If How we run our VA team on Claude Code is the strategy piece, this is the tactical companion. Five commands you can copy-paste tomorrow.

What you need before you start

A laptop with Node.js installed (node --version returns something; if not, get it from nodejs.org)
A Claude Pro account ($20/mo)
A folder of work files (SOPs, exports, drafts, whatever)
5 minutes to install

Installation, in your terminal:

npm install -g @anthropic-ai/claude-code
claude login

Follow the prompts. You’ll authenticate via browser. Done.

To use it, cd into your work folder and run claude. You get a chat interface that has direct access to the folder. Type your prompt.

Command 1: Audit a folder

When: quarterly SOP review. New VA onboarding to a client. After a tool change.

Prompt:

Walk this folder and check every file. For each markdown or text doc:
- Does it reference any tools we don't use anymore? (Cross-check against ./active-tools.md)
- Are any internal links pointing at files that no longer exist?
- Does anything reference a person no longer on the team? (Cross-check against ./team.md)
- Are any "last updated" dates older than 6 months?

Output a markdown report grouped by file, with file:line references.

What it does: Claude reads every file, finds stale references, produces a punch list. Takes 30-90 seconds for a folder with 20-40 documents. The same audit done manually takes 45-60 minutes.

Failure mode: Claude assumes things exist when they don’t. Verify before deleting or “fixing” anything.

Command 2: Clean up data

When: the boring CSV work. De-duplicating customer lists. Merging exports from two tools. Reshaping data for an upload.

Prompt:

Take customers-shopify.csv and customers-klaviyo.csv. For every row in
shopify.csv where the email also appears in klaviyo.csv:
- Output a new file `cleaned-customers.csv` with the original Shopify
  columns plus a new column `klaviyo_lists` listing all the lists that
  email is in from the Klaviyo file
- Skip any row where the Shopify customer has unsubscribed
- Print the row count and the top 5 rows so I can sanity check

If the input files have problems (different column headers than expected,
encoding issues, etc), flag those before processing.

What it does: Claude writes a small Python or shell script, runs it, returns the file plus a sanity check. About 90 seconds for files with 5-10k rows.

Failure mode: Claude writes the script confidently for a file structure it hasn’t seen. Always paste the first 3 rows in the prompt so Claude can pattern-match the actual structure.

Command 3: Batch-process files

When: renaming images, resizing photos for Shopify, converting file formats, organising photos by date.

Prompt:

In ./raw-product-photos/, rename each .jpg file sequentially using the
naming pattern from product-list.csv where:
- column 1 is the current filename (without extension)
- column 2 is the new product slug
Then resize each renamed photo to a max 2048 on the longest side, convert
to webp at quality 82, output to ./processed/

After: print the input/output counts and any files that failed.

What it does: Claude uses ImageMagick or similar to batch the operation. About 30-60 seconds for 50 photos.

Failure mode: Claude assumes the conversion library is installed. On a fresh machine, you may need to run brew install imagemagick first. Tell Claude what’s installed if there’s any doubt.

Command 4: Summarise long documents

When: Loom transcripts. Meeting notes. Long email threads. PDF reports.

Prompt:

Read all the .txt files in ./transcripts/. For each:
- Identify the meeting type (sales call, internal, client review) from the
  content
- Extract action items (with owner: me / VA / client / unassigned)
- Extract key decisions
- Flag any commitments to deadlines

Output a single markdown summary, one section per transcript, in the order
the files were created (oldest first).

Use Australian English. No em-dashes.

What it does: Claude reads each transcript file and produces a structured digest. For 5-10 transcripts, takes 1-2 minutes vs the 30-60 minutes the VA would spend doing it manually.

Failure mode: Claude invents action items if the transcript is ambiguous. Train your VA to verify by sampling 2-3 entries against the source.

Command 5: Draft from templates

When: weekly client reports. Vendor reports for real estate vendors. Recap emails after meetings. Anything templated.

Prompt:

Use the template at ./templates/weekly-vendor-report.md.

Fill in the variables from ./data/vendor-week-data.csv where:
- {{property_address}} = column "address"
- {{ofi_count}} = column "ofi_attendees"
- {{enquiry_count}} = column "enquiries"
- {{portal_views}} = column "views_total"
- {{listing_age_days}} = column "days_listed"
- {{market_temp}} = derived from "enquiry_count / portal_views" — write one
  short sentence, e.g. "Engagement is tracking ahead of similar listings"
  if rate > 1.5%, "in line with similar listings" if 1-1.5%, "below similar
  listings" if <1%

Output one filled report per row, in ./reports/ named by property address
(slugified).

Australian English. No em-dashes.

What it does: Claude generates as many reports as you have rows. 12 reports in 2 minutes vs 90 minutes manually.

Failure mode: Claude invents data when a row is missing a value. Train your VA to spot null cells in the source CSV before running.

What Claude Code is NOT for

The honest list:

Anything requiring a current login. Claude Code cannot log into Shopify, Klaviyo, or your CRM. Your VA still does click-throughs in a browser for live tool work.
Long multi-tool orchestrations. Great at one well-scoped task. Bad at “run this across four tools and reconcile the outputs”. Break it down.
Anything visual. Cannot see designs, cannot evaluate layouts. Stick to text and structured data.
Highly proprietary domain reasoning. Doesn’t know your business. Executes the workflow you teach.

If you have heard “agents will do everything in 2027” — maybe. In 2026, Claude Code is a focused tool for focused tasks, and that’s where the productivity gain is real.

The skill, in one sentence

The skill we teach our VAs is knowing when to use Claude Code versus the Claude web chat versus doing it manually. Each has a sweet spot. Each is wrong for some tasks.

Web chat: one-off prompts, no files involved, fast.
Claude Code: file-touching tasks, repeatable workflows, structured outputs.
Manual: novel context, sensitive judgement, anything customer-facing requiring real-time tone.

A VA fluent in all three is roughly 1.5-2x more productive than a VA who only uses the web chat.

What’s next

For the strategy piece, see How we run our VA team on Claude Code.

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

For the SOP format that works with Claude Code, see Writing SOPs your VA AND Claude can both follow.

If you want a placement that comes with Claude Code already running, book a discovery call.