Ben Liquornik · companion guide
9 minute video · written breakdown

AI agents 101

Most people build an agent once and never touch it again. Here's the part they skipped.

The reason is almost always the same: they jumped straight to asking the AI to do something. Everything below is what comes first, written down so you don't have to rewatch the video.

01 — the agent loop

Every agent runs the same three steps

Whatever the task is, it loops until it decides the goal is done. Then it hands you the result.

01
Evaluate
Collect what's available: your files, connected apps, tools, whatever else you fed it.
02
Think
Reason through a plan of action that gets to the goal you defined.
03
Do
Execute it. Research, pull files, write, take the action.
02 — two ways people use AI

Conversation vs. autonomous

About 99% of use is conversation. You ask, it answers, it asks a follow up, you answer, and you keep going until the output is close enough. That works, because you're steering every turn. It just gets your own work done slightly faster, and it needs you present for all of it.

Autonomous use is goal in, result out. That's what an agent is for, and it only works when you hand over a defined goal plus the inputs that get to the final result. Those inputs are three things.

Context+Skills+Tools
Like any employee, an agent is only as good as what you equip it with. A high IQ generalist with no training gives you generic work. Train it, and you get a specialist.
03 — context

Background knowledge on you and your business

It comes in two flavors.

Static context
The parts of your business that don't change.
  • What you do
  • Your ideal customer
  • Products and services
  • Partners and vendors
  • Past projects
  • Brand positioning and voice
  • Your working style
Hot context
What's live right now, updated as it moves.
  • Your current marketing campaign
  • Ongoing projects
  • Prospective clients
  • Email conversations
  • Meeting notes

Write all of it as markdown files, not PDFs or Word docs. Keep them in folders locally or in the cloud and give the agent access. Skipping this step is like asking a random smart person to build your marketing campaign without telling them what your company does.

04 — skills

How you do the work

Context tells the agent who you are. Skills tell it how you do the work. Think SOPs, the kind you'd hand a new salesperson or a VA so they can run a process without asking you twice.

Take writing a proposal. Your company has a specific way of building one, so spell it out like you're onboarding a real employee. Don't let it assume anything:

Document it once and it's saved. You stop retyping the same instructions every time a proposal comes up.

Option A · build the skill first
Start a new chat, tell Claude what skill you want, and let it walk you through defining it. This works, though it means planning everything upfront and it feels kinda unnatural.
Option B · do the work first (better)
Use Claude in conversation and play ping pong. Iterate, correct it, refine until the output is what you actually wanted. Then say:
build a skill for what we just did
Claude picks up the whole process, including what you rejected along the way, so the skill knows what good and bad look like.
05 — tools

The easy 10%

With context and skills in place you're about 90% there. Your agent knows the job, now it needs access. Connect Claude through MCP or API to the platforms you actually run on.

PowerPoint / Google Slides Gmail / Outlook Calendar Notion Slack / Teams Your CRM
Start here this week

Build it up brick by brick

  1. Write one static context file about your business. One page is fine.
  2. Pick the task you repeat most and do it with Claude in conversation.
  3. When the output is right, ask it to build a skill from what you just did.
  4. Connect the 2 or 3 platforms that task touches.
  5. Hand it the goal and let it run. Fix the inputs, not the output.
← benliquornik.com © 2026 Ben Liquornik