Most articles about Claude Code are written for software engineers. This one is written for the people who run a virtual assistant agency in Australia and have started using it as a productivity layer for our team.

If you are not a developer, do not be put off by the name. Claude Code is a CLI tool from Anthropic that lets Claude touch files, run shell commands, and stay grounded in the actual artefacts you work with. Email drafts, SOPs, spreadsheets, image batches. Anything on your machine, basically.

Here is how we actually use it.

The setup, in 60 seconds

Install Claude Code on the VA’s machine (npm install -g @anthropic-ai/claude-code or the platform installer). Claude Pro subscription gets you a generous Claude Code quota.
Drop the VA into a working folder for the client. SOPs, templates, exports, drafts all live in this folder.
The VA runs claude from that folder and gets a session that already has the folder’s context loaded.

That is the entire setup. No coding required.

The five workflows that pay back fastest

1. SOP audits

Every quarter we audit a client’s SOP doc against three things: is it still accurate, are the templates still current, are there gaps a new VA would trip over.

Claude Code makes this 8x faster. The VA opens the SOP folder, runs:

Audit this SOP folder against these questions:
- Are any links broken or pointing to old documents?
- Are any tools mentioned that we don't use anymore (cross-check with the active-tools.md list)?
- Are there templates referenced in the SOP that don't exist as files in this folder?
- For each main section, is there an example output? If not, flag.
Output as a markdown report with file:line references.

Claude walks the folder, opens every file, runs the checks, produces a report. 5 minutes vs 40.

2. Content QA before publication

We publish blog posts, newsletters, and client deliverables every week. Before anything ships, the VA runs a content QA pass.

Read this blog post draft and check:
- Australian English throughout (organised, colour, analysed, programme)
- No em-dashes anywhere
- Typographic apostrophes (don't, won't, it's), not straight
- All headings use sentence case, not Title Case
- Every external link works (use WebFetch on each)
- Reading time at the top matches the actual word count (250 wpm)
- Any factual claim about pricing, regulations, or our policies is cross-checked against the source-of-truth files in this folder

Output a punch list of fixes, in order from biggest to smallest.

Claude flags issues a human eye would miss after the third pass. The VA fixes, ships.

3. Data wrangling without Excel

VAs often need to reshape data: a CSV from Klaviyo segments needs to merge with another CSV from Shopify, then de-dupe by email, then output for a Klaviyo upload.

Pre-Claude: 30-45 minutes in Excel with VLOOKUPs and pivot tables.

With Claude Code:

Take klaviyo-segment.csv and shopify-customers.csv. For every row in shopify-customers.csv where the email also appears in klaviyo-segment.csv, output a new file `tagged-customers.csv` with the original Shopify columns plus a new column `klaviyo_segment` set to the value from klaviyo-segment.csv.

After: print the row count and the top 5 rows so I can sanity-check.

Claude writes a small Python script, runs it, returns the file and the sample. 90 seconds total.

4. Batch image renaming and resizing

A client sends 47 product photos named IMG_3941.jpg through IMG_3988.jpg. The VA needs them renamed (product-slug-01.jpg through product-slug-47.jpg), resized to a Shopify-friendly 2048x2048, and converted to webp.

In ./raw-images, rename each .jpg sequentially as product-slug-NN.jpg where NN matches the file's order in the folder when sorted by name. Resize each to max 2048 on the longest side. Convert to webp at quality 82. Output to ./processed-images.

After: print the count of input vs output and any files that failed.

ImageMagick + a one-line bash loop, written and executed in 20 seconds.

5. Weekly metrics summarisation

Every Monday our VAs send a one-page summary to each client: orders, support tickets, returns, social engagement, key KPIs from the last 7 days.

Pre-Claude: 25-30 minutes per client.

With Claude Code, the VA drops the raw exports into a weekly-data/ folder and runs:

Read every CSV in ./weekly-data/. For each, identify the metric type (orders, support tickets, returns, social engagement) by the column headers. Output a one-page markdown summary in our standard format:

## Last 7 days at [client name]

**Sales**
- Orders: [N] (vs prior 7d: ±[%])
- Revenue: $[X] AUD (vs prior 7d: ±[%])
- AOV: $[X] (vs prior 7d: ±[%])

**Support**
- Tickets opened: [N]
- Tickets closed: [N]
- Median first-response time: [N] hours

**Returns**
- Returns processed: [N] ($[X] AUD)
- Top reason: [reason]

**Notes**
- Anything anomalous worth flagging?
- Anything that broke vs the prior week?

Australian English. No em-dashes. Numbers in AUD.

10 minutes per client, including the human sanity-check pass.

What Claude Code is bad at

The honest list, so you have realistic expectations.

Anything requiring a current credential. It cannot log into Shopify, Klaviyo, or your CRM. Your VA still does the click-throughs in the browser.
Long, complex, multi-system workflows. It is great at one well-scoped task. It struggles with “run this thing across four tools and reconcile the outputs”. Break the task down.
Anything visual. It cannot see designs, fonts, or layouts. Stick to text and data tasks; use Figma, Canva, or the human eye for the visual stuff.
Highly proprietary domain reasoning. It does not know your business well enough to make judgement calls. It executes the workflow you teach it.
Cost reasoning at scale. For VAs running 100+ commands a day, you will start to feel the API-quota wall. Claude Pro is enough for our team’s normal use; if a VA is hitting limits, the workflow is wrong.

What we would change if we started again in 2026

Standardise the folder structure across clients. Each client gets a folder with: sop/, templates/, weekly-data/, drafts/, outputs/, and a CLAUDE.md at the root that briefs Claude on the client’s voice, rules, and current priorities.
Build a CLAUDE.md for the VA, not just the AI. A short doc the VA reads when they take over a client account. It mirrors what Claude sees so the VA and the AI are working from the same playbook.
Train every VA in week 1. Not in week 4. The compound benefit of AI fluency hits within month one. Delaying training is delaying ROI.
Be ruthless about the no-AI tasks. Customer complaint tone, sensitive escalations, anything TGA/ACCC-adjacent. Keep these in the human’s lane forever.

The skill we actually teach

The skill we teach is not “Claude Code”. It is:

How to write a prompt that wins on the first try. (Specific role, specific output format, at least one example, an escalation rule.)
How to recognise when Claude is bluffing. (Vague numbers, generic phrasing, references to sources Claude has not actually checked.)
How to break a 30-minute task into a 3-minute prompt + a 27-minute review. (The review is the value-add.)

Our prompt library for AU VAs is the starter pack. Our why-we-dont-replace-vas-with-ai post is the philosophical companion piece.

What this means for hiring

A VA who is fluent in Claude Code is worth roughly 1.5x a VA who is not, at the same hourly rate. We train every VA on placement so the multiplier flows to the client, not to us.

If you are evaluating a VA agency in 2026 and they do not train on AI tooling, you are looking at last decade’s offer. Ask the question.

What’s next

For the prompt library that drives daily use, see Prompt library for Australian VAs.

For the Shopify-specific workflows that we run with Claude every day, see Shopify automation for VAs.

For the contrarian take on where AI loses, see Why we don’t replace VAs with AI.

If you want a placement that comes with Claude Code training in week one, book a discovery call.