissions-based, granting access to specific data, not blanket. Skills are instruction files that teach Claude how to do a specific task in a specific way. Plugins bundle skills, connectors, and helpers into a single install.

What it isIn one line
ConnectorsWiringWire Claude into your tools — permissions-based, specific data not blanket.
SkillsInstructionsInstruction files that teach Claude how to do a task in a specific way.
PluginsBundleSkills, connectors, and helpers in a single one-click install.

Loading them into a session

There are two ways to load a skill, connector, or plugin. Explicit: click the + button in the conversation panel and pick exactly which one to activate — same tool every time, same data path. Implicit: just ask the question, and Claude scans the available skills and picks one it thinks fits.

How can I help you today?
opens the picker
Add files or photos
Skills
Connectors
Plugins
The + button in the conversation panel opens the picker.
ExplicitImplicit
HowClick + and pick the exact skill, connector, or plugin.Ask the question; Claude scans and picks one it thinks fits.
PredictabilitySame tool every time, same data path.Convenient, but less predictable when several could fit.
Best forHigh-stakes work and recurring workflows.Exploration — figuring out what is available.
Default to explicit — use explicit invocation for anything you will act on; lean on implicit when you are figuring out what is available.
Day-one plugin
Your firm's BI Assistant — the plugin you are most likely to reach for. Use it for any question about your organisation's data — revenue, margin, inventory, AR, AP, budget variance, supplier payments are all in scope. A router slash command at the start of a session loads it, then picks the right specialised skill for whatever you ask — queries, finance, analytics, investigations, branded deliverables.
Section 4 · Projects

Projects

A project in Cowork is a folder paired with a written brief that tells Claude how to work on anything inside it. Every session you open against the project starts with the brief already loaded — the same agent shows up each time, knowing your work, your conventions, what was decided last week.

Three pieces define a project.

1
The folder
Whatever you already organise on disk — your finance area, an M&A diligence room, the board pack folder — becomes the project. Cowork reads what is there, drafts back into the same place, and keeps notes scoped to that folder only.
2
The Instructions
A short written brief — the project’s CLAUDE.md equivalent — telling Claude who it is in this project, what it should do, what it should not do, and how to communicate. It loads automatically at every session start.
3
The memory
What Claude has learned across previous sessions in this project — decisions, data shapes, the conventions you settled on. Project-scoped, so what it learns in one project does not bleed into another.
Project-scoped memory
— what Claude learns in one project does not bleed into another.

Open a new session two weeks later and the agent is already up to speed: its context engineering routes through the Instructions, the folder, prior memory, and the connected capabilities. As the project evolves, ask Claude to update the brief, restructure the folder, or capture new conventions — the agent helps maintain its own context over time.

Context engineering is not a one-time setup; it is a discipline the agent shares with you.On context engineering

AI·BI
Section 5 · The session lifecycle

The session lifecycle

A Cowork session has an arc — four phases inside the session, a bridge between sessions, an automation layer above. Run the arc deliberately and the work compounds; drift and the next session burns half its budget re-establishing context. The toolkit in the next section is what you reach for inside this arc.

The session lifecycle Open, Work, Validate and Close run left to right inside a session. A Bridge hands off to the next session. An Automate band sits above the arc, encoding recurring patterns so the agent runs the bridge for you. AUTOMATE Encode recurring patterns as skills, then schedule the run — the agent works the arc while you don’t. THE SESSION Open plan & load Work iterate Validate cross-check Close handoff BRIDGE handoff doc next session Inside the session: Open → Work → Validate → Close. Between sessions: the Bridge. Above them: Automate. Run the arc deliberately and the work compounds; drift, and the next session burns its budget rebuilding context.
Open / Work / Validate / Close happen inside the session. Bridge happens between sessions. Automate sits above them — recurring patterns encoded so the agent runs the bridge for you.
Run the arc deliberately and the work compounds; drift and the next session burns half its budget re-establishing context.On running the arc
Open
Plan and load
Retrieve a plan from the project or draft a fresh one; review, approve, then run against it. The plan becomes the brief.
Work
Iterate
Execute against the plan without re-deciding structure mid-flight. You steer rather than direct turn-by-turn.
Validate
Cross-check
Trace each figure to source and spot-check a calculation by hand before anything leaves the session.
Close
Handoff
Summarise what was done, what was touched, what is still open. Save it somewhere durable — it becomes the bridge into the next session.
Section 6 · Managing a session

Managing a session

The lifecycle is the map. What follows is the toolkit you reach for, roughly in this order.

1
Pick the model and effort
Opus is the deeper thinker — default it for analytical and strategic work. Sonnet trades depth for speed on iteration and drafting. Effort tunes depth versus speed on top: Opus High for planning, Opus Medium for execution, Sonnet for quick drafts. Reach for Xhigh on hard one-offs, Max only when the answer justifies the wait, Low for routine batches.
2
Plan first, then execute
On session open, define the plan — retrieve one from the project or ask Claude to draft a fresh one. Review, approve, then run against it. The plan becomes the brief, so Work executes without re-deciding structure mid-flight and you steer rather than direct turn-by-turn.
3
Watch the window
A session’s working memory is one million tokens on Opus — about fifteen hundred pages. /context shows how full it is; /compact summarises older messages when the topic is alive but bloated; /clear starts fresh when the topic has shifted. When it is genuinely full, bridge to a new session.
4
Validate before it leaves
Outputs that change a shared system or land in front of a stakeholder need cross-checking. Ask Claude where each figure came from. Spot-check at least one calculation by hand. Plausible is not proof; the model can be confidently wrong.
5
Close with a handoff
Before walking away, ask Claude to summarise what was done, what was touched, what is still open. Save the summary somewhere durable — it becomes the bridge into your next session.
6
Decompose large work across sessions
A quarterly programme does not fit in one session, and should not. Break it into tasks each scoped to a single session, run them in sequence, and let the handoff doc carry the thread. A written specification — stored in the project folder — keeps every session pointing the same way.

Step 1 — Opus or Sonnet, and how hard to think

Opus is the deeper thinker — it defaults for analytical and strategic work, where being right matters more than being fast. Sonnet trades depth for speed on iteration, drafting, and exploration, on the same context budget. Effort tunes depth versus speed on top of the model.

OpusSonnet
RoleThe deeper thinker.Faster and lighter.
When to useAnalytical, strategic — where being right beats being fast. Most of the time.Iteration, drafting, exploration — trade depth for speed.
Context budgetOne million tokens.Same context budget as Opus.
Effort cheat sheet — model × effort, and where each earns its keep
SettingWhere it earns its keep
Opus HighPlanning — laying out the steps for a non-trivial piece of work, where the plan is the brief.
Opus MediumExecution — once the plan is clear, each step does not need maximum reasoning.
SonnetQuick drafts and follow-ups.
XhighHard one-off problems where reasoning is the whole task.
MaxOnly when the answer justifies the wait.
LowRoutine batches.

Step 3 — Watch the window

A session's working memory is one million tokens on Opus — about fifteen hundred pages of text. Three commands keep it in hand.

Session-management commands
# show how full the window is, and what is in it/context# summarise older messages to free room when the topic is alive/compact# start fresh when the topic has shifted/clear
One million tokens — a session's working memory on Opus is about fifteen hundred pages of text. When the window is genuinely full and compaction has run its course, bridge to a new session — close, save the handoff doc, paste it back in.

Step 4 — Validate before it leaves

Outputs that change a shared system or land in front of a stakeholder need cross-checking. Ask Claude where each figure came from. Spot-check at least one calculation by hand.

Plausible is not proof — the model can be confidently wrong. A number you act on is a number you have traced to source yourself.

Step 6 — Decompose large work across sessions

A quarterly programme does not fit in one session, and should not. Break it into tasks each scoped to a single session, run them in sequence, and let the handoff doc carry the thread between them.

P1Write a specification for multi-session work
The leverage tool is a written spec — objective, scope, acceptance criteria, constraints, non-goals — stored in the project folder so every session loads it as part of the Instructions. Open session four of a six-session programme and Claude reads the spec and immediately knows the destination, what is done, and what is out of bounds. Without one, each session re-litigates the brief and drifts a little further. With one, every session points the same way.
Section 7 · Worked examples

Worked examples

Three patterns cover most of what you will do in Cowork. They scale from a single-session ask to a programme that spans weeks.

command session / spec output
Quick investigation single session, iterative
/initiate iterate cross-check branded XLSX
Load the BI Assistant and work the problem. Ask, read, ask the follow-up the answer surfaces, drill in where it gets interesting. When you have nailed it, spot-check a figure against source and ask for a branded Excel — it ships to your downloads.
Plugin workflow skills chained in one plugin
/select /enrich /review 7-sheet XLSX
A multi-step deliverable that lives across plugin skills. A web-product enrichment run is a clean example: select identifies the brand and product cohort; enrich scrapes supplier sites, reads data sheets, and generates web copy; review consolidates the run into a branded multi-sheet workbook ready for import. Each skill knows what came before and what is needed next.
Multi-session programme spec-anchored, sessions chained
spec session 1 session 2 session N synthesis
A quarterly inventory analysis across every business unit is too big for one session. Open a project folder; write a spec — objective, scope, acceptance criteria, what is out of bounds — and store it in the project's Instructions. Decompose into tasks: one per business unit, plus a synthesis. The spec keeps every session pointing the same way; the synthesis pass pulls the outputs into a board-ready deliverable.
NextStart with the Quick investigation pattern
It is the shortest path to a result you can trust. Load the BI Assistant, work a single question end-to-end, spot-check the figure, and ship a branded workbook. Once that rhythm is familiar, the plugin and multi-session patterns are the same arc at larger scale.
Section 8 · Three rules to stay safe
Safety

Three rules to stay safe

Three rules carry most of the safety load in Cowork. Anthropic's guide to using Cowork safely covers the rest.

P11 · Verify before trusting
When Cowork hands you a number, definition, or recommendation you will act on, ask where it came from. Plausible is not proof; the model can be confidently wrong.
P12 · Drafts, not sends
Cowork can write to your files, email through connectors, and overwrite documents. Default to draft-and-show for anything that sends, posts, deletes, or modifies a shared asset. Local-file access is powerful and unforgiving.
P13 · Fresh session for new tasks
Sessions accumulate framing. New task, new session — sharper answer, less chance of cross-contamination.
Plausible is not proof; the model can be confidently wrong.The spine of rule 1

AI·BI — Engineered Clarity · Claude Cowork: Getting Started · Training material · 16/06/2026

Claude, Claude Cowork and Anthropic are trademarks of Anthropic PBC. This is independent AI·BI training material — not affiliated with, sponsored by, or endorsed by Anthropic. Product names and links are provided for reference only.