I say "create new page" to my AI agent Tim, and he goes off and creates a page in my automation platform, builds a data table, sets up three workflows, names everything according to convention, creates Facebook credentials, links them to the real page, and activates the schedule. Seven steps. One command. No follow-up questions.
It's not because AI is magically brilliant. It's because I wrote him an SOP — a step-by-step standard operating procedure — stored as a skill file that he loads and executes on demand. Same concept as an employee handbook, except the employee is an AI and the handbook is executable.
Why I Started Building Skills
When I started running 24 Facebook pages simultaneously, every new page meant I had to walk Tim through the same process. "Create the page in Loom. Name it exactly like the Facebook page. Create a data table with this naming convention. Set up three workflows — Video, Image, Reel..."
Every. Single. Time.
Tim has a memory system that remembers everything. But memory stores context — what happened, what I prefer, what went wrong. It's not designed to hold a 15-step sequential procedure with API calls, payload formats, and validation checks. Memory and skills serve different purposes.
So one day I told Tim: "Write the entire page creation process as an SOP. Save it as a markdown file. Every time I say 'create a page,' load that file and just do it."
That was skill number one — page-creation.
What a Skill Actually Looks Like
A skill is a single markdown file that tells the AI:
- What the skill is called and when to use it
- The exact steps to follow, in order
- Which APIs to call and what payloads to send
- What to watch out for — edge cases, constraints, forbidden actions
- How to verify the work before reporting "done"
The page-creation skill is 130 lines covering 7 steps:
- Create a Page in Loom (name must exactly match the Facebook page)
- Create a Data Table (name = page-id + "-content")
- Create a Facebook Credential (access token + page ID)
- Create 3 Workflows (Video, Image, Reel — following a naming convention)
- Link workflows to the data table and credential
- Test-run one workflow cycle
- Activate the schedule for automatic posting
Every step has the API endpoint, payload format, error handling, and validation criteria written out. Tim reads it and executes. No guesswork, no questions.
From 1 Skill to 34 — 7,058 Lines Total
Once page-creation worked, the pattern was obvious. Every repeatable process in my business could be a skill.
Running Facebook ads — A skill called run-ads tells Tim how to create campaigns, set targeting (age_min must be 25 because Facebook Advantage+ requires it), pick the right pixel, and manage budgets. Tim runs ads on his own.
Organizing receipts — A 56-line skill called receipt-rename specifies the file naming format ("Service - DDMMYY"), the correct Google Drive folder, and the sorting rules. Tim pulls receipts from Gmail and organizes them automatically.
Cutting live stream clips — At 292 lines, this is one of the longest skills. It covers transcription, viral moment detection, clip cutting, writing captions for 4 platforms, and scheduling 28 posts spread across a week.
Writing ebooks — A 400-line skill covering niche research, content writing, cover generation, KDP formatting, and getting everything ready to publish on Amazon.
Today Tim has 34 skills totaling 7,058 lines. They cover everything from page creation and ad management to clip cutting, ebook publishing, accounting, email marketing, competitor analysis, and server security audits.
The Skill That Creates Other Skills
This is the part that sounds like science fiction.
As the skill count grew, I noticed that writing new skills was itself a repeatable pattern — define the trigger, write the steps, test it, iterate, optimize the description so it fires at the right time.
So I told Tim: "Build a skill that creates new skills."
The result is skill-creator — 485 lines, the longest skill in the system. It handles:
- Interviewing me about what the new skill needs to do
- Writing the first draft
- Creating test cases and running them
- Having me review the output
- Iterating until the skill works correctly
- Optimizing the description so it triggers at the right moment
AI that builds more AI. It sounds like a movie plot, but it's just good process automation.
Deep Thinking on Demand — 7 Modes
One of my favorite skills is thinking. Normally, AI answers instantly — which is great for speed but terrible for decisions that need depth. This skill forces Tim to think deliberately in one of 7 modes:
- First Principles — Strip the problem down to fundamentals
- Brainstorm — Generate ideas without judgment
- Red Team — Find weaknesses in a plan
- Council Debate — Simulate 3-4 perspectives arguing
- Hypothesis — Form a hypothesis and find evidence for/against
- Pre-mortem — "If this project fails, why did it fail?"
- Opportunity Cost — What am I giving up by doing this?
I used this when deciding to turn my personal tools into a SaaS. I use it before launching features, before killing features, before making any decision where speed of answer matters less than quality of thinking.
Why Naming Conventions Matter More Than You Think
One skill that sounds boring but turned out to be critical: name-sop — the naming convention for everything in the system.
When you run 24 pages, each with 3 workflows + 1 data table + 1 credential, that's 120 entities. If naming is inconsistent, the AI can't find anything.
The naming rules are simple:
- Page ID = page name, lowercase, spaces become dashes
- Workflow = "Page Name - Output" (e.g. "Income in Click - Video")
- Data Table = page-id + "-content"
- Credential = "fb-" + page-id
With these rules, Tim can derive any entity's name from the page name alone. No guessing, no searching, no asking me.
This Is What ChatGPT Can't Do
The comparison is worth making explicit.
Standard ChatGPT / Claude: You type "create a new Facebook page in Loom." It asks "What's Loom? Where's the API? What settings?" You paste in your SOP. Every new session, you paste it again. Character limits mean you can't even fit a 400-line skill.
Tim with the Skill System: I type "create page Income in Click." Tim loads page-creation + name-sop, executes 7 steps, and reports back. No questions. Because every edge case, every naming rule, every API endpoint is already written down.
Three fundamental differences:
1. Scale. ChatGPT's Custom Instructions have a character limit. My skill library is 7,058 lines. And only the relevant skills load per task — no cramming everything in at once.
2. Execution. Custom Instructions tell AI how to talk. My skills tell Tim how to act. He calls APIs, creates campaigns, deploys code, posts content, sends emails. Real execution, not just text.
3. Ownership. My skills live on my own server. I write, edit, and delete them with full control. I don't wait for a platform to add features.
Skills Aren't Perfect on Day One
I don't want to make this sound flawless. Many skills started rough.
The run-ads skill originally didn't specify that age_min must be 25. Tim kept creating campaigns that Facebook rejected. I spent two hours debugging before discovering Advantage+ requires it. The fix went into the skill immediately.
That's how skills improve. Every time I hit a new edge case, it goes into the skill file. Like a good employee handbook — you don't write it once and call it done. It accumulates knowledge from real mistakes over time.
Same pattern as the memory system where Tim has accumulated 60+ entries from our daily work together. But they serve different roles: memory stores context ("who is the owner, what happened before"), skills store procedures ("how to do this specific job").
Memory + Skills = A Real AI Partner
If memory is what Tim knows about me and my business, skills are what Tim knows how to do.
Memory means Tim knows I prefer short responses, that I hate emoji, that my name is Pond not "Paul." Skills mean Tim can create pages, run ads, cut clips, write ebooks, and build entire tools from scratch.
Together, these two systems turn Tim from "a smart AI" into something much more useful — an AI that understands me and can do my work.
If you're a business owner who wants an AI agent like this — your own private server, persistent memory that learns from you, custom skills you build yourself, and real execution instead of just chat — take a look at Newton. It's the same system I use every day, packaged so you can set it up on your own server in minutes. See how it works →
— Pond