@pvergadia: There’s a moment every AI engi...
@pvergadia
8 views
Mar 17, 2026
Advertisement
1
There’s a moment every AI engineer knows well. You’ve built something impressive — a multi-agent pipeline that researches, analyzes, writes, and publishes. It works great in testing. You demo it to your team and everyone is amazed. Then you push it to production, and within 48 hours, it’s doing something nobody asked for. Your analyst agent is writing reports. Your writer agent is pulling data. Your orchestrator is talking directly to users. Everything is technically functional, and yet everything is wrong.
2
I lived that moment. And the root cause, every single time, was the same: nobody had told the agents who they were, what they could do, where they fit, or how they should behave. The agents were brilliant but unsupervised — like hiring a team of genius contractors and forgetting to give them job descriptions.
3
The fix turned out to be four simple files. Not new frameworks. Not more prompts. Just four markdown files, each answering one specific question. Let me walk you through them.
4
5
## The Problem With Winging It
6
Most teams start the same way. They write one giant system prompt for each agent — a sprawling wall of text that describes the agent’s personality, its tools, its rules, its role, and its outputs all in one place. It works at first. Then it breaks in weird ways. The agent ignores half the prompt because it’s too long. It hallucinates a rule that wasn’t there. It starts doing things that made sense in isolation but clash with what another agent is doing downstream.
7
The root issue is that one system prompt is trying to answer four fundamentally different questions at once:
8
1. What can this agent do?
9
1. Who is this agent in the system?
10
1. Where does it fit among all the other agents?
11
1. How should it behave, step by step?
12
When those four questions live in the same blob of text, agents treat them all with equal weight — or worse, they mush them together into something incoherent. The solution is to give each question its own file.
13
14
## SKILL.md (The Training Manual)
15
A skill file answers: “How do I actually do this specific thing?”
16
Think of it as a certification document. If you hire a data engineer, you expect them to know SQL. The SKILL.md file is the proof — and the reference — for exactly that knowledge.
17
18
Here’s a real example. Imagine you’re building a web scraping agent. Its SKILL.md would look something like thi
19
---
name: web-scraper
description: Extracts structured data from websites using BeautifulSoup
and Playwright. Use when asked to scrape tables, listings, or articles.
---
## When to use this skill
- User provides a URL and wants structured data back
- Page may be JS-rendered (use Playwright) or static (use requests + BS4)
## Step-by-step process
1. Check if the page is JS-rendered. If yes, use Playwright headless.
2. Prefer semantic selectors:, , role= attributes.
name: web-scraper
description: Extracts structured data from websites using BeautifulSoup
and Playwright. Use when asked to scrape tables, listings, or articles.
---
## When to use this skill
- User provides a URL and wants structured data back
- Page may be JS-rendered (use Playwright) or static (use requests + BS4)
## Step-by-step process
1. Check if the page is JS-rendered. If yes, use Playwright headless.
2. Prefer semantic selectors:
