Working with Claude — A Brown Bag Session

Segment 01 · 12:00

One Claude, several rooms

Same model underneath — different surfaces for different kinds of work. The skill is knowing which room to walk into.

Chat

The conversation

Thinking, drafting, analysing, learning. You steer turn by turn. Best when the value is in the exchange — exploring options, refining an argument, reviewing a document.

Claude Code

The workshop

An agentic coding tool in your terminal, IDE, desktop app, or browser. It reads your codebase, edits files, runs commands, and commits — you describe outcomes, it does the work.

Claude Cowork

The colleague

Claude Code's agent architecture, pointed at knowledge work instead of code. Give it folders, connectors, and a task; it plans, executes end-to-end, and hands back finished files.

Chat

Claude Code

Cowork

You provide

Questions, drafts

A repo + an outcome

Folders, tools + a task

Claude returns

Answers, artifacts

Working code, commits

Docs, decks, sheets, done-ness

Interaction

Turn by turn

Delegate, review diffs

Delegate, steer checkpoints

Best when

Value is in dialogue

Output is software

Output is a deliverable

Segment 02 · 12:08

Claude Code: an agent in your terminal

Not a chat window bolted onto an editor. It takes action — edits files, runs tests, resolves conflicts, writes release notes — from wherever you already work.

~/house-of-wealth — claude

A 60-second feel for the loop: describe → plan → act → verify. Press “Run demo”.

Getting started

Install in a minute

Native installer or npm install -g @anthropic-ai/claude-code (Node.js 18+)
Sign in with your Claude account, cd into a project, run claude
Also available as VS Code / JetBrains extensions, in the desktop app, and on the web

The core loop

Describe outcomes, not steps

Build: "Add an endpoint returning user risk profiles, with tests"
Debug: paste the stack trace, ask it to find and fix the cause
Navigate: "Walk me through how auth works in this repo"
Automate: lint fixes, merge conflicts, release notes, CI chores

A CLAUDE.md file in your repo acts as persistent memory: build commands, code style, branch conventions, gotchas. Claude reads it every session, so you stop re-explaining your project. Keep it short and ruthlessly curated — it's loaded into context every time.

Project-level: ./CLAUDE.md — team conventions, checked into git
Personal: ~/.claude/CLAUDE.md — your own preferences, everywhere
Ask Claude to maintain it: "Add what we just learned about the build to CLAUDE.md"

For multi-file changes, ask for a plan before any edits (plan mode). Reviewing a five-line plan is cheaper than reviewing a five-hundred-line diff — and steering happens where it's cheapest. Approve, adjust, then let it run.

The Model Context Protocol connects Claude Code to external systems: design docs in Google Drive, tickets in Jira, messages in Slack, your own internal tooling. The agent stops guessing at context it can simply go and read — or update.

Claude Code composes with pipes and runs headless in scripts and CI. One documented example: tail -f app.log | claude -p "Slack me if anomalies appear in this stream". Anywhere you'd write a fiddly script, consider writing a sentence instead.

If you build anything — analysis notebooks, internal dashboards, a personal site — Claude Code is the fastest way to iterate. Tech-savvy non-coders were using it for file wrangling and research long before Cowork existed; that usage is literally why Cowork was built.

Segment 03 · 12:20

Claude Cowork: hand off the work itself

Cowork sits next to Chat. You point it at folders and tools, describe the deliverable, and it runs the task end-to-end — no copying out of a chat window.

1 · Brief

Tell Claude what you need, not how. Give it the folder with the raw material and name the output format: deck, doc, spreadsheet, tidy folder.

2 · Plan & run

Claude analyses the request, shows a plan, then executes — opening files, using tools, browsing when needed. Every step is visible; redirect any time.

3 · Review

You get finished files for review, not a wall of chat text. Start at your desk, check in from your phone; the session keeps going with your laptop closed.

⚠ Worth knowing: Cowork tasks consume noticeably more usage than chat — batch related work into one session, keep quick questions in Chat, and watch Settings → Usage. Claude's shell work runs in an isolated environment on Anthropic's servers, reaching your local files only through the desktop app with the access you grant.

Connectors plug Cowork into where your work already lives — Slack, Microsoft 365, Jira, Salesforce, Google Drive, internal tools. With one enabled, Claude can read and act: update a ticket, draft a reply, post to a channel, save a file. Authorise once from the Customize panel; use it in any session. On enterprise plans, admins control which connectors exist and whether each is read-only or can write.

The unlock: a connector plus your filesystem means Claude can pull data from a system and save it locally, or use local files to drive external actions — the two halves of most real workflows.

Skills are packaged procedures — a folder of instructions and templates Claude loads when relevant. Invoke one by name (/skill-name) or just describe the task. The compounding trick: finish a workflow once, then say "Package what we just did into a skill" — the built-in skill-creator captures the steps so next month's version is one prompt.

Plugins bundle connectors + skills so a whole team starts from the same setup. Anthropic ships role-based ones — e.g. Product (Jira, Confluence, Slack with skills for feedback clustering, PRDs, sprint updates) and Sales (Salesforce, Outlook with meeting prep and pipeline reviews).

Recurring reporting

"Every weekday at 5pm, read the six squad standup notes in this folder and produce the consolidated status report in our template, flagging blockers." Schedule it daily, weekly, or monthly.

Folder archaeology

"Sort this downloads folder", "pull the expense lines out of these receipt screenshots into a spreadsheet", "rename these 200 files to our convention."

Deck & doc assembly

"Turn these meeting notes, this data extract and last quarter's deck into a first-draft steering committee update."

Cross-tool chores

"Read this week's Jira movement for the platform project and draft the stakeholder email; save a copy to the project folder."

Segment 04 · 12:30 · audience participation

Which Claude for this?

Click a task. Argue with the verdict — that's the point of a brown bag.

Pick a task ↖

The honest answer to "which tool?" is almost always about where the output lives and how much back-and-forth you want.

Segment 05 · 12:38

Maximising Claude — the habits

Beyond picking the right surface, a handful of habits compound. These are the ones that change how much you get back.

Create a Project per workstream and load it with the standing context — style guides, templates, key docs, custom instructions. Every chat inside it starts warm. One for your platform programme, one for your writing, one for hiring.

The quality ceiling is set by the brief. State the audience, the format, the constraints, and what "good" looks like — ideally with an example. Ask for a plan or outline first on big deliverables; iterate on structure before prose. Positive and negative examples beat adjectives every time.

End big briefs with: "Before you start, ask me anything that would change your approach." Cheap insurance against confident work in the wrong direction — exactly what you'd want from a new hire.

Interactive HTML pages, working prototypes, Word docs, decks, spreadsheets — ask for the finished artefact, not text to paste somewhere. If you're demoing an AI concept to stakeholders, a clickable artifact lands harder than any slide.

Anything you've asked for three times is a candidate: package it as a skill, or make it a scheduled Cowork task. Your prompt library should shrink over time as it graduates into automation.

For market scans, vendor comparisons, or regulatory questions, use web search or the deeper Research feature rather than trusting trained knowledge. You get citations you can check — which matters in a bank.

Some of the highest-value prompts are adversarial: "Red-team this proposal as a sceptical risk officer." "What would a regulator flag here?" "Argue the opposite." Cheaper than learning it in the steering committee.

Verify numbers, citations and legal claims before they leave the building; keep client-identifying data out of prompts unless your deployment is approved for it; review agent actions with write access the way you'd review a junior's. Delegation without review isn't leverage, it's exposure.

Segment 06 · 12:44 · take-away

The prompt shelf

Copy these on your way out. Adapt freely — they're patterns, not incantations.

Claude Code · onboarding a repo

I'm new to this codebase. Give me a guided tour: what it does, the architecture,
the 5 files that matter most, and how a request flows through the system.
Then add a CLAUDE.md capturing build/test commands and conventions you observed.

Claude Code · safe delegation

Plan first, don't edit yet: [describe feature]. List the files you'd touch,
the approach, the tests you'd add, and any risks. Wait for my go-ahead.

Cowork · recurring status report

Every weekday at 17:00: read the standup notes in /ProjectX/standups added today,
produce the daily status report using /ProjectX/templates/status.docx,
flag blockers and cross-squad dependencies at the top, save to /ProjectX/reports.

Cowork · workflow → skill

Package what we just did into a skill called weekly-steerco-pack:
capture the steps, the template locations, and the formatting rules,
so next week this is a single prompt.

Chat · pre-mortem critique

Here's a proposal I'm presenting to [audience]. Red-team it as a sceptical
[risk officer / regulator / CFO]: the 5 hardest questions I'll get,
the weakest claim, and what evidence would strengthen it.

Any surface · better briefs

Before you start, ask me up to 5 questions whose answers would
materially change your approach. Then proceed.