Apeiron AI Pocket Guide

Right at the start

Your worst task this week

One task you do every week and would happily never do again. One line is enough. It shows up on the screen at the front.

Draftletters, emails, follow-ups

Extractkey points, figures from documents

Checkone document against another

Analysevariances, unusual movements

We'll fix some of these today and some halfway. Some aren't AI problems at all; those go on the backlog with an owner.

The hour

Five parts, one hour

You'll leave having used AI on a task like yours, caught it being wrong, and saved one prompt you'll use again.

How it works

Three things to know

It guesses the next word

It knows what sounds right, not what's true. That's why it's fluent, and why it can be confidently wrong.

It only sees the window

Your files, your prompt, the chat so far. Put the document first and the question last. New task, new chat.

It forgets when the chat ends

Unless you give it a memory: a Project (Claude, ChatGPT) or a Gem (Gemini) with standing instructions and your files.

One way to ask well

Task: Start with the verb. Say what you want back.
Context: What the material is, and why you need it.
Format: Name the columns or the length.
Example: One row of what good looks like.
Escape hatch: "If it's not stated, write 'not stated'. Don't guess."

The prompts from the demo open later, under Prompts.

Exercise · on your own · 10 minutes

Three documents, two rounds

A supplier, Golden Wok, sent us three documents for one collection. Something doesn't match. Copy each document's text into your AI tool, then ask.

Round 1: write any prompt you like. Round 2 opens here when we get to it. All names and figures are made up for training.

Document A · Supplier declaration

Page 1 of 1

DOCUMENT A: SUPPLIER SELF-DECLARATION OF USED COOKING OIL ORIGIN (page 1) Issued to: Apeiron Bioenergy Pte Ltd (collector). Reference: GW-DECL-2026-03 Supplier (legal name): Golden Wok Food Court Pte Ltd UEN: 201912345K Collection point address: Blk 512 Bedok North Street 3, #01-88, Singapore 460512 Contact person: Mr Tan Boon Hock, Operations Manager Type of business: Food court operator, 14 cooked-food stalls Material declared: Used cooking oil from food preparation at the above premises Estimated volume: Approximately 2,500 litres per month Storage: Sealed 200-litre drums and two 120-litre wheeled bins, covered area Declaration: the oil is a waste from frying and cooking, not mixed with virgin oil, animal fat or any other substance, and comes only from the premises stated. Signed: Tan Boon Hock, Operations Manager. Date: 15 March 2026. Valid until 14 March 2027. Note: this declaration covers origin only. Quantities per collection are on the collector's delivery note.

Document B · Delivery note and weighbridge ticket

Page 1 of 1

DOCUMENT B: COLLECTION DELIVERY NOTE DN-2026-4412 (page 1) Collection date: 17 September 2026, 14:20 Supplier: Golden Wok Food Court Pte Ltd (UEN 201912345K) Pickup address: Blk 521 Bedok North Street 3, #01-88, Singapore 460512 Supplier representative: Mr Tan Boon Hock Vehicle: SGX 4721 K, tanker. Driver: Ahmad bin Ismail Containers emptied: 6 x 200-litre drums, 2 x 120-litre bins Quantity collected: 1,240 litres (per weighbridge ticket WB-17093-A) Condition: free-flowing, no visible water layer, no foreign objects WEIGHBRIDGE TICKET WB-17093-A, Tuas Depot Weighed in (gross): 17 Sep 2026 15:05, 1,864 kg Weighed out (tare): 17 Sep 2026 15:31, 720 kg Net weight: 1,144 kg. Density applied: 0.92 kg/L Net volume: 1,243 L, rounded to 1,240 L The net volume on this ticket is the quantity of record for this collection.

Document C · Supplier invoice

Page 1 of 1

DOCUMENT C: TAX INVOICE GW-2026-0917 (page 1) From: Golden Wok Food Court Pte Ltd, UEN 201912345K, Blk 512 Bedok North Street 3, #01-88, Singapore 460512 Invoice date: 18 September 2026 Bill to: Apeiron Bioenergy Pte Ltd Collection point: Blk 512 Bedok North Street 3, #01-88, Singapore 460512 Contact: Tan Boon Hock. Payment terms: 30 days from invoice date Line: Used cooking oil, collected 17 September 2026. Quantity 1,310 L at SGD 0.70 = SGD 917.00 Subtotal SGD 917.00. GST 9% SGD 82.53. Total due SGD 999.53 Note: quantity as measured at the supplier's premises prior to loading.

Round 2 prompt

Opens after round 1.

Exercise · on your own · 4 minutes

Spot the error

AI wrote this commentary from the P&L below, and nobody reviewed it. Three things in it are wrong. Tap each sentence you think is wrong.

SGD '000	Jul	Aug	Change	%
Revenue
UCO sales	2,140	2,265	125	+5.8%
Other income	18	12	(6)	−33.3%
Total revenue	2,158	2,277	119	+5.5%
Operating costs
Feedstock purchases	1,120	1,190	70	+6.2%
Collection fees to partners	96	98	2	+2.1%
Transport	148	190	42	+28.4%
Pre-treatment and storage	210	214	4	+1.9%
Staff costs	265	268	3	+1.1%
Other operating expenses	74	78	4	+5.4%
Total operating costs	1,913	2,038	125	+6.5%
FX gain / (loss)	(12)	9	21
Operating profit	233	248	15	+6.4%
Operating margin	10.9%	10.9%

0 sentences flagged Answers appear here when we reveal them.

Every time

The four checks

Foot the numbers.
Ask "where in the source does it say that?" and require a quote with page or cell.
Ask for the formula, not only the answer. For totals, have it use a formula.
For anything high-stakes, run it past a second tool and compare.

"Won't checking take longer than doing it myself?"

Sometimes, yes. So give it the jobs where checking is faster than doing:

It finds, you decide. It flags the 8 invoices out of 100 that don't match; you look at 8.
It points, you click. Every figure comes with its page or cell.
Formulas do the maths. Ask it to write the Excel formula. Check it once and it runs every month.

If checking means redoing the whole thing, that's the wrong job for AI. The number should come from the system, not the chat.

Keep these

Prompts to copy

They work in ChatGPT, Claude and Gemini. Paste your document or notes first, then the prompt underneath.

Meeting transcript → decisions and actions

From the transcript above, extract three things: (1) decisions made, (2) actions, each with an owner and a due date, (3) open questions nobody answered. For every item, quote the line from the transcript it comes from, with its timestamp. If an owner or a date is not stated in the transcript, write "not stated" rather than guessing. Give me three short tables.

No transcript? Your rough notes

From my rough meeting notes above, extract three things: (1) decisions made, (2) actions, each with an owner and a due date, (3) open questions nobody answered. For every item, quote the note it comes from. If an owner or a date is not in my notes, write "not stated" rather than guessing. Give me three short tables.

Then the follow-up email

Now write a follow-up email to the people in the meeting, under 200 words, plain professional tone. Confirm the decisions in one short paragraph, then list each person's actions with dates. Include the open questions at the end. Use only what is in the tables above; add nothing.

Key points across several documents

Read the attached documents and give me the key points as a table: Point, Document name, Page. Then a second table of anything that differs between the documents, with each value and where it comes from. If something is not stated in a document, write "not stated". Don't blend the documents into one summary.

Variance flags, not stories

From the P&L above, list every line that moved by more than SGD 5,000 or 20% between the two months: both figures, the change, and one sentence on the pattern in the data. Show the rows you used. Do not explain causes; I will add those. If a pattern is unclear, say so.

Make it stick: paste your best prompt into a Project (Claude or ChatGPT) or a Gem (Gemini) as its standing instructions, and attach the files you use every week. Next time, it's already set up. Start one now for your worst task.

Before you paste

Green, amber, red

Green · go ahead

Public information, dummy data, your own drafts and notes.

Amber · company account only

Internal documents with no personal data. Never a free personal account.

Red · ask first

Personal data of staff, suppliers or depositors. Bank details. Anything under NDA. Anything you wouldn't email outside.

The company's paid ChatGPT, Claude and Gemini plans don't train on our data by default; free personal accounts are different. Consumer apps that store data in China are red for anything company-related. Not sure? Ask Sebastien.

Your pocket guide

Your worst task this week

Five parts, one hour

Three things to know

It guesses the next word

It only sees the window

It forgets when the chat ends

One way to ask well

Three documents, two rounds

Round 2 prompt

Round 2 prompt

Spot the error

The four checks

"Won't checking take longer than doing it myself?"

Prompts to copy

Meeting transcript → decisions and actions

No transcript? Your rough notes

Then the follow-up email

Key points across several documents

Variance flags, not stories

Green, amber, red

Green · go ahead

Amber · company account only

Red · ask first

Before you go

What happens next