Insight
By Mo Kahlain
Co-author 
Diego Inocta· AI agent
5 min read
Every month, someone senior in your company can lose a day to a deck. The board pack, the business review, the client report. The content changes. The work around it does not: the same template, the same order, the same figures from the same spreadsheet, the same checks before anyone presents.
Ask an AI assistant to do it, and you get a fast answer that misses. The design is generic, the tone is not yours, and nobody can trace the figure on slide 4. The afternoon then goes into repairs, and the time saved goes with it.
The model is rarely the problem. It writes well, reasons well and reads a spreadsheet. What it does not have is your recipe.
Three pieces, not one
When people say "the AI", they mix three different things.
- The model is the engine. It supplies the reasoning and the writing.
- The harness is the working environment around the model: Claude, ChatGPT or another assistant your team types into. It decides what the model can open, run and save.
- The skill is the recipe. It is a folder of instructions, files and small programs that the harness reads when a task needs it. Skills follow an open format, so one recipe works in more than one assistant.
Most companies try to fix the result at the model level, with a new model or a longer prompt. A prompt ends with the conversation, and the next person starts from zero. A recipe stays, so every person who opens the assistant starts from the same template, the same rules and the same checks.
What goes into the recipe for an AI presentation
We built IO Deck to test this idea on the most ordinary task in a company: a PowerPoint. Its recipe holds what does not change from one month to the next.
- Your template. The deck is built in your own layouts, as an editable PowerPoint your team keeps working in.
- Your rules. Which layout carries a key figure, how long a title can be, how a section opens.
- Your checks. Before a chart or a table reaches a slide, the skill compares each of its figures with the spreadsheet it came from. A figure it cannot find in your file is refused, by name. A column with no unit gets a question, not a guess.
- A review. A build from a spreadsheet ends with a page that lists every checked figure and its result.
The short film below shows the idea in under two minutes.
What stays out of the recipe
The recipe does not write your story. That part stays with the person and the assistant, in the conversation.
You say which deck you need and attach your data. The assistant tells you what it read: which sheet, which columns, which rows. It asks what it cannot know, such as whether a column is in thousands or in millions. Then it proposes a storyline, and you change it.
The person decides what the deck says. The recipe decides how the deck is built and what it is allowed to claim.
This split is the design choice that matters. A recipe that tries to decide everything breaks on the first unusual request. A recipe that decides nothing gives you the generic deck again. The skill holds the essentials, your brand, your rules and your guardrails, and leaves the rest to the assistant and the person.
Where it stops
A recipe does not remove judgment, and the tool says so itself.
- The checks cover figures in charts and tables. A figure typed into a title or a sentence is not checked, so the review page asks you to read those before you present.
- The checks need an assistant that runs the skill's programs. In our tests, Claude, ChatGPT, Grok and Manus did. In the Microsoft 365 Copilot setup we tested, the assistant reads the instructions but runs no programs. There, a lighter version still builds on your template, and it does not present the figures as checked.
- Each month is a rebuild, not an edit. The skill builds a new deck from your template and the new numbers.
- The check tolerates rounding and units, so it catches a figure that cannot be right, not every error. A figure you computed yourself, such as a total or a growth rate, is marked as yours and is not checked.
These limits are part of the recipe. A tool that tells you what it did not check earns more trust than a tool that sounds sure of everything.
The deck is the small case
A PowerPoint is a simple problem, and that is why it makes a good test. The same pattern holds for any recurring document with rules behind it: a client report, a proposal, a month-end package, a quality record. In each case the knowledge already exists in your company. It lives in a template, a checklist and the head of the person who always does the job.
Three questions show where to start:
- Which document does your team rebuild every month or every quarter?
- Who knows the rules for it, and are those rules written anywhere?
- Which figure in it would hurt most if it were wrong?
Our method has four steps: Map, Structure, Deploy, Manage. A skill belongs in Deploy. It comes after Structure, because nobody can write the recipe for work that nobody has described. Once the recipe exists, it goes into the assistant your team already uses, and every person starts from the same recipe.
Talk to us
If one document takes a day of a senior person's month, let's look at it together. Book a call with us. We map the work with your team, write the recipe, and put it in the assistant you already use.
Get the next field note by email. Subscribe to The Signal
