A skill is a set of instructions. In modern AI, skills have two jobs. The first is that you can save them and reuse them yourself — essentially a saved prompt. The second is that the AI can find and use them automatically when it's doing work on your behalf.
Back when I was a researcher, I did a lot of branding work and had a completely standard way of going about it. One of the things I needed to do was create market share data. I've built a skill that replicates the process I used as a market researcher — it has four basic steps. Let me show you how that works inside Displayr.
You'll see a new box called Skills. I'll create a new skill by clicking New, give it a name — note I'm using dashes between words, which is just the convention. You can add a description that both humans and the AI can understand, paste in instructions about how to calculate market share, and also specify when you want the skill to be used, so the AI can reason more precisely rather than triggering every time the word "market share" appears.
The last thing I'll do is paste in the actual skill. I want you to notice that what I've pasted in is just plain text — there is no computer code of any kind. It's a very detailed set of instructions, but designed so that anybody can implement them, whether that's a computer or a human.
You can save the same skill in multiple places. You might save one at your home folder level, which applies across your whole company, and then save a customised version in a specific client folder. The AI will use the client-specific version when working in that folder.
Why Skills Matter: Closing the Knowledge Gap and the Execution Gap
Skills are important because they close two significant gaps in research organisations.
The first is the knowledge gap. There are always people in a company who know a lot more about how to do things than others. When I was a researcher, I was part of what was called an advanced quant team — now more commonly called a data science team. We were brought in to click all the buttons that nobody else knew what order to click in. That was our special skill.
In the pre-AI world, that was a full-time job. People would come to the advanced quant team, we'd take over, and they'd wait a couple of weeks for results. In the AI world, the job of the advanced quant team is to create the skills and tools that other teams can use — yes, some troubleshooting here and there, but really it's about building systems for other people rather than doing the work yourself.
The second gap is the execution gap. This is when everyone knows what they're meant to do, but they don't actually do it — because they're busy, or it's tedious. The AI is simply more diligent than humans at this stage. If you've got tedious, well-defined work that you know exactly how you want done but the team tends to skip, AI is now closing that gap.
Six Factors That Make a Good AI Skill
I consulted with Gemini, Claude, and ChatGPT — the highest tiers of each — and together we identified six factors that determine whether you've created a good data skill.
1. Model Intelligence
The first factor is the intelligence of the models. Just like with humans, the more capable an agent is, the better the quality of its work. For skills, quality really comes down to the number of errors made — and the smarter the model, the lower the error rate.
But it's a curve with diminishing returns. Every new generation of Claude or ChatGPT gets better at following instructions, but errors can never be fully eliminated. Why? Because there's often ambiguous wording, missing context, or capabilities the agent simply doesn't have. The other five factors are how we address each of these.
2. Precision
We want to tell the AI how to do things in a way that leaves no room for ambiguity.
A colleague of mine gave the instruction: "Make the variable label for Q5 intuitive." The AI changed the label to the word intuitive. That makes sense — it's perfectly literal. If we want to be precise, we need to say something like: "Change the label field in the variable properties panel to a concise description of what Q5 measures."
How do you remove ambiguity? Find a pedant — someone in your team from an advanced quant or data processing background who is naturally precise. You can also test in chat, or paste your skill into another AI and ask it to critique it.
3. Correct Problem Structuring
There are two common mistakes here. The first is guessing the process when you don't actually know it. The second is asking the AI how to solve a problem and then instructing it to follow that process. This is circular: if the AI knew the correct process, you wouldn't need to give it the structure. And if it doesn't know, you can't ask it — because it doesn't know.
The right approach: only give process instructions when the AI is actually failing without them. Consult a genuine domain expert. Don't write skills for processes you don't understand yourself.
4. Embedded Context
Rules, cutoffs, and institutional knowledge belong inside the skill. If your organisation uses specific cut-offs for concept testing, or has a house view on how to handle certain data, put them in the skill. When the skill knows them, the AI applies them automatically — the user doesn't have to remember.
5. Explicit Tool Use
Modern AI platforms have many tools available, and often multiple ways to accomplish the same task. Being explicit about which tool to use — and how to invoke it — is one of the most powerful ways to improve skill quality. For example, Displayr has at least a dozen different ways to categorise text, but there is one best method. If you want the AI to use that specific approach, you tell it exactly which UI path to follow.
6. Reusable Data Assets, Not Just Outputs
The goal is not to get a number on your screen — it's to create data that can be cross-tabbed, filtered, visualised, and updated without running the whole skill again.
When an external coding tool calculates top-of-mind awareness, it produces a static output. There's no variable set, no way to cross-tab it, no way to extend the analysis. The same calculation done inside Displayr using a skill creates a variable set you can immediately work with — filter, cross-tab, visualise, add commentary. Good skills don't just produce answers. They produce data structures that stay useful.
Traceability: The New Standard for AI-Assisted Research
Skills improve over time — but improvement requires verification, and verification requires traceability.
When you delegate work to AI, it can look right but be wrong. For creative work, verification is easy — you look at the output and can tell immediately. But with something like spontaneous awareness numbers, how do you verify? If you have to recreate the entire analysis by hand to check it, you haven't saved any time.
You need to be able to trace every result back through the chain — from the output, to the calculation, to the input data, to the prompt that generated it — without rebuilding it yourself. In Displayr, you can click through from a chart to its input variable set, see exactly how that variable was created, read the text categorisation rules that were applied, and even see the AI-generated prompt that drove the analysis. That's full traceability.
There's a broader shift worth naming here. For most of our careers, software meant reliability — limited in what it could do, but consistent. Today, great software can be remarkable, but the trade-off is that it's no longer fully reliable. Our role has changed. We're less like operators of reliable machinery, and more like the manager of a capable but fallible colleague. We guide, we check, we course-correct. We accept that errors will happen and build systems to catch them. If we reject AI every time it makes a mistake, we'll never realise its benefits.
Advanced Automation: Building Blocks at Scale
The real power — especially for organisations with a technical team — comes from composing skills into automated workflows. The principle is: everything is Lego blocks.
A skill can call other skills. The brand awareness skill calls a separate "add page with purple bar chart" skill, which applies a branded template automatically. Each piece does one thing well and can be reused across any project.
Other building blocks worth using: templates (for charts, calculations, and entire multi-page document structures), external agents like Claude Code (excellent for integrating Displayr with non-analysis workflows, though a step up in complexity), and Q script (Displayr's scripting language for precise, deterministic automation).
The key principle is always the same: build reusable, auditable components. Don't solve the same problem twice — encode the solution once and let it scale.
