Back to Resources
Claude AI11 min read

Five Claude Workflows That Eliminate Daily Admin

The recurring admin most Singapore SMEs still do by hand — inbox triage, meeting prep, quote drafting, reconciliation chasing, weekly reporting — and the exact prompts that remove them.

Haojun See
Haojun See

Founder & Director, On The Ground

Updated 5 August 2026

The admin that survives every efficiency drive

Ask a Singapore SME owner where their week goes and the answer is rarely the work they're paid for. It's the sediment around it: reading an inbox to find the three things that matter, reconstructing what happened in a meeting last month, retyping the same quote with different numbers, chasing a supplier who hasn't replied, and assembling a Friday update nobody asked for but everyone expects. None of that is hard. All of it is expensive, because it eats attention in small pieces at exactly the times you needed to concentrate. These five workflows are the ones we see removed most often, in roughly the order they're worth doing. Each one is a prompt you save once and reuse. None requires a developer.

1. Inbox triage — decide once, in one pass

The problem isn't volume, it's re-reading. Most people open the same email four times before acting on it. The fix is to classify everything in one pass, then work the list. Connect Gmail, then run this: *"Read my unread email from the last 24 hours. Classify each into exactly one of: REPLY-TODAY (someone is blocked on me), REPLY-THIS-WEEK (matters, not urgent), FYI-NO-ACTION (read and forget), IGNORE-OR-UNSUBSCRIBE. For REPLY-TODAY only, draft a two-sentence reply in my voice. Give me a table, newest first. Do not send anything."* Two details make this work. Exactly one category forces a decision — without it you get things filed in two places and you're triaging again. And draft only the urgent ones, because drafting all of them produces more reading, not less. The habit that makes it stick: run it once, in the morning, and then close your inbox. If you run it four times a day you've built a more elaborate version of the problem.

2. Meeting prep — fifteen minutes of context in ten seconds

Walking into a meeting and remembering who someone is, what you last agreed, and what's outstanding takes about fifteen minutes of digging. It's also the prep most often skipped, which is why meetings restate things. With Calendar and Gmail connected: *"My next meeting is [name/time]. Give me: who's attending and their role, the last three exchanges I've had with them, anything I committed to that I haven't done, and two questions worth asking. Half a page maximum. Flag anything that looks unresolved."* The last instruction is the one that earns its place. *"Flag anything that looks unresolved"* surfaces the thing you'd have remembered halfway through the meeting. The anti-pattern: asking for a summary of the meeting's agenda. You wrote the agenda. What you need is the history you've forgotten.

3. Quote and proposal drafting — the same document, different numbers

Most SMEs send a variation of the same five documents forever. The variation is real — different scope, different price — but the structure never changes, and retyping structure is pure waste. Do this once: paste your best recent quote and say *"extract the structure of this as a reusable template. Mark everything that changes per client as [SQUARE BRACKETS]. Don't include this client's specifics."* Save the result. Then each time: *"Using the attached quote template, draft a quote for [client] for [scope]. Price [amount]. Their situation: [two sentences]. Match the tone of the template exactly. Leave anything you're unsure about as a bracketed placeholder rather than guessing."* That last sentence prevents the failure mode that makes people abandon this: Claude inventing a plausible delivery date or a payment term you don't offer. Told to leave a gap, it leaves a gap.

4. Chasing what's outstanding — the reconciliation nobody enjoys

Unpaid invoices, unreturned deposits, POs without a delivery note. The information exists, in a spreadsheet or an export, and the work is comparing two lists and writing polite emails. This is the one workflow worth doing in Claude Code rather than the browser, because the output should be a file you keep, not text you read. Export both lists to CSV, then: *"Compare invoices.csv against payments.csv on invoice number. Produce outstanding.csv with: invoice number, client, amount, days overdue, and a MATCH / SHORT / MISSING status. Then draft one chase email per row over 30 days overdue — firm, not aggressive, no apology for chasing. Save the drafts as a single markdown file. Don't send anything."* *"No apology for chasing"* matters more than it looks. Left to its own devices, an AI writes *"sorry to bother you about this"* on an invoice that's sixty days late.

5. The weekly update — written from evidence, not memory

The Friday summary is usually written from whatever you happen to remember on Friday, which is why it over-reports the last two days. Set it as a recurring task rather than a prompt you rerun: *"Every Friday at 4pm: review my sent email, calendar, and any documents I edited this week. Write a five-bullet summary: what moved, what's stuck, what I committed to, what needs a decision from someone else, and what slipped. Base it only on what you can see — if the week looks quiet, say so rather than padding it."* The final clause is the whole point. A summary that inflates a quiet week is worse than no summary, because you stop trusting it. Prune these ruthlessly. A recurring task you skim and dismiss is costing you attention every week. If you stop reading it, delete it.

What not to automate

Three honest exclusions, because the failures are as instructive as the wins. Anything that sends without you. Every workflow above stops at a draft. Automated sending fails rarely, but the failures are the kind you have to apologise for. Anything you do monthly or less. The instruction takes longer to write than the task takes to do, and you'll have forgotten the prompt exists by the time it recurs. Anything where being wrong is expensive and hard to notice. Arithmetic on figures that feed a filing, for instance. Compute in your system, and use Claude for the narrative around the numbers. The pattern across all five: Claude removes the blank page and the cross-referencing. It doesn't remove the judgement, and the workflows that try to are the ones people quietly stop using.

Where to take this next

For the full cross-industry prompt set, see The Singapore Prompt Library. For prompts that carry Singapore business context properly, see Singapore System Prompts for Claude. If you'd rather build the tool than run the prompt, the five-session Build School cohort covers going from nothing to a deployed, instrumented app — including what to do when it breaks. If you want us to build it instead, that's Groundwork.

Frequently asked questions

Do these workflows need Claude Code, or just Claude.ai?

Four of the five run in Claude.ai with connectors — no code, no terminal. Only the reconciliation-chasing workflow benefits from Claude Code, because it reads a spreadsheet export and produces a file. Start with the browser; move to Claude Code when you want the output saved rather than read.

How long before these actually save time?

The first run of any of these takes longer than doing the task by hand, because you're writing the instruction. The saving starts on the second run. If a task doesn't recur at least weekly, don't automate it — write it down and move on.

Is it safe to point Claude at my company inbox under PDPA?

Connectors read data on your behalf under your own Google permissions, and Claude Pro/Team/Enterprise exclude your inputs from training by default. The judgement call is scope: connect the account you actually work from, not a shared admin mailbox holding other people's personal data. See our [PDPA Prompting Checklist](/resources/pdpa-prompting-checklist).

What if the output is wrong?

Assume it will be, sometimes. Every workflow below produces a draft for you to accept, not an action it takes on your behalf. That's deliberate — the value is in removing the blank page, not in removing the review.

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.