Keep more of it: 10 AI skills for the bills, not the menu

stock-loss

where the stock actually goes

How the two work together

Claude thinks it through. Paste the Claude prompt into Claude Code, or drop the folder into your skills folder. Claude does the judgement: what to look for, what is worth doing, what is right.

Codex gets it done. At the hand-off point Claude runs Codex on your machine with one command and passes it the Codex prompt. Codex does the mechanical part and hands the result back. Claude checks it before you see it.

No API key to set up: Claude calls the Codex you already have installed. If Codex is not installed, Claude does that half itself and tells you.

Prompt for Claude

---
name: stock-loss
description: Compares what your till says you sold against what actually left the store room, and produces a costed line-by-line list of the stock that cannot be accounted for. Use when the gross profit keeps missing, when a stocktake keeps coming up short, or before you accuse anybody of anything.
---

# Find out where the stock actually went, in pounds, before anyone gets blamed

You give this two stock counts with dates on them, the delivery invoices for the period between them, your till sales export, and your recipe or serve specifications. You get back one table: for every line, what the till says you should have used, what you actually used, and the difference in units and in pounds, sorted by money. Underneath it sits the list of losses you already know about and have recorded, subtracted line by line, so that the figure at the bottom is genuinely unexplained rather than merely uncounted. It never names a person. It produces the number that tells you whether there is something to investigate at all.

## What it does

1. **Close the period properly, because a sloppy count is the commonest cause of a fake variance.** You need an opening count and a closing count, each taken at a stated date and time, with the site closed and no deliveries in progress. Write both timestamps into the pack. Count every location: bar, cellar, store room, kitchen shelves, the chest freezer in the yard, the office cupboard with the spirits in it. A bottle in a location nobody counted is indistinguishable from a bottle that was stolen, and it will appear in your variance as a loss every period until someone counts it. ILTSA, the Institute of Licensed Trade Stock Auditors, which describes itself as "the UK's professional body for independent stock auditors", says that "Most hospitality businesses benefit from a monthly stocktake, but some choose bi-weekly or quarterly audits depending on their size, turnover, and control needs". Monthly is the right starting cadence, because a quarterly variance is too old to investigate and a weekly one is mostly counting error.

2. **Count in the unit you buy and convert to the unit you sell, and record both.** For every line record: the product, the pack as delivered, the unit you sell in, and the conversion between them. A 70cl bottle of gin sold in 25ml measures is 28 servings before any allowance for spillage; sold in 35ml measures it is 20. A 50 litre keg is 88 pints. A 5kg bag of chips at a 200g portion is 25 portions. Write the conversion factor on the line and keep it, because next period the same factor must be used or the variance is meaningless. Where the site pours 35ml as standard, record that explicitly: the permitted quantities for spirits such as gin, rum, vodka and whisky are "25 ml, 35 ml, multiples of 25 ml or 35 ml", and a bar that advertises 25ml and pours 35ml has a twenty-eight per cent loss built into its own price list rather than a theft problem.

3. **Build theoretical usage from the till, not from anybody's recollection.** Export every sale line for the period. Multiply each sold item by its recipe or serve specification to get the units of each stock line that should have left the shelf. A burger is a bun, a patty, a slice of cheese and a portion of chips, and each of those is a separate stock line. Where an item has no recipe attached, do not estimate one: list it under "sold with no specification" and total the retail value of those sales, because that figure tells the owner how much of their business is currently unmeasurable. This step is the one that makes the whole exercise work, and it is also the one people skip, which is why so many kitchens have a gross profit target and no idea which dish is missing it.

4. **Do the arithmetic in the open, one line at a time.** Actual usage equals opening count plus purchases minus closing count, with purchases taken from delivery invoices and matched to signed delivery notes. Variance equals actual usage minus theoretical usage. Cost the variance at the invoice price you actually paid in that period, excluding VAT, not at the retail price and not at last year's cost. Show all five numbers on every line: opening, purchases, closing, actual, theoretical, then the variance in units and in pounds. An owner who is shown only the final pound figure cannot tell a counting error from a loss, and will either dismiss the whole report or act on the wrong line.

5. **Subtract every loss you already know about and have written down, and name the ones you have not.** Recorded kitchen waste, staff meals, comped dishes and drinks, manager discounts, breakages, returned dishes, line cleaning losses, and samples poured for customers are all real stock that left with your knowledge. Each one must come off the variance with a document behind it: a waste sheet, a comp authorised in the till, a line cleaning log. What is left after those subtractions is the unexplained variance, and it is the only figure worth investigating. If any of those categories has no record at all, say so in one line each and put a value of zero against it, because a category recorded as zero because nobody writes it down is the single most common reason an honest site looks like it is being robbed.

6. **Check the measure before you look at any person, because the pour is usually the answer.** Draught beer, lager and cider must be sold in "1/3 pint, 1/2 pint, 2/3 pint, multiples of 1/2 pint", wine by the glass in "125 ml, 175 ml, multiples of 125 ml or 175 ml", and spirits in "25 ml, 35 ml, multiples of 25 ml or 35 ml". Measuring equipment must be government stamped, marked with a crown symbol, a CE mark or a UKCA mark. On brim-measure glasses, drinks must be "filled to the required level, having regard to the nature of the drink", and customers may ask for a top-up "to a reasonably acceptable level". So before anyone is spoken to: check the optics and the thimble measures are stamped and are the size the price list claims, check the glassware is the right line, weigh six standard pours and six standard kitchen portions against the specification, and check the beer line length and the cellar temperature. Record what you find as a measured figure. A free-poured spirit, an unlined wine glass and a 240g chip portion sold as 200g will each produce a large, consistent, entirely innocent variance.

7. **Check the till before you look at any person either.** Pull the exception report for the period: voids, no-sales, refunds, price overrides, discounts, training-mode transactions, and sales rung after close. Count them by operator and by hour, and compare against the variance by day. A stock loss with no matching cash loss points at the store room; a stock loss with a matching pattern of voids points at the till. HMRC treats the deliberate version of this as fraud: electronic sales suppression "is where a business manipulates their till systems to hide or reduce the true value or amount of sales", HMRC states plainly that "ESS is tax fraud", and the Finance Act 2022 powers carry penalties of up to £50,000 for making, supplying or promoting an ESS tool. That is why this step matters even for an owner who trusts everybody: if a till has functions that suppress sales, the owner is the one HMRC will be talking to.

8. **Rank by pounds, then prove it by repeating the count.** Sort the unexplained variance by cost, biggest first. Take the top five lines only. For each, write what the variance is worth per week, the three most likely causes in order, and the single cheapest test that would separate them. Then run the same count again over a shorter period, a fortnight or even a week, on those five lines only, with the measures corrected and the recording gaps from step 5 filled. A variance that survives a second clean count with accurate specifications and complete waste records is a real finding. A variance that disappears was a measurement problem, and finding that out costs a week and saves a relationship.

9. **If the evidence still points at a person, stop here and hand it to a fair process.** Do not confront anybody, do not search a locker, and do not put a name in this document. Acas is clear that "When there is a possible disciplinary or grievance issue at work, the employer should carry out an 'investigation'", and that "If the employer does not carry out a reasonable investigation, any decisions they make in the disciplinary or grievance case are likely to be unfair". It advises that "The employer should get somebody who's not involved in the case to carry out the investigation, for example another manager or someone from HR" so that "there is no conflict of interest", and that the investigation plan should set out what needs to be investigated, who is carrying it out, who needs to be spoken with, and "any sources of evidence, for example work records, emails or CCTV recordings". What this skill hands over is the evidence pack: dates, counts, costed variances, exception reports and the measurement tests, with no conclusion attached. The conclusion belongs to the investigation, and an owner who reaches it first has usually just made an unfair dismissal expensive.

## Then it checks

1. Both stock counts carry a date and a time, the list of locations counted is written down, and every location that holds stock appears on it.
2. Every line shows opening count, purchases, closing count, actual usage, theoretical usage, variance in units and variance in pounds, with the conversion factor between the buying unit and the selling unit printed on the line.
3. Purchases on every line trace to a delivery invoice number, and any invoice with no matching signed delivery note is listed separately rather than being included silently.
4. Every known-loss category (recorded waste, staff meals, comps, breakages, line cleaning) appears with either a document reference or an explicit "no record kept" and a zero value, and none is omitted from the table.
5. Sales lines with no recipe or serve specification are listed under their own heading with their total retail value, and they contribute nothing to the theoretical usage figure.
6. The measurement tests from step 6 appear as recorded figures, six weighed pours and six weighed portions with the specification beside each, and the pack states whether the measures are stamped.

Any check fails: name it, redo that step once. Failed twice: say what is wrong and stop.

## Rules
- Public information only.
- Never invent a fact, a number or a quote.
- Anything sent in someone's name says whose name it is.
- Never name a person in this document, and never attach a cause to a variance line. The output is counts and pounds. The moment a name appears, the document stops being evidence and starts being an allegation, and it will be read out in a tribunal.
- Never treat an unrecorded loss as a theft. A site with no waste sheet and no comp log will always show a large unexplained variance, and the fix is a sheet on the wall, not an accusation.
- Never cost a variance at retail price. The business lost the cost of the goods, not the menu price, and a report inflated to retail values will be dismissed by anyone who checks one line of it.
- Never search a person, their bag or their locker, and never put a hidden camera in a staff area. Both are decisions for the owner with their HR adviser, and getting either wrong converts a stock problem into a legal one.
- This output is a working document prepared for the owner's accountant, stock auditor and HR adviser to check before it is used. It counts stock and states arithmetic; it is not evidence of wrongdoing by anybody and it is not employment advice.

## Built from
- Institute of Licensed Trade Stock Auditors (ILTSA), "Hospitality Stocktaking Services", https://www.iltsa.co.uk/hospitality-stocktaking.html, no publication date shown on the page, read 14 September 2026: the definition of what a hospitality stocktake covers and the monthly cadence recommendation that set step 1.
- Business Companion, "Selling alcohol in licensed premises", https://www.businesscompanion.info/en/quick-guides/weights-and-measures/selling-alcohol-in-licensed-premises, no publication date shown on the page, read 14 September 2026: the prescribed quantities for beer, wine and spirits, the stamped-equipment requirement and the rule on filling brim-measure glasses, which are the whole of step 6 and the conversion discipline in step 2.
- HM Revenue & Customs, "Electronic sales suppression" collection, https://www.gov.uk/government/collections/electronic-sales-suppression, read 14 September 2026, and HM Revenue & Customs, "New powers introduced as HMRC targets till fraud", https://www.gov.uk/government/news/new-powers-introduced-as-hmrc-targets-till-fraud, read 14 September 2026: the definition of electronic sales suppression, HMRC's statement that it is tax fraud, and the £50,000 penalty, which is why step 7 checks the till exception report before anything else.
- Acas, "Investigations at work", https://www.acas.org.uk/investigations-for-discipline-and-grievance-step-by-step and https://www.acas.org.uk/investigations-for-discipline-and-grievance-step-by-step/step-2-preparing-for-an-investigation, no publication dates shown on the pages, read 14 September 2026: the requirement for a reasonable investigation, the independent investigator and the investigation plan, which is what step 9 hands the pack over to and why no name appears in the output.

Prompt for Codex

# stock-loss

## You are given
A folder for one UK hospitality site containing any of: two stocktake counts as CSV or spreadsheet (an opening and a closing count, each with a date and time), delivery invoices for the period between them, a till or EPOS sales export for the same period, a recipe or serve-specification file, and any waste, comp or line-cleaning records. Figures may be incomplete. Never estimate a missing one.

## Produce
Write these files into an `output/` folder beside the inputs.

1. `variance.csv` - one row per stock line. Columns: `line_id, product, supplier_code, buying_unit, selling_unit, conversion_factor, opening_units, purchases_units, closing_units, actual_usage_units, theoretical_usage_units, variance_units, unit_cost_gbp_ex_vat, variance_gbp, known_losses_units, unexplained_units, unexplained_gbp, invoice_refs`. Sort descending by `unexplained_gbp`.
2. `known-losses.csv` - columns: `line_id, product, category, units, gbp, document_ref`. `category` is one of `recorded_waste, staff_meals, comps, breakage, line_cleaning, samples, other`. A category with no records for the period still gets one row with `units=0` and `document_ref=NO RECORD KEPT`.
3. `unspecified-sales.csv` - every sold menu item with no recipe or serve specification. Columns: `item, quantity_sold, retail_value_gbp_ex_vat`. Include the total.
4. `till-exceptions.csv` - columns: `date, hour, operator_id, exception_type, count, value_gbp`. `exception_type` is one of `void, no_sale, refund, price_override, discount, training_mode, after_close`. Use an operator ID or terminal number, never a person's name.
5. `stock-loss.md` - one page. The total unexplained variance in pounds first, then the top five lines with the weekly cost of each, then the known-loss categories with no records kept, then the count of unspecified sales lines and their retail value.

## Rules
- British English. No em dash characters. Use " - " or a comma. No emoji.
- Never write a person's name, initials or nickname into any file. Operator IDs and terminal numbers only.
- Never write a cause, a theory or an accusation into any file. Counts, costs and dates only.
- Cost every variance at the invoice price paid in the period, excluding VAT. Never at retail price.
- Print the conversion factor on every row. If a line has no conversion factor, leave the theoretical columns empty and list the line in `stock-loss.md` as unmeasurable.
- If either stocktake has no date and time, write every file, leave the affected columns empty, and say so in the first line of `stock-loss.md`.

## Return
The absolute path of the `output/` folder, the total unexplained variance in pounds, the five largest lines by unexplained value, the number of known-loss categories marked NO RECORD KEPT, and a numbered list of every input you could not find.

Built from the best public work on this

Sources for stock-loss

Everything below was opened and read on 14 September 2026. Nothing is cited that could not be loaded.

1. Institute of Licensed Trade Stock Auditors (ILTSA), "Hospitality Stocktaking Services"

https://www.iltsa.co.uk/hospitality-stocktaking.html, no publication date shown on the page, read 14 September 2026.

ILTSA describes itself as "the UK's professional body for independent stock auditors", a body that "promote[s] best practice through training, industry standards and support for our members". It is the trade body whose members do this work for a living in British pubs, hotels and restaurants, which is why it is here rather than a software vendor's blog. Two things were taken from it. The definition of scope: a stocktake is "A detailed audit of all food, drink, and consumable stock held by a hospitality business", which is broader than most owners run, because consumables and kitchen dry goods are usually left out and they are where a surprising amount of money sits. And the cadence: "Most hospitality businesses benefit from a monthly stocktake, but some choose bi-weekly or quarterly audits depending on their size, turnover, and control needs." Step 1 takes monthly as the default and step 8 uses a short second count on the top five lines only, which is the practical compromise for an owner who cannot afford to shut the bar for a full count every fortnight.

Where the skill deliberately departs from the source: ILTSA's page exists to sell the services of its members, and its logic runs towards engaging a professional stocktaker. That is often the right answer, and the skill says so. But the page offers no method, no arithmetic and no numbers, and an owner who cannot yet justify a stocktaker's fee still needs to know whether they have a problem. The skill therefore provides the method ILTSA does not publish, and it is deliberately written so the output is the pack a stock auditor would want on day one rather than a substitute for one.

2. Business Companion, "Selling alcohol in licensed premises"

https://www.businesscompanion.info/en/quick-guides/weights-and-measures/selling-alcohol-in-licensed-premises, no publication date shown on the page, read 14 September 2026.

Business Companion is the official Trading Standards guidance site for UK businesses, which makes it the authority an owner would actually be measured against by the enforcement officer who walks in. It supplied the legal quantities that make a conversion factor real rather than assumed. Draught beer, lager and cider must be sold in "1/3 pint, 1/2 pint, 2/3 pint, multiples of 1/2 pint". Wine by the glass must be "125 ml, 175 ml, multiples of 125 ml or 175 ml". Spirits, specifically gin, rum, vodka and whisky, must be "25 ml, 35 ml, multiples of 25 ml or 35 ml". All measuring devices must be government stamped, marked with a crown symbol, a CE mark, or a UKCA mark from 1 January 2021, covering automatic meters, spirit measuring instruments, thimble measures and lined glasses. On froth, for brim-measure glasses drinks must be "filled to the required level, having regard to the nature of the drink", and customers may request topping up to a "reasonably acceptable level". The legislation named includes the Weights and Measures (Intoxicating Liquor) Order 1988, the Licensing Act 2003 and the Measuring Instruments Regulations 2016.

Step 6 is built entirely on this, and it sits before the steps that look at people on purpose. A 70cl bottle yields 28 measures at 25ml and 20 at 35ml. A site that prices at 25ml and pours 35ml generates a consistent, enormous, completely innocent variance, and every period it will look exactly like theft to anyone reading only the bottom line.

Where the skill departs: the source is a compliance guide, concerned with whether the customer got what they paid for. The skill uses it in the opposite direction, to test whether the business is giving away more than it is selling, and it makes no compliance finding. If a measure turns out to be unstamped or the wrong size, the skill records the measurement and tells the owner to raise it with their Trading Standards adviser, rather than telling them whether they are committing an offence.

3. HM Revenue & Customs, "Electronic sales suppression" collection, and "New powers introduced as HMRC targets till fraud"

https://www.gov.uk/government/collections/electronic-sales-suppression and https://www.gov.uk/government/news/new-powers-introduced-as-hmrc-targets-till-fraud, read 14 September 2026.

The collection page defines the problem: "Electronic sales suppression (ESS) is where a business manipulates their till systems to hide or reduce the true value or amount of sales", involving "use of ESS tools such as misusing built in till functions or installing software specifically designed to suppress sales". HMRC's position is not hedged: "ESS is tax fraud. You are involved with ESS if you have made, supplied, promoted, possess or have access to an ESS tool." The news release explains the mechanism in operator terms, that "ESS users will either have access to specialist software or will configure their Electronic Point of Sale (EPOS) device in a specific way that allows them to consciously hide true sales and the resulting tax that is due", and gives the penalty scale of up to £50,000 for making, supplying or promoting an ESS tool, with the powers introduced in the Finance Act 2022 in February 2022. Marc Gill of HMRC is quoted as saying "Electronic Sales Suppression gives the appearance a business is trading legitimately, when in fact they're really just stealing money from taxpayers."

This drove step 7, and it is the step most stock-variance methods leave out. Stock and cash are two ends of the same transaction. If the goods left the building and the sale was never recorded, the stock count shows a loss and the bank shows nothing missing, and an owner who only counts bottles will chase the wrong end of it for months. The step also protects the owner: an owner who has never audited their own till exception report does not know what functions it has, and under this regime the liability lands on the business.

Where the skill departs sharply: the HMRC material is about deliberate fraud by the business. The overwhelming majority of till exceptions in a small restaurant are training-mode transactions, mis-rings corrected by a void, and refunds for a genuinely wrong order. The skill therefore treats the exception report as a pattern to compare against the stock variance by day and by operator, not as evidence of anything on its own, and it never tells an owner that a void count means fraud.

4. Acas, "Investigations at work"

https://www.acas.org.uk/investigations-for-discipline-and-grievance-step-by-step and https://www.acas.org.uk/investigations-for-discipline-and-grievance-step-by-step/step-2-preparing-for-an-investigation, no publication dates shown on the pages, read 14 September 2026.

Acas is the statutory advisory body for workplace relations in Great Britain, and an employment tribunal will measure an employer against its Code of Practice. The guidance states that "When there is a possible disciplinary or grievance issue at work, the employer should carry out an 'investigation'", and gives the consequence that decides how this whole skill ends: "If the employer does not carry out a reasonable investigation, any decisions they make in the disciplinary or grievance case are likely to be unfair." On who does it: "The employer should get somebody who's not involved in the case to carry out the investigation, for example another manager or someone from HR", to make sure "there is no conflict of interest", and where that is impossible the employer should "try and make the procedure as impartial as they can". On the plan: it should cover "what needs to be investigated, who is carrying out the investigation, 'witnesses' - anyone who needs to be spoken with to find out about the issue, any sources of evidence - for example work records, emails or CCTV recordings, any time limits..., timeframes, policies or organisation guidelines to follow, setting out the importance of confidentiality".

Step 9 exists because of this, and so does the hardest rule in the skill, that no name appears in the output. In a small hospitality business the owner is usually the person who would both investigate and decide, and Acas's conflict-of-interest point is exactly the trap. An owner who has already written a name next to a variance figure has formed a view before the investigation started, and the document proving they did so is the one they created themselves.

Where the skill departs: Acas guidance is written for the whole disciplinary process and this skill stops well before any of it. It deliberately produces no conclusion, no suspect and no recommendation about a person. It stops at a costed, dated, unexplained variance and a list of the measurement and till tests already done, which is what a fair investigation would want as its starting evidence rather than its conclusion.

Best public prompt we found for this job

The closest public artefact is the `audit-support` skill in Anthropic's `knowledge-work-plugins` repository, at https://raw.githubusercontent.com/anthropics/knowledge-work-plugins/main/finance/skills/audit-support/SKILL.md. The repository has 24,016 stars, read from api.github.com. It is a SOX 404 control testing skill for large companies, nothing like a pub cellar, but it contains the one habit that makes a variance investigation honest:

Define the population (all transactions subject to the control during the period)

That is step 1's list of locations and step 3's requirement to export every sale line rather than a sample. A stock variance is arithmetic on a population, and if the population is incomplete, which is what happens when the freezer in the yard is not counted or a dish has no recipe attached, the arithmetic still produces a confident number and the number is a lie. Its distinction between design effectiveness and operating effectiveness also shaped step 6: checking that the site has a 25ml optic fitted is design, and weighing six actual pours is operation, and only the second one tells you what left the bottle.

What was deliberately not copied: the whole materiality and deficiency-classification apparatus, which assumes an audit function, a materiality benchmark and a reporting chain that a twelve-cover bistro does not have. Nor the sampling methodology, because a small site's stock population is small enough to count in full and a sample would only import uncertainty into a figure that is about to be used to decide whether to distrust someone.

Want this running in your business?

I optimise how businesses run — your sales, your visibility, your social media — and build bespoke software where nothing off the shelf fits. The first conversation is free. Work starts from £150 a day.