TL;DR: AI makes it trivial to build agents that process data. The problem? Most of them are just noisy data dumps with no conclusion. The fix isn't more data — it's architecting judgment. Three checks: one-sentence directive, operational hierarchy, and the 50% deletion test.
I am James, CEO of Mercury Technology Solutions.
From my office in Wanchai— July 2026.
The Pretty Fluff Problem
I've been watching how enterprises build AI agents and define their "skills." And it reminds me exactly of the infographic craze from a decade ago.
Back then, content creators discovered they could compress a 2,000-word report into a visually stunning graphic. Pretty. Shareable. And mostly useless — because it summarized without concluding.
Today, developers are doing the same with AI. Throw massive data and API connections into a system prompt, expect a brilliant autonomous agent to emerge. The code is elegant. The output is garbage.
I see it daily. The agent ingests a spreadsheet of sales data, spits out three pie charts, and stops. Immense processing power. Zero judgment.
This happens because we misunderstand what an "Agent Skill" actually is. To fix it, we need to look back at how humans have structured information for 30,000 years.
The Evolution of Information: From Caves to Code
Structuring data to drive decisions isn't new. It evolved through four distinct phases:
1. Survival (30,000 BC) Cave paintings weren't art. They were prehistoric databases — recording volcanic eruptions, hunting grounds, migration patterns. The goal was simple: live or die.
2. Frameworks (1637) Descartes invented the coordinate system. X and Y axes. The foundational architecture for plotting abstract data. He gave us the structure to think spatially about numbers.
3. Clarity (1786) William Playfair invented the bar chart and pie chart. His goal wasn't aesthetics. It was to force politicians who couldn't read spreadsheets to understand the economy at a glance.
4. Action (1858) Florence Nightingale didn't just plot Crimean War casualties. She designed the "Rose Chart" specifically to prove that poor hygiene was killing more soldiers than enemy bullets.
Her chart wasn't a summary. It was a weaponized argument that changed military medical policy.
**The Standard:** When you build an AI agent, don't aim for a data summary. Aim for Nightingale's Rose Chart. Your agent's skill isn't to "process data" — it's to **make a definitive judgment.**
The 3 Crucial Checks for AI Agent Skills
At Mercury, we deploy Mercury Muses AI — agents that perform tasks across content, operations, and sales. Before any agent goes live, we run its skills through three checks.
If you're writing system prompts or building agents, ask yourself:
Check 1: What Is the One-Sentence Directive?
If you can't summarize your agent's ultimate goal in a single sentence, your agent will fail.
Bad skill definition:
"Analyze May and June sales data across eight product lines and provide a summary."
Good skill definition:
"Identify which product caused the June revenue dip and recommend whether to increase ad spend or cut the product."
AI models are eager to please. If you don't give them a definitive stance or conclusion to reach, they vomit data back at you. Architect the conclusion first.
Check 2: What Is the Operational Hierarchy?
When your agent looks at the prompt and data, what does it process first?
In our internal F.I.N.D.S. Framework, the 'I' stands for Information Structuring — architecting content into "citable chunks" that AI can extract and reason over.
If your agent is distracted by irrelevant background or bloated context windows, its cognitive hierarchy is broken. Design the priority of instructions so the AI knows exactly which variables matter most.
Check 3: The 50% Deletion Test
This is the most painful, most effective test.
Look at your agent's system prompt and data. Now delete half of it. Edge cases, redundant instructions, "nice-to-have" context — all of it.
If the agent's core directive survives, and its output becomes faster and clearer, that 50% was pure noise.
In the AI era, token efficiency is everything. The value of an agent's skill isn't determined by how much data it can swallow. It's determined by how much noise it can ignore.
AI Simplifies Execution, Not Judgment
Yes, AI makes building tools incredibly easy. You can generate a functioning agent in minutes. But because the barrier to execution has vanished, the barrier to *judgment* is now fully exposed.
AI will faithfully execute whatever you tell it to do. But if you don't know what you're actually trying to say, the AI cannot decide for you.
This is why, when we implement solutions for clients, we don't start with the data. We start with the business intent. We define the exact decision the system needs to make — then we structure the agent to deliver it.
Stop building agents that draw pretty charts.
Start building agents that make decisions.
Mercury Technology Solutions: Accelerate Digitality.
