You are probably burning 7,000 tokens on every single message you send your AI agent.
Not on the task. On instructions the model never asked for and mostly does not need.
I run 11 AI agents across Profitable Founder, and I still got schooled on how Claude skills actually work by one 35-minute video: Ross Mike on Greg Isenberg's podcast, drawing diagrams until the whole thing finally clicked. It has 763K views for a reason.
Ross almost didn't record it. He texted Greg saying he had no "banger" tool to review that week. Greg's answer was to pull up a message from a viewer who watched one of their older episodes, learned to code from it, and now runs a cake business doing $150,000 a year.
So they recorded the basics. And the basics are what most founders get wrong.
Here's my breakdown of his system, plus where I'd push back.
The models are good. Your context is the problem
Ross opens with a claim that would have started a fight a year ago: the models are exceptionally good now. Opus is great. GPT is great. The "which model codes better" debate is mostly noise.
What separates quality output from slop is context, and you control the context.
Every time your agent acts, it assembles a context window out of the same ingredients:
- The system prompt from the model provider (Claude Code's leaked, and Ross read the whole thing)
- Your agent.md or claude.md file
- Your skills
- The tools the use exposes (read, write, and so on)
- Your codebase
- The conversation itself
That stack might start around 20,000 tokens and can grow toward a limit of roughly 250,000, which is when Claude Code and Codex compact the conversation.
The part nobody tells you: the model gets dumber as the window fills up. Ross tries to stay under about 70% of the window, because past 80 or 90% the quality visibly drops.
He compares it to cramming the night before an exam after skipping class all year. You can stuff information in. It won't go well.
So the whole game is: put less in the window, and make what's in there count.
The claude.md confession: 95% of people don't need one
This was the line that got me. Ross says 95% of people don't need an agent.md or claude.md file at all.
His analogy: imagine telling Greg "remember to bring a microphone" before every podcast recording. Greg has recorded hundreds of episodes. He knows.
Same with your stack. If your codebase uses React, the agent can see that. It has the code in context. Writing "this project uses React and Next.js and Supabase" in a config file adds nothing the model didn't already know.
The math makes it worse. A 1,000-line claude.md is around 7,000 tokens, and it gets injected into every single turn of every conversation. You pay that tax whether the instructions are relevant to the task or not.
The 5% exception: proprietary information the model genuinely needs on every turn. Your company's specific methodology. A currency convention. Things it could never infer from the code.
Everything else, Ross says, should be a skill.
How Claude skills actually work: progressive disclosure
A skill is a markdown file with a name, a description, and then the actual instructions underneath.
The magic is what loads into context: not the whole file, just the name and description.
Ross demos this with his own code structure skill, a 116-line file he uses to reorganize AI-generated code so it's easier for him to review. The full skill is 944 tokens. The name plus description? 53 tokens.
So the agent carries a 53-token index card that says "code structure: use when multiple workflows duplicate the same operational logic." When Ross asks it to clean up code, it scans its skills, matches the description, and only then pulls in the full file.
That's called progressive disclosure. Pay for the instructions only when the task needs them, instead of on every turn like a claude.md.
It matters because of what a language model actually is. Ross draws this out too: an LLM doesn't think or understand. It maps your words onto a giant vector space and finds what sits closest. Ask "what is the capital of France?" and Paris happens to sit nearby. It feels like understanding. It's math.
Which means the model will mimic you perfectly, but only if you've given it something to mimic. That's where his skill-building method comes in.
The recursive method: never write the skill first
Most people identify a workflow and immediately write a skill for it (or worse, have the AI write one cold). Ross calls this the worst thing you can do.
The skill has no idea what a successful run looks like, because there's never been one.
His counter-example is the sponsor pipeline for his YouTube channel. Sponsor emails come in, some legit, some scams, and vetting them eats time. So he set up an agent with its own email address (he's been hacked before, so it doesn't touch his inbox) and told it to check every 15 minutes for forwarded sponsor emails.
The first run was useless. It accepted everything. No research, no rejections, no "these guys are a scam."
Instead of rage-quitting, he walked with it. Do the research. Here's what a good company looks like. Now mark that one as bad in the Google Sheet. When it's a strong fit, email me.
Only after a successful run did he say: review what you just did, and create the skill.
Then the loop continues:
- The skill will still mess up somewhere (there are gaps)
- Ask the agent why it failed. It will tell you, specifically ("I got an error, insufficient credits")
- Feed the failure back, let it fix itself
- Tell it to update the skill file so it never happens again
Ross calls it recursively building skills, and claims a 100% hit rate on specific tasks now. His weekly report skill pulls from 8 data sources, YouTube Analytics and Twitter included, runs for 10 minutes, and executes flawlessly. It took 5 iterations to get there.
This maps to how I'd onboard a human, and it maps to how the best business automations get built in general: run the process manually with the system watching, then codify what worked. Skipping the manual run is why most people's automations produce garbage.
If you want to see the day-to-day version of this, Ross has a whole episode on running his life with skills. I covered it in my skillsmaxxing breakdown.
Don't download skills. And earn your sub agents
Two takes from the episode that will annoy people.
First: skill marketplaces are mostly a trap. Ross literally tells viewers not to download his own published skill ("I just put it up to get GitHub stars"). A random skill is a security attack vector, and more importantly, it carries someone else's context. Your agent needs the memory of YOUR successful run, not a stranger's. Read other people's skills for ideas. Build your own.
Second: you don't need a fleet of agents on day one. Ross started with one agent that did everything. Only after workflows were proven did he split off sub agents. Today he runs five: marketing, business, personal, and a couple more.
His line: "I didn't scale for what looks cool, I scaled for productivity."
I learned that one the expensive way. My own setup only got useful after I stopped adding agents and started teaching the ones I had. If you're at the multi-agent stage, my notes on running an AI agent workforce pair well with this episode.
And the payoff for putting in these unsexy reps? Ross's take: knowledge that took 20 people 20 years to acquire now costs 20 bucks a month. The cake business viewer turned one podcast episode into $150K a year. The gap between founders who know how these tools work and founders who complain that "AI doesn't work" is getting wider every month.
FAQ
What are Claude skills?
Claude skills are markdown files that teach an AI agent a specific workflow. Each has a name, a description, and step-by-step instructions. Only the name and description load into context (about 53 tokens in Ross Mike's example). The full instructions load when the agent decides the skill matches the task, which keeps your context window small.
Do I still need a claude.md or agents.md file?
Probably not. Ross Mike's rule: 95% of people don't. A claude.md gets injected into every turn (a 1,000-line file costs roughly 7,000 tokens each time), so reserve it for proprietary information the model needs constantly. Everything workflow-shaped belongs in a skill instead.
How do I create my first Claude skill?
Don't write it upfront. Do the workflow with your agent step by step, correcting it as you go. After the first fully successful run, tell the agent: "Review what you just did and create a skill from it." When it later fails, ask why, feed the error back, and have it update the skill file. Repeat until stable. Ross needed 5 iterations for his 8-source weekly report skill.
Should I download skills from a marketplace?
Ross advises against it, including his own. Downloaded skills are a potential attack vector, and they lack the context of your successful runs, which is what makes a skill reliable. Study them for structure, then build yours from a real workflow.
The bottom line
Strip your setup down. Kill the bloated config file, teach one agent one workflow by hand, and turn the win into a skill.
It's unsexy work (Ross apologizes for that mid-episode). It's also the version that holds up.
I interview bootstrapped founders every week about the systems behind their revenue, the AI-powered ones included. If this breakdown helped, the podcast goes deeper.
