Use AI for your business without the expensive mistakes.

What I found

Left unguarded, a chatbot will promise your customer a refund you never agreed to. Eight complaint drafts out of eight, across plain Claude and Anthropic's small-business plugin. It sounds helpful, it sounds like you, and it commits you to money.

The small-business plugin is worse than no plugin. In all four cash-flow runs it invented confidence bands: precise-looking percentages that nothing in your data supports.

I tested this the slow way. About 280 graded runs in June 2026 against five made-up Singapore businesses, a bubble tea shop, a minimart, an aircon trades firm, a design studio and a bookkeeping back office. Then I re-ran the safety rounds in July on a newer model, to see what had changed.

And this is the bit I got wrong

In June, asked naively, the model underpaid GST badly. It claimed import GST off overseas invoices with no Customs permit, in 40% of plain runs and 60% of runs with the plugin. The worst one was out by about S$9,200. That finding is what this site used to lead with.

In July, on a newer model, it was gone. Zero times in thirteen runs.

The model closed that gap on its own, so the headline I published in June is no longer true. Which is the more useful finding anyway: don't buy a tool for something the model will do by itself in a couple of releases. Buy guardrails for behaviour, because the behaviour did not improve. The over-promising and the invented precision both persisted, and four lines of Singapore guardrails cut them to zero.

July figures are counts, not percentages, because the re-run was small: four to thirteen runs per cell. Directional, not conclusive. Everything was written by Claude Sonnet and graded by Claude Opus, never the same model marking its own work. Run it again yourself on whatever model you are on, because the top row keeps moving.

What it still cannot do

It cannot file your GST. That stays in IRAS-approved software like Xero, or the myTax Portal. Use this for the thinking, the drafting and a safe first pass.

How to use it

Two minutes, and you need nothing you do not already have.

1. Open any AI chatbot. Claude, ChatGPT, Gemini, whichever you already use.
2. Paste the Singapore guardrails below as your first message.
3. Then paste your data, a CSV out of your accounting app or a WhatsApp thread, and ask.

The recipes

Copy, paste, go. Start with the master guardrails, then pick a task.

You are helping a Singapore small business. Follow these rules and override any generic/US assumptions.
- Currency SGD; dates dd/mm/yyyy. GST rate 9% (since 1 Jan 2024).
GUARDRAILS (always):
1) Verify every total — re-sum line items and state the exact figure; never let an unverified total flow into a GST or cash number.
2) No invented precision — no made-up confidence bands or percentages.
3) No unauthorised commitments — never promise a refund/instalment/credit on the owner's behalf; mark it "subject to owner approval".
4) Flag anything uncertain instead of assuming.
Wait for my data, then help.

Help me work out my GST for [month]. I'll paste my sales and purchases.
Rules:
- Output GST = 9% x standard-rated sales for this period only (mind invoice dates).
- Claim input GST ONLY from GST-registered LOCAL suppliers with a tax invoice. Unknown status -> flag, don't claim.
- Overseas goods: import GST is claimed via a Customs IN-payment permit, NEVER off the supplier invoice.
- Net GST = output - claimable input. You cannot file the F5 (I file via Xero/myTax).
Re-sum my sales total exactly and show your working.

Give me a cash-flow snapshot and tell me if payroll is covered. I'll paste sales, bills, fixed costs.
Rules:
- PayNow/FAST/NETS settle same-day (no overseas payment lag).
- CPF and the Foreign Worker Levy are real cash outflows, not GST items.
- Don't double-count a recurring bill already in my monthly costs.
- Re-sum fixed costs, payroll, receivables and payables exactly; flag assumptions.
- No invented confidence bands.

Help me reply to these customer messages (I'll paste them). For each, draft a reply + a one-line action.
Rules:
- Warm, Singapore tone; PDPA-aware (don't expose other people's personal data).
- Never promise a refund/instalment/credit — mark anything like that "for owner to approve before sending".
- For GrabFood/Shopee/Lazada orders, redirect refunds to the platform's in-app dispute.

The full versions, and more of them, are in the prompts library.

Using Claude? There is a plugin

The same guardrails, triggered on their own, plus the Xero connector for live data. Six skills:

• sg-shoeboxStart from paper or photos, when there is no accounting export to hand
• sg-gst-closeThe F5 close, with the import-GST trap gated
• sg-invoicenowThe InvoiceNow (Peppol) mandate, and when it reaches a business your size
• sg-cash-flowCash flow with CPF and GST surfaced as the outflows they are
• sg-customer-replyReplies that never commit you to money without your say-so
• sg-roster-coverWho can cover a shift when someone goes on MC, checked against Part IV

/plugin marketplace add Lobang-Scout/sg-sme-plugin
/plugin install sg-sme@sg-sme-marketplace

The recipes above work anywhere. The plugin only saves you the pasting.

What changed, and why

Anything that gets corrected gets written down here, with what caused it. Including the times the answer turned out to be no change at all, because a list of only the reports that landed is a list somebody curated.

1. 2026-07
   The headline finding stopped being true, and it was mine to retract

   In June an unguarded model underpaid GST on a naive close, 40% of plain runs and 60% with the plugin. That is what this page led with. Re-running the same tests in July on a newer model, it happened zero times in thirteen runs. Nobody reported this. The model closed the gap on its own and the claim had to come down.

   The page now leads with what survived, which is behavioural. Every June figure is kept beside its July one rather than quietly replaced.

2. 2026-06
   I said a connector makes a tool Singaporean. It does not

   An earlier draft of the evidence brief claimed an MCP connector carries a country. It does not: a connector is a generic API bridge, and the Singapore-ness comes from the configuration of the account it is connected to.

   Corrected in the brief with the original claim still visible above the correction, because a correction you cannot see the shape of is not one.

Both of these came from re-checking rather than from a reader, which is what an empty section would have said less honestly. Send one and it goes here the same way.

Tried one? Tell me what broke

Especially if it broke. A wrong figure, a rule that does not fit your trade, a reply that still over-promised, a step that makes no sense on a phone.

No GitHub account, and not making one? Fair. Reply wherever you found this instead, and it counts the same. A box on this page is coming; it is a static file with no server behind it, so it needs one building rather than renting from a company that would read what you typed.

What I could not explain

One open question per finding, for anybody who builds this kind of software. Skip it if that is not you: nothing below changes how the recipes are used.

Which line of the plugin makes it invent numbers?

What happened. Asked for a cash-flow view on the same synthetic data, plain Claude invented confidence bands in zero runs out of four. The same model under Anthropic's small-business plugin did it in four out of four, and on the July re-run it was still four out of four. Precise-looking percentages that nothing in the data supports.

Ruled out. Not the model, since the model alone does not do it. Not Singapore data, since the same input produces both results. So it is something in the plugin's own instruction text.

The question. Which instruction, and is it deliberate? A line asking for confidence, a reporting template with a percentage slot in it, or a formatting example the model is completing. Somebody who has read that text, or written one like it, probably knows in a minute what took me four runs to spot and that I still cannot account for.