TLDR:
- A skill is a saved set of instructions that tells your AI tool how to run a specific fraud workflow using the same methodology, with different inputs every time
- The fastest way to build one is the interview method: describe the workflow, let AI ask detailed questions and your answers become the skill instructions
- Good skills ask before they assume, calculate baselines before they analyze and define exactly what the output should look like
Most skills take two or three iterations to get right — with this guide, you’ll have something useful by tomorrow morning.
You’ve been running the same type of query every morning in a row. Pull the last 24 hours of declines. Group by firing rule. Flag anything that looks like a false positive. Highlight patterns worth investigating.
The transactions change daily. Your analytical framework does not. And yet, every time you open your AI tool, you re-explain the whole process: what to pull, how to group the results, what counts as a false positive in your environment, what the output should look like. AI follows the instructions, produces a useful result and then forgets everything when you close the conversation.
A skill fixes that. You encode your methodology once and AI applies it consistently every time, without you walking through the steps again.
What a Skill Is
Think of it like onboarding an entry-level analyst. When you train someone to run decline spot-checks, you don’t hand them raw data and say, “figure it out.” You walk them through the process: pull the batch, identify the firing rules, flag anomalies, assess false positive risk and format the output. A skill encodes that exact walkthrough, so AI follows the same process each time you trigger it.
Each AI tool uses its own name for the concept, but the function is identical across all of them.
| AI Tool | What they call it | Where to build one |
| Claude | Skills | Customize → Skills → Create skill |
| ChatGPT | Custom GPTs | Explore GPTs → Create |
| Gemini | Gems | Gem manager → New Gem |
You write a set of instructions that define how AI approaches a specific type of task, save it and trigger it whenever you need it. The interfaces differ, but the concept doesn’t.
The Interview Method
Most people open a blank document and try to write skill instructions from scratch. They produce generic, surface-level instructions that miss the edge cases and operational nuances that make a skill useful in practice. The better approach is letting AI interview you.
Step 1: Tell AI what skill you want to build
Open a new conversation and describe the workflow you want to automate. Be specific about what the skill should do, but don’t worry about writing perfect instructions. Explain it the way you’d describe the process to a colleague sitting next to you.
“I want to build a skill for my morning decline spot-check. Every day I pull a batch of declines on key action types, look at which rules fired, flag false positives and note any patterns. Interview me about my workflow so you can buildthe skill instructions.”
Step 2: Let AI interview you
A well-prompted AI tool will start asking detailed questions about the workflow. Let it dig. The interview is where the real value gets captured, because AI asks about dimensions you might not think to include on your own.
| Category | What AI will ask |
| Scope | Which action types do you check? How many days back? Do you look at all declines or are they filtered by score range? |
| Methodology | How do you group the results? By firing rule? By score band? What do you check first? |
| False positive criteria | What signals tell you a declined transaction might be legitimate? Email age? Device uniqueness? Phone history? |
| Edge cases | What happens when there are zero declines in a period? What if a rule fired on hundreds of transactions? What if labeling data is missing? |
| Output | What does the final output look like? A summary table? A prioritized list? Grouped by rule? How detailed should each entry be? |
| Context | What scoring model do you use? What are your approve/review/decline thresholds? Any signal quirks the skill should know about? |
Answer honestly, including the messy parts. If your process has a step that’s inconsistent (“sometimes I check velocity first, sometimes I skip it if volumes are low”), say so. AI can build conditional logic into the skill to handle both scenarios.
Step 3: Review the draft and test it
After the interview, ask AI to compile everything into skill instructions. The first draft won’t be perfect. Test it against real data and note where it falls short. Common issues on the first run: the skill references signals you don’t have access to, uses thresholds that don’t match your model, produces output in a format that doesn’t match what you need or skips a step you described during the interview. All fixable with a follow-up: “The output needs to be grouped by firing rule, not by score band.”
Step 4: Iterate and save
Most skills take two or three iterations before they’re reliable. Each test against real data reveals a gap. Each gap leads to a refinement. After three rounds, you typically have a skill that runs the way you actually work. Save it in your AI tool and use it the next time the task comes up. After a week of daily use, you’ll likely make one more round of adjustments. Then the skill is locked in.
Key takeaway: The interview-based approach captures operational knowledge you didn’t realize you were carrying. Your thresholds. Your triage shortcuts. Your mental heuristics for separating real fraud from noise. Writing instructions from a blank page misses all of that. Letting AI ask the questions surfaces it.
Design Choices That Make Skills Work
The difference between a skill that’s mildly helpful and one that saves real time comes down to a handful of decisions. Build these in from the start.
Have the skill ask questions before it starts
The best skills don’t assume inputs; they ask. Before running any analysis, the skill should prompt you for the parameters that change each time: the date range, action types, score thresholds and whether you want a summary or detail view. A skill that assumes “last 7 days” when you actually need “since Friday” gives wrong results silently. A skill that asks “What date range should I pull?” gives correct results every time.
Example instruction to include in your skill:
Before running the analysis, ask me:
1. What date range should I pull?
2. Which action types should I focus on?
3. Should I use the summary view or full detail?
Then proceed with the analysis using my answers.
Have the skill calculate baselines first, then compare
Raw numbers without context are useless. A decline rate of 12% means nothing until you know whether your trailing average is 10% or 15%. Build baseline calculation into the skill’s first step so every finding comes with automatic context.
Example instruction:
Step 1: Before analyzing the current data, calculate the trailing 7-day average for each metric (volume, decline rate, average fraud score) per action type. Use these baselines as the comparison point for all subsequent analysis. When reporting a finding, always include the baseline next to the current value so I can see the delta immediately.
Define what good output looks like
If you don’t specify the output format, AI will guess. Sometimes a wall of text when you needed a table, sometimes a table when you needed a ranked list. Specify it explicitly.
Example instruction:
Format the output as:
1. Executive summary (3 sentences max: what’s normal, what’s not, where to look first)
2. Anomaly detail (one section per flagged metric, each with: current value, baseline, delta, and the 5 transactions most worth investigating)
3. Skip anything that’s within normal range. Don’t report good news. Only flag what needs attention.
Tell the skill what to do with edge cases
Skills break when they hit data they weren’t designed for. No declined transactions in the period, a rule that fired on 10,000 transactions, missing labeling data that makes precision calculations impossible.
Example instruction:
If there are no declined transactions in the requested period, say so and suggest expanding the date range.
If a single rule accounts for more than 50% of all declines, call that out as a separate finding before the main analysis.
If labeling data is missing, skip precision and recall calculations, note that they can’t be calculated, and recommend enabling labeling.
Version your skills
When you refine a skill, note what changed and why. After a month of daily use, you’ll have made a dozen small adjustments. Without versioning, you can’t tell which change fixed a false-positive issue or which accidentally broke the output formatting.
Example:
# Decline Spot-Check Skill v3
# v3: Added edge case handling for empty results
# v2: Changed output to group by firing rule instead of score band
# v1: Initial version from interview
Tell the skill what it should NOT do
Negative instructions are as important as positive ones. The skill should not fabricate data it doesn’t have, and you need to say so. If it should not recommend specific score values (because that requires backtesting you’ll do separately), say so.
Example instruction:
Never fabricate statistics. If you don’t have enough data to calculate a metric, say so.
Do not recommend specific score values for rules. Recommend direction only (raise, lower, investigate). Exact scores require backtesting.
Do not average metrics across different action types. Always report per action type.
Mixing registration and login metrics into one number hides the signal.
The best skills are opinionated. They don’t just process data; they encode your judgment about what matters, what doesn’t, what to flag and what to ignore. The more specific your instructions, the less time you spend correcting the output.
Start here
Open your AI tool. Copy this prompt:
“I want to build an AI skill for [name the task]. Interview me about my workflow. Ask me detailed questions about what data I pull, how I analyze it, what I’m looking for, which edge cases I encounter and what the output should look like. Then compile my answers into a set of skill instructions I can save and reuse.”
Spend 15 minutes in the interview. Test the draft against real data. Refine what doesn’t work. After three iterations, you’ll have a skill that saves you 30 minutes every time it runs. Then build the second one. Each skill you build compounds, and by the end of the month, AI runs your methodology. The analyst who builds five skills does in 30 minutes what used to take half a day.
Frequently Asked Questions
An AI skill is a saved set of instructions that tells your AI tool how to run a specific fraud workflow. The methodology stays consistent while the inputs change each time. In Claude, skills are saved in the Skills menu. In ChatGPT, they’re called Custom GPTs. In Gemini, they’re called Gems. You build a skill once and trigger it whenever you need that workflow.
Use the interview method: tell your AI tool which workflow you want to automate, then ask it to interview you about your process. AI asks detailed questions about your scope, methodology, false positive criteria, edge cases and output preferences. Your answers become the skill instructions. Most skills take three iterations to become reliable.
They’re the same concept with different names across AI tools. Each one saves persistent instructions that define how AI approaches a specific task. The implementation details differ by platform, but the function remains the same: a repeatable workflow you trigger on demand.
The initial interview takes about 15 minutes. Testing the first draft against real data takes another 10-15 minutes. Most skills need two or three iterations before they’re reliable, with each refinement round taking 5-10 minutes. You can have a working skill within an hour of starting, and a polished one within a few days of daily use.
