Back to Resources
Prompt Engineering10 min read
Singapore System Prompts for Claude
Reusable system prompts that carry Singapore context — GST, CPF, PDPA, local spelling, SGD, and the regulator names — so you stop correcting the same four things every session.
The four corrections you're making every session
If you work in Singapore and use Claude daily, you're probably correcting the same things repeatedly:
- Currency defaulting to US dollars, or worse, being ambiguous about which dollar
- American spelling — *organize*, *analyze*, *center*
- Employment and tax assumptions imported from elsewhere — 401(k) instead of CPF, sales tax instead of GST
- Regulator and framework names that don't apply here — GDPR instead of PDPA, SEC instead of MAS
Each correction takes seconds. Multiplied across a working week it's a meaningful tax, and it introduces errors when you miss one in something you send.
A system prompt fixes it once. What follows are the ones we actually use, in ascending order of specificity.
The baseline — for everything
Short enough to keep in your profile settings permanently:
*"I am based in Singapore and work with Singapore businesses. Default to: British/Singapore English spelling (organise, analyse, centre, licence as a noun); SGD written as S$; dates as 5 August 2026; Asia/Singapore time. Assume Singapore statutory context unless I say otherwise — GST not sales tax, CPF not 401(k), PDPA not GDPR, MAS not SEC, IRAS not IRS, ACRA for company matters. If something genuinely differs by jurisdiction and you don't know mine, ask rather than assuming a US default."*
The final sentence does the heavy lifting. Without it, a model that doesn't know will pick the most statistically common jurisdiction, which is not this one.
For professional services
Adds tone and the disclaimers that matter:
*"I run a professional services firm in Singapore. Write in plain English — no consultant-speak, no 'leverage', no 'synergies'. Prefer short sentences and specific nouns. When advice touches law, tax or accounting, note where a qualified professional should confirm it rather than stating it as settled. Never invent a statute name, section number, or case citation — if you're not certain it exists, say so. Refer to 'the firm' rather than 'our team'."*
Never invent a citation is the one to keep even if you drop everything else. Fabricated case references are the single most damaging failure mode in professional use, precisely because they look correct.
For SME operations and finance
*"I run operations for a Singapore SME (under 50 staff). Practical over comprehensive — I'd rather have three things I can do this week than a framework. Assume limited budget and no dedicated IT. When you suggest a tool, note roughly what it costs in SGD and whether there's a free tier. For anything involving figures that end up in a filing or an invoice, do not compute — tell me what to calculate and where, and write the narrative around the numbers I give you."*
That last instruction reflects a real limitation. Arithmetic across many rows is where language models are least reliable and the errors are hardest to spot. Compute in your system; use Claude for the sentences.
For anything touching personal data
*"Everything in this conversation may involve personal data under Singapore's PDPA. Before using any identifier — NRIC/FIN, full name, address, phone, bank account, medical detail — tell me it should be redacted rather than working with it. If I paste something containing identifiers, flag it before continuing. Never repeat an identifier back to me in full. Assume 'reasonable security arrangements' is a standard I have to meet, and tell me when something I'm asking for would make that harder."*
This one is deliberately obstructive. It's meant to interrupt you, because the moment personal data ends up somewhere it shouldn't is almost always a moment nobody noticed.
For Claude Code — the CLAUDE.md version
In a repo, the same job is done by a
CLAUDE.md file in the root. Claude reads it before doing anything:
``markdown
# Project conventions
- Singapore English spelling throughout, including in code comments and UI copy.
- Currency is SGD, formatted S$1,234.56.
- Dates display as "5 Aug 2026"; store as ISO 8601 with an explicit +08:00 offset.
- Timezone is Asia/Singapore. Never store a local time without an offset.
- Personal data: no NRIC/FIN in logs, ever. Not even redacted.
## Things I don't want
- No new dependencies without asking.
- Don't refactor code I didn't ask you to touch.
- Don't add tests to throwaway prototypes; do add them to anything with money in it.
``
The *"Things I don't want"* section is consistently the most useful part. Constraints are signal — a model told what to avoid produces better output than one given only goals.
One warning worth stating: don't write *"no database"* as a permanent constraint. It's a common line in starter templates, and it stops Claude suggesting one at the point you genuinely need one. Write *"no database yet"* instead.What system prompts can't fix
Three honest limits.
Drift. Over a long conversation, particularly after you paste in American-spelled material, conventions slip. Do a final pass on anything you're sending.
Currency ambiguity in source data. If your spreadsheet says "$1,200" the system prompt can't tell you which dollar that was. Fix it at source.
Recency. Statutory thresholds, GST rates and grant schemes change. A system prompt that names a rate will be wrong eventually — name the concept, look up the number.
Where to take this next
For the task-level prompts these sit underneath, see The Singapore Prompt Library. For getting started with Claude Code specifically, see Claude Code in Singapore: A Non-Developer's First-Week Guide. For handling personal data properly, see the PDPA Prompting Checklist.
Session 1 of the Build School cohort covers writing your own CLAUDE.md as pre-work.
Frequently asked questions
What's the difference between a system prompt and a normal prompt?
A system prompt sets standing context that applies to everything in the conversation — who you are, what conventions to follow, what never to do. A normal prompt is the task. Putting Singapore context in the system prompt means you stop restating it.
Where do I put these in Claude.ai?
Settings → Profile lets you set persistent preferences that apply across chats. For per-project context, use a Project and put the system prompt in the project instructions. In Claude Code, the same job is done by a CLAUDE.md file in the repo root.
Will this stop Claude using American spelling?
Mostly. Spelling drifts back over long conversations, particularly after you paste in American-spelled source material. The reliable fix is a final pass: "check this for American spellings and fix them."
Should I put confidential company details in a system prompt?
Standing context like your industry, entity type and conventions is fine. Client names, NRIC numbers and financial specifics are not — they persist across every chat, which is exactly what you don't want for sensitive data. See our [PDPA Prompting Checklist](/resources/pdpa-prompting-checklist).
PreviousClaude for Strategic Advisory: Frameworks for Useful AI OutputNext AI Governance and the EU AI Act: What Singapore Businesses Actually Need
Want to Apply This to Your Business?
We're a Singapore AI development and automation agency. Let's discuss how we can help solve your specific challenges.