Your instructions, with precedence over ours.
Ready-made AI bookkeeping tools ask you to trust a decision you cannot see. Numbers Game ships its procedures as readable text, and lets your firm write a section that outranks them.
What this means for your firm.
Numbers Game ships a library of roughly three dozen bookkeeping SOPs as readable prompt recipes, and a firm can write its own instruction section with its own trigger phrase and rules. The firm’s section takes precedence over the standard instructions, and every publish is versioned, acknowledged by a named person, recorded in a publish log, and reversible.
Whose name is on the return?
Every accountant evaluating AI is asking a question most vendors avoid. Not can it categorize this transaction, which is table stakes. The real question is: when it gets it wrong, whose name is on the return.
It is the accountant’s name. It is always the accountant’s name, and that single fact should shape every product built for this profession. Mostly it does not. You hand over the books, an opaque system makes decisions, you get a confidence score, and you cannot read the instructions it followed because those instructions are the vendor’s product.
Accountants notice this immediately. You have spent a career building firm-specific judgment: materiality floors, capitalization thresholds, which vendors are ever 1099-eligible, which client wants classes on every revenue line. That judgment is the firm. Handing it to a black box and hoping the box agrees is not automation, it is outsourcing without a contract.
An SOP library you can actually read.
Roughly three dozen SOPs across seven categories, covering the work a bookkeeping team repeats every month.
- Clearing the uncategorized queue, and vendor mapping
- Bank and credit card reconciliation
- Pre-close review, and locking the period
- Branded financials and board reporting
- Handling a new vendor, and class and department tagging
Each one has the same shape, and the shape is the point. What the work is, in plain accounting language, including the preconditions and the tie-outs you check before approving anything. The prompt, as actual text you can read, copy and change. Inputs and outputs: what the AI needs to see, and what artifact you should end up holding.
We call these prompt recipes rather than step paths, deliberately. A language model is probabilistic. Selling an accountant a “deterministic workflow” is selling something that does not exist, and the first time it deviates they will never trust it again. The reproducible artifact is the prompt, not the path. Give a firm the prompt and you have given them something they can audit. Give them a button and you have given them a leap of faith.
When a firm says our SOP is too thin, that is the product working.
The strongest signal we get is a firm telling us our procedure does not match how they work. One CPA firm on the platform read our month-end close SOP, decided it did not match how they close, and wrote their own: twenty pages, seven steps, their own materiality floors, their own rule for dormant accounts, their own capitalization threshold per client.
That is not a support ticket. So we shipped the layer underneath it.
Your own trigger phrase
Up to two words, 32 characters. The phrase your team says to invoke your firm’s way of doing the work rather than the standard one.
Your own instructions
Up to 8,000 characters, appended to the skill file your team downloads. Your thresholds, your exceptions, your client-specific rules, in your words.
Precedence, not suggestion
Where your section and the standard instructions disagree, yours wins. That is what makes it control rather than a preferences panel, and it is why the publish is governed the way it is.
Governed like a control change
Every publish is versioned, requires an explicit acknowledgement from a named person, is written to a publish log, and can be rolled back to any earlier version.
Publishing changes the next download. Nothing updates remotely.
This is the single most common misunderstanding, so it is worth being blunt about. When you publish a new version of your firm’s section, it changes what the next download contains. It does not reach into an installed skill and update it.
Every person on your team has to download the skill again and choose Replace. Until they do, they are running the version they installed. If you have just changed a materiality floor and it matters that everyone is on it, publishing is step one of two, and step two happens on each person’s machine.
Plan a re-download whenever a published change is material. A publish log that says the rule changed is not evidence that anyone is running it.
Split your operating knowledge into three homes.
This emerged from the customisation work and it generalizes well beyond accounting.
- The SOP is what a person does. It lives in a document, in human language, and it names no tool identifiers, so a vendor renaming something cannot silently break your close.
- The firm’s instruction section is what the AI does. It lives in the versioned skill, because that is the artifact with precedence, publish history and rollback.
- The client file holds facts only. This client’s accounting basis, thresholds and quirks. No procedure, no prose. Facts change on their own schedule and should not drag a procedure with them.
Most firms start with all three fused into one heroic master prompt. It works until it does not, and then nobody can tell whether the model misbehaved, the procedure was wrong, or a client-specific fact was stale. Splitting them makes each failure diagnosable, which is the property you need when the failure has your signature on it.
If you want to see the raw material, we publish the actual prompts our accountants send, organized by the work.
The model stops being the differentiator.
Frontier models are converging, and they are already good enough for the accounting work in question. What will separate firms is the quality of the instructions they feed one: their thresholds, their close sequence, their client-specific rules, written down properly for the first time in most firms’ history.
Firms are about to discover that their institutional knowledge was never documented, only remembered, and that the person who remembers it is retiring. The firms that write it down get to run it across a hundred clients at once. The firms that do not get to keep asking that person.
The details firms ask first.
Can our firm change how the AI does the work?
Yes. A firm writes its own instruction section with its own trigger phrase, up to two words, and up to 8,000 characters of its own rules. That section is appended to the skill your team downloads and takes precedence over the standard instructions.
What happens when our instructions disagree with yours?
Yours win. That precedence is the point, and it is why publishing is governed like a change to a control: versioned, acknowledged by a named person, written to a publish log, and reversible to any earlier version.
If we publish a change, does everyone get it automatically?
No, and this is the most common misunderstanding. Publishing changes what the next download contains. Nothing updates an installed skill remotely, so each person must download the skill again and choose Replace before they are running your new version.
What is in the SOP library?
Roughly three dozen SOPs across seven categories: clearing the uncategorized queue, vendor mapping, bank and credit card reconciliation, pre-close review, locking the period, branded financials, board reporting, handling a new vendor, and class and department tagging. Each gives the work in plain accounting language, the prompt as actual readable text, and the inputs and outputs.
Why call them prompt recipes instead of workflows?
Because a language model is probabilistic. A deterministic workflow is not a thing you can sell honestly here, and the first time one deviates the accountant stops trusting it. The reproducible artifact is the prompt, not the path, and a prompt is something a firm can audit.
Do the instructions remove the reviewer?
No. Nothing writes to a client’s books until a person approves it, whatever the instructions say. The instruction layer decides how work is prepared, not whether a human signs off on it.
Bring a real client workflow. See what changes.
Book a walkthrough to see Numbers Game on the work your team is doing now.