Most VA agencies and AU founders we talk to are now provisioning Claude or ChatGPT for their assistants on day one. Fewer are actually training them. The result: VAs with two new tabs open who barely use them, or VAs who use them too eagerly and ship hallucinations to customers.

Here is the exact five-day plan we run with every new placement.

Before day one

The VA needs:

A Claude Pro account ($20/mo, paid by us)
A ChatGPT Plus account ($20/mo, paid by us)
Access to our internal prompt library (the public version is at /ai-with-vas/prompt-library-for-australian-vas-40-prompts/)
A blank “AI practice” Google Doc to log their day 1-5 outputs

Total cost to us: $40/mo per VA. Total return per VA: 5-8 hours of recovered productivity per week.

Day 1 — Prompt structure (30 min)

Objective: the VA understands why some prompts win first-try and others don’t.

Teach (15 min via Loom, recorded once, reused):

The four parts of a winning prompt:

Role (“You are a customer support specialist for [brand]”)
Task with specific constraints (“draft a reply under 80 words, Australian English, no em-dashes”)
Example of what good looks like (one paste-in)
Escalation rule (“if this requires a refund decision, output [ESCALATE]”)

Show them three before/after prompt examples:

Before: “write a reply to this customer” → generic AI mush
After: full 4-part prompt → useable output

Practice (15 min):

Three drills. VA picks one of their actual current tasks (an email reply, a product description, a meeting summary) and:

Writes a “before” version of their prompt
Writes an “after” version using the 4-part structure
Runs both, pastes outputs into their AI practice doc
Sends both versions + outputs to their DotVA team lead by end of day

Pass criteria: the “after” output is visibly more usable than the “before”. Team lead reviews and either approves or asks for a redo.

Day 2 — Customer-facing writing (45 min)

Objective: the VA can draft a customer email, a Klaviyo flow email, and a product description, in your brand’s voice, in under 5 minutes each.

Teach (15 min):

Show the prompt library entries for:

Customer support first-line reply (prompt #11)
Klaviyo flow copy (prompt #5 of Shopify)
Product description (prompt #22)

Walk through the bracketed placeholders — what to fill in, what to leave.

Practice (30 min):

Real client work. Pick three actual customer-facing pieces from the client’s backlog. VA drafts each using the relevant prompt. VA edits each before submitting.

Submit to team lead with: the prompt used, the AI output, the VA’s final edit, and a note on what they changed and why.

Pass criteria: the VA’s final edits are noticeable improvements on the AI output (tone, specific detail, removed cliches). If the final edit is identical to the AI output, the VA is rubber-stamping, which is failure mode #1.

Day 3 — Data + admin (45 min)

Objective: the VA can use AI for CSV wrangling, inbox triage, and meeting summaries.

Teach (15 min):

The “Claude is good at structured text” lesson. Show:

A CSV merge task using the prompt #29 pattern
An inbox triage task using prompt #1
A meeting note extraction from a transcript using prompt #4

Emphasise: AI doesn’t replace your spreadsheet skills. It’s faster than VLOOKUPs for one-off tasks.

Practice (30 min):

Three real tasks from the client’s queue:

One CSV reshape (de-duplicating a customer list, pivoting a Klaviyo export, etc)
One inbox triage on yesterday’s emails
One meeting note extraction from a Loom transcript

Pass criteria: outputs are accurate. We sense-check the CSV row counts and email counts. Inaccuracies = redo.

Day 4 — The judgement layer (60 min)

Objective: the VA can identify which tasks should NOT go through AI and which should.

Teach (20 min):

Walk through Why we don’t replace VAs with AI. Specifically the seven losses:

Escalation
Ownership over time
Novel context
Ambiguous priorities
Sensitive communication
The soft no
Follow-through

For each, give a worked example from a real placement.

Practice (40 min):

Show the VA 10 real tasks (mix of AI-suitable and not). For each, they must:

Decide: AI-draft / AI-skip
If AI-draft: write the prompt
If AI-skip: write the manual approach

We expect 8-9 out of 10 to be classified correctly on day 4. Below 7 = failure mode #2 (over-reliance on AI), and we run a second week of practice.

Day 5 — Failure mode recognition (45 min)

Objective: the VA recognises when AI is bluffing, hallucinating, or off-tone, and corrects before shipping.

Teach (15 min):

The three failure modes:

Confident invention. AI states a number, source, or fact that wasn’t in the prompt. Always verify against the input.
Tone drift. AI defaults to US conversion-copy tropes (“don’t miss out”, excessive urgency). Strip these.
Generic phrasing. AI gives an answer that could apply to any business. Re-prompt with more specific context.

Practice (30 min):

Three drills, each with a deliberately AI-broken output. VA must identify the failure mode and fix.

Drill 1: AI output containing an invented testimonial. VA must spot it.
Drill 2: AI output with US tone defaults. VA must rewrite to AU voice.
Drill 3: AI output that’s generic. VA must add specificity using the client’s actual context.

Pass criteria: VA spots all three failures and produces an acceptable corrected version.

Day-5 competency check

After the day-5 drills, the VA produces three outputs unsupervised:

A customer support reply for an actual customer email (must pass the “would Jenn send this?” test)
A spreadsheet reshape on real client data (must be accurate)
A judgement call on whether a third task should use AI or not (must match what the team lead would decide)

Pass → VA goes into normal placement rhythm in week 2, with weekly AI quality spot-checks.

Fail → second week of supervised practice. We do not push a VA to “go live with AI” until they pass the check.

Failure modes we teach for explicitly

Three patterns we see in untrained VAs:

The eager paste-through. VA gets AI output, pastes directly into the customer email, sends. Caught either by us at QA or by a customer review complaint. Fix: every AI output is edited before shipping. “Edit” must be visible — at least 2 sentences different from the AI version.

The fearful avoider. VA opens Claude once on day 1, gets a mediocre output, decides “I’m faster on my own”. Goes back to manual work for the rest of the placement. Caught at the 30-day review when their output velocity is below benchmark. Fix: practice forced via assigned AI-augmented tasks for the first 30 days.

The over-trustor. VA accepts AI’s numbers, dates, and source attributions as facts. Ships a product description with a fabricated study citation. Caught when the client notices. Fix: mandatory “verify before publishing” for every output containing a specific claim.

What we don’t teach in week one

Custom GPTs / Claude Projects (week 4 if relevant)
Image generation workflows (week 3 if the role needs it)
Code-adjacent Claude Code workflows (week 4 for VAs who’ll use them)
Advanced agent or multi-step automations (month 2)

Week one is foundational. Layer the rest on once the foundation is solid.

Why we pay for both tools from day one

The temptation to save $20/mo by giving the VA only one AI subscription is real. We don’t.

Two reasons:

The VA needs to learn the strengths of each (see Claude vs ChatGPT for VAs). Forcing them to pick one tool means picking a tool that’s wrong for half the tasks.
The $40/mo per VA is roughly 0.3 per cent of the placement value. The downside of an under-trained VA shipping a bad output is much higher than the saving.

Pay for both. Train on both. Win on both.

What’s next

For the working prompt library, see Prompt library for Australian VAs.

For the SOP format that pairs with AI execution, see Writing SOPs your VA AND Claude can both follow.

For the failure modes module in more depth, see AI hallucination guardrails for VAs.

If you want a placement that comes with this training week already complete, book a discovery call.