Using Claude to Run Your Export Business

Why an exporter should care

Most of the desk work in an export business is language and documents: researching which companies import your product, drafting and following up outreach, building price and landed-cost models, turning a product list into a catalogue or a company profile, keeping invoices and shipping papers in order, checking what competitors are doing. AI assistants — Claude in particular — can now do large parts of that with you, working directly on your own files, not just answering questions in a chat box. This guide explains the pieces and how a non-technical exporter can use them. You do not need to know how to code.

Product names, model names, prices, and limits in this space change fast. Treat the specific names below as a snapshot; check anthropic.com and claude.com for what is current, and use the /model and /status commands (below) to see what you actually have.

The ways you can use Claude

• Claude app / web chat (claude.ai) — the familiar chat box. Ask questions, paste text, upload a PDF or spreadsheet, get an answer. Good for drafting, learning, quick analysis.
• Projects and connectors — inside the app you can create a Project (a workspace with its own instructions and files) and connect Claude to services like Google Drive, Gmail, and Calendar so it can read from them.
• Claude for Chrome — a browser extension that lets Claude act on the web for you: open sites, search, filter, fill forms, take screenshots, pull results into a document. Useful for structured research (e.g. "find importers of X in country Y on this directory and list company, site, contact"). Still an early capability — supervise it, and never let it enter payment or login details on your behalf.
• Claude's desktop agent (the course calls it "Cowork") — a desktop experience that brings the same agentic ability to your own folders. You point it at a folder, describe a task, and it works through it: reads the files, makes a plan, creates sub-tasks, writes Excel/Word/PowerPoint/PDF output, can use connectors, and can run on a schedule. It runs in an isolated environment on your machine. Needs a paid plan.
• Claude Code — the most capable agentic tool. Runs in a terminal or inside the VS Code editor. Despite the name, you drive it in plain English; it does any coding itself. Best for anything involving many files, a website/landing page, repeatable workflows, or automation. This guide covers it in detail below.
• Claude in Excel and PowerPoint — add-ins that put Claude inside those apps to build, edit, summarise, and reorganise spreadsheets and decks in place.

Two concepts worth understanding first

Tokens and the context window

Everything you send or receive is measured in tokens — roughly 4 characters of English per token, about 750 words per 1,000 tokens. You are billed (or your plan usage is counted) by tokens, not words. The context window is how much the model can hold in view at once — input plus output. Modern models have very large windows, but a folder of hundreds of documents still won't fit at once, which is why agents load only what a task needs and will compact (summarise) a long conversation to free space. Practical effect: keep prompts focused, start a fresh session for a new task, and don't dump irrelevant files into the agent's folder.

Human in the loop

Agentic tools ask permission before editing files, running commands, sending email, or deleting anything — and you approve each step (or approve for the session). You can turn this off ("auto-accept"), but for anything that touches real customers, money, or files you can't easily recreate, leave it on and read what it proposes.

Getting started with Claude Code (no coding needed)

1. Install the VS Code editor (or use a terminal), then install Claude Code with the one-line command from code.claude.com/docs (there's a Windows line and a macOS line).
2. Make an empty folder for your project (e.g. price-model or germany-buyers). In VS Code: File → Open Folder → pick it.
3. Open the terminal panel, type claude, press Enter, and trust the folder.
4. Type your request in plain English. Claude Code reads/writes only inside that folder.

It looks plain — a text terminal — but it is doing real work: reading files, planning, writing, running checks.

The three modes (cycle with Shift+Tab)

• Normal — Claude can read, write, and run commands, asking permission each time. Default.
• Auto-accept ("accept edits") — same, but skips the permission prompts. Faster, less safe.
• Plan mode — Claude explores and proposes a plan but changes nothing. Use it to start a non-trivial task or to look at unfamiliar files safely, then switch to Normal to execute. Skip plan mode for a small, obvious change.

Built-in slash commands

Type / to see them. The useful ones:

• /model — see and switch the model. Tiers: a most-capable model (Opus tier) for hard work, a balanced fast model (Sonnet tier) for everyday tasks, a lightweight fast model (Haiku tier) for quick answers. Names and version numbers change — this menu is the source of truth.
• /context — how much of the context window is in use, and by what.
• /compact — summarise the conversation to free up context.
• /usage — your plan usage this session and this week.
• /status — version, account, plan, connected MCP servers.
• /diff — review the file changes Claude has made.
• /init — scan the project and write a CLAUDE.md (see below).
• /help — list everything.

CLAUDE.md — your project's standing brief

CLAUDE.md is a plain markdown file at the root of a project that Claude Code reads every time it starts in that folder. Put the things you'd otherwise repeat every session: what this project is, who you are, your company, the product, conventions to follow, things never to do. /init drafts one from the project; then edit it by hand. It's the difference between re-explaining yourself daily and having Claude already know the context.

Skills — teaching Claude a repeatable task

A skill is a folder with a SKILL.md file: a short description at the top (a name and a one-line "what this does / when to use it") followed by detailed instructions, examples, and optionally templates or scripts. Claude loads the description always, and reads the full skill only when a task needs it — so many skills can be available without clogging the context. Skills make output consistent: define the method once, reuse it forever.

• Built-in skills cover Excel, Word, PowerPoint, and PDF creation, plus a **front-end design** skill that makes generated web pages look professionally designed instead of generic, and a skill-creator skill that helps you write new skills.
• Official skills live at github.com/anthropics/skills.
• Third-party skill marketplaces exist with thousands of community skills. Useful, but a skill can instruct the agent to take real actions — only install from sources you trust, and read a skill before using it.
• Your own skills — write a SKILL.md for a workflow you repeat (e.g. "turn a product spec into a line-sheet entry in our format", "draft a buyer follow-up in our voice") and add it. See [[branding_for_exporters]] for what a brand-voice skill should encode, and [[finding_buyers]] / [[linkedin_for_exporters]] for outreach method it could bake in.

Custom slash commands — your own shortcuts

A custom slash command is just a markdown file in the project's .claude/commands/ folder (e.g. .claude/commands/my-). Running /my-brand makes Claude read that file. A common use: store your brand identity — name, tagline, colours, fonts, tone of voice, words to use, words to avoid, sign-off phrases — so every email, caption, flyer, or page it produces is on-brand without re-typing the guidelines. You can ask Claude Code to create the command file for you by describing the brand in one prompt.

MCP and connectors — plugging Claude into your tools

MCP (Model Context Protocol) is an open standard (from Anthropic, late 2024) for how an AI agent talks to outside tools and data — the analogy is a universal USB-C port. Because a tool "speaks MCP", any MCP-capable agent can use it without custom glue code. Connectors are the ready-made ones: Google Drive, Gmail, Calendar, Slack, GitHub, and more. Connect them in settings, and Claude can, for example, read a shipment tracker in Google Sheets, build a summary, and email it — with your approval at each external step. Only connect what a task needs; a connector grants real access to that account.

Plugins — bundles of capability

A plugin packages several things into one installable unit: skills, MCP tool connections, slash commands, sub-agents, and hooks (small checks that auto-run at set points, e.g. "before publishing, verify no API keys or secrets are in the files"). Where a skill teaches one task, a plugin equips a whole role. Install them from a marketplace inside Claude Code with /plugin. Anthropic ships plugin sets for functions like engineering, finance, legal, marketing, sales, and data.

Plugins an exporter might actually use:

• code-review — reviews changed files for quality, security, and missing tests before you publish. Useful even for a simple website.
• feature-dev — turn a short written description of a feature into a working implementation.
• playwright — lets the agent drive a real browser: open competitor sites, capture full-page screenshots, fill forms, run end-to-end checks. Good for competitor and market research.
• marketing — brand positioning, content calendars, landing-page recommendations, outreach/email sequences, SEO audits.
• ralph loops ("Ralph Wiggum") — runs the agent through repeated build-improve-commit cycles for long autonomous sessions; each loop refines the previous output.
• context7 — pulls live, version-correct documentation for a library into the session so generated code isn't based on stale training data.
• superpowers — a combined plan → build → debug → review lifecycle in one.

A realistic combination: use playwright to research two or three competitors (it visits the sites and writes up messaging, colours, calls-to-action, gaps), then feed that to the marketing plugin to produce a positioning and content plan as a formatted page.

Practical uses for an export business

• Buyer and market research — point the browser tool or the desktop agent at trade directories and company sites; have it compile a structured list (company, country, site, product fit, named contact and role) and a short profile of each. Confirm contacts before relying on them — see [[finding_buyers]].
• Pricing and landed-cost models — give the agent your cost inputs (or a folder with quotes and a freight rate) and ask for an Excel model with FOB/CIF build-up, a sensitivity table, and charts. Then check the numbers yourself.
• Document handling — point it at a folder of invoices, POs, and shipping documents; ask it to sort them into subfolders, extract totals into a spreadsheet, flag currency mismatches, and (via a connector) email you the summary.
• Catalogues, line sheets, company profiles, decks — turn a product list plus specs and photos into a formatted catalogue or a company-profile PDF; build or restructure a pitch deck inside your own template with Claude in PowerPoint.
• Outreach at consistent quality — with a brand-voice slash command or skill, generate first emails, follow-ups, and LinkedIn posts that all sound like your company; fold in the sequence and cadence from [[finding_buyers]] and [[linkedin_for_exporters]].
• A simple website or landing page — describe the page (hero, product sections, an enquiry form) and Claude Code builds a single self-contained page; the front-end design skill makes it look designed. See [[branding_for_exporters]] for the identity to feed it.
• Scheduled routines — e.g. every Monday, research named competitors' new activity and send you a brief.

Safety and good practice

• Keep backups. Before letting an agent reorganise or edit a folder, have a copy elsewhere. Agents can delete and overwrite.
• Keep sensitive data out of agent folders. No customer lists, contracts with confidentiality terms, banking details, or passwords in a folder you hand to an agent, and don't paste them into chat. These tools are still maturing.
• Verify anything factual or numerical. Models sound confident and still make mistakes. Check figures in a financial model, dates, duty rates, and any legal or tax statement against an authoritative source or your customs broker.
• Stay in the loop for outward actions. Sending email, submitting forms, publishing, paying — approve these yourself; don't hand over credentials or card details.
• Don't pass AI output off as genuine. AI-generated product photos, customer reviews, or testimonials presented as real are dishonest — use your own.
• Only connect and install what you trust. Each connector is real account access; each third-party skill or plugin can instruct the agent to act. Prefer official sources; read before installing.
• Watch your usage. Agentic work consumes far more tokens than chat. Check /usage, and expect to need a higher plan tier for heavy agent use.

Where this connects

This complements the rest of the guidance rather than replacing the exporter's own judgement: the buyer-finding method is in [[finding_buyers]], LinkedIn in [[linkedin_for_exporters]], and brand identity in [[branding_for_exporters]] — AI tools help you execute all of them faster and more consistently. EximAI on eximhub.pro can answer trade and tooling questions directly.