How to conquer everyone with AI (FULL Course)

Cover image

founders and employees using agents instead of chat are running 5 to 10 times more productive. take the middle number, 7x, and a single working day starts doing a full week's worth of output. this is everything I know about building with AI, condensed from a two hour masterclass into one place.

stage one was chat. stage two is agents.

chat is question to answer. you ask something, it replies, and you still have to go do the actual work yourself. an agent is goal to result. you hand it an outcome, it breaks that down into a plan, executes it, and hands you a finished product.

think of chat as a brilliant consultant, genuinely useful advice, but you're still the one doing the work after. an agent is that same brilliant person hired in house, actually doing the work end to end.

there's still a real place for chat. quick questions, small decisions, the kind of thing that would be overkill to hand an agent, still belong there. anything that counts as meaningful work, actual output that used to eat real hours, belongs with an agent instead.

what's actually happening inside that agent step

hand an agent a goal and it runs through three steps on a loop, over and over, until the task is done: observe, think, act. look at the situation, decide the next move, do it, then look again.

it keeps looping until it determines the task actually matches what you defined as done. vague about what "done" looks like, and the result comes back vague too. the app that runs this loop is called an agent harness, claude code, cowork, codex, whatever you're using. none of these tools built the loop themselves, that's the same underlying mechanism across all of them. what actually separates a good outcome from a bad one is how well you've tuned the agent running through it.

three levers control that tuning, one for each step of the loop:

→ the observe step gets tuned with context and skills→ the think step is just the model itself, the brain making the decision → the act step gets tuned with tools

think of the different harnesses like cars, codex is a ferrari, claude code is a range rover, another one's a g wagon. know how to actually drive, steering, brakes, accelerator, and you can hop into any of them and figure it out in minutes. a volvo with the context, skills, and tools properly dialed in beats a lamborghini running on defaults.

onboarding your agent like an actual hire

picture hiring a genuinely brilliant stranger for $100 a month, infinite bandwidth, works around the clock. no matter how smart he is, he's useless without three things: understanding your business, the tools to actually do the work, and training on your exact processes.

that's context, tools, and skills. get all three right and you've got the best employee you've ever had.

context: everything the agent should already know before you give it a task

this lives in markdown files, plain formatted text, one step above a raw txt file, headings and bullets and bold text but nothing fancier. markdown matters specifically because it's the easiest format for an agent to actually ingest. hand it a pdf and it has to unpack and extract the text first. hand it markdown and it just reads straight through, faster and cheaper every time.

the most important file is your claude.md (called agents.md in most other harnesses). this one auto loads into every single session by default. everything else, your about me doc, your ideal customer profile, your brand voice, your offer catalog, lives in a separate context folder that your claude.md points to.

proven live: open a completely blank folder with no context files and ask "who am I?" it has no idea, doesn't even try to guess. add a claude.md pointing at a real context folder, ask the exact same question, and it comes back knowing the founder, the business, the three products, the background, all of it.

worth being deliberate about what goes in these files. rushed, ai slop context documents create real problems downstream, a sloppy ideal customer profile with an inaccurate pain point baked in will quietly bend every piece of marketing copy toward the wrong angle. read every line before approving it.

why fresh sessions matter more than people think

every session has a fixed amount of usable space, think of it as brain power that fills up as a conversation runs longer. push a single chat for hours and something called context rot sets in, the same failure mode behind an endless chatgpt thread that starts hallucinating or forgetting things it said an hour earlier. information from early in the conversation quite literally falls out the top as new information gets crammed in.

the fix is simple and most people never do it: start fresh sessions deliberately rather than living in one endless thread. a well built claude.md and context folder means a brand new session still knows everything it needs to, without carrying the baggage of a conversation that's slowly degrading.

the memory layer, and why chatgpt "knowing you" isn't magic

past the static context files sits a memory.md, a running log of corrections and preferences the agent learns over time. tell it once to always sign off emails a certain way or never write formally, and it gets added as a permanent rule, checked before every future task.

this is genuinely the same mechanism behind chatgpt "knowing you" so well. it's not trained on you specifically, it's just quietly building its own memory markdown file in the background that you can't see or control. that's exactly why it sometimes bleeds unrelated context into the wrong place, a personal conversation influencing a landing page draft. owning your own memory file, in your own folder, fixes that entirely.

one small but genuinely useful trick, tagging a file directly with an @ symbol loads it straight in, no search loop required, functionally the same as the file already being in the session.

why lean context files actually matter

running the /context command inside claude code pulls up a real usage bar. a single small task can already burn 6% of total session capacity, and that's before accounting for everything auto loaded at the start. stack a hundred bloated markdown files and a fifth of the entire window is gone before a task even begins. keep files lean and information dense rather than sprawling.

tools: giving your new hire hands

context alone doesn't let an agent actually act. mcp, model context protocol, is the translator sitting between an agent and every external tool. claude speaks english, gmail speaks french, notion speaks spanish, slack speaks chinese. without a translator, teaching an agent every language directly takes real custom development. mcp removes that step entirely, one universal translator, connect anything through it.

a handful of tools are genuinely close to non negotiable:

→ apify, a marketplace of thousands of prebuilt scrapers, instagram, linkedin, google maps, reddit, virtually anything on the internet becomes reachable

→ firecrawl, the difference between an agent that can search the web and one that can actually read a specific page, extract branding, run a real cro audit

→ composio, a bridge that sits above every individual mcp connection, connect once and it travels across every harness instead of rebuilding the same connections in claude, then again in codex, then again anywhere else

→ chrome devtools and playwright, both give an agent an actual browser, spinning up a fresh session or controlling an already logged in one, useful for anything with no clean api, like sending a linkedin connection request

→ higgsfield, image and video generation routed through one connection instead of juggling a dozen separate model subscriptions

on security specifically, the honest approach is starting with read only access, letting an agent see data without acting on it, then escalating permissions as trust builds. most of the real risk sits with sensitive financial tools and company ip, not the everyday stuff. weigh the productivity gain against the actual risk, and decide deliberately rather than either avoiding tools entirely or granting full access blindly.

skills: teaching your hire exactly how you like things done

skills are just sops for ai, the same kind of step by step doc you'd hand a new virtual assistant. without one, every session starts from zero, correcting the same preferences over and over. save that correction process as a skill, and it never needs re-teaching.

every skill is a folder containing a skill.md file with three parts: a name, a description, and the full content. the name and description are the only pieces loaded by default, think of them as a spine on a bookshelf. the moment a task matches what a skill's description covers, the full content gets pulled and read, a mechanism called progressive disclosure, keeping hundreds of skills available without bloating every session's starting context.

proven live, building a real skill from scratch: connect an email platform's mcp, then in plain natural language, ask the agent to find the most recent unpublished newsletter, pull historical subject line performance data, and write ten new options based on what's actually worked before. it delivered ten ranked options with real patterns identified, bold claims outperforming generic ones, stacked tool names underperforming clean single concepts.

the specific patterns pulled out were genuinely useful on their own, independent of the skill itself. subject lines jamming two tool names together underperformed consistently. a single bold stance stated plainly beat a boring, descriptive task summary every time. one line that scored surprisingly well: a flat declaration that the business refused to be trapped inside one single ai tool, a stance, not a feature list. that's the kind of pattern only surfaces by actually looking at real historical performance data rather than guessing at what sounds clever.

then, a single instruction, build this exact process into a skill, and it's permanently repeatable from that point on.

skills rarely land perfectly the first time. run one, notice something off, correct it directly in the same chat, tell it to update the skill so it never happens again. most skills in daily use have gone through five to ten rounds like this before they stopped needing adjustment.

orchestrator skills, one skill that calls several others in sequence, are where this compounds hardest. a youtube publish workflow skill that calls a titles skill, then a thumbnail skill, then a description skill, one instruction instead of four separate ones.

proven live, an orchestrator skill actually running: a skill called ads analyst does exactly this for competitor research. give it nothing more than a competitor's ad library link, and it chains together several skills underneath it, scraping every active ad, downloading every video and image asset, finding and screenshotting every landing page those ads point to, then compiling all of it into a single readable report.

the finished report broke competitor activity down into real numbers, dozens of active ads split between video and image, the price point being tested, how long the longest running ad had been live, when the newest campaign launched. past the raw numbers, it pulled out genuine strategic insight too, flagging that a competitor had reframed their product's category specifically to justify a higher price, not by changing the product, just by changing what it was being compared against.

worth being honest about the first pass, the initial formatting wasn't right, some sections displayed too large and clunky to actually read comfortably. rather than accepting that, the fix was as simple as saying so directly in the same chat and asking for the layout to be corrected. that's the same iteration loop every skill goes through, rarely perfect on the first run, consistently better by the second or third.

one skill worth building specifically for decisions

a single skill built around better decision making has quietly become one of the most useful ones in daily rotation. ask claude directly whether something's a good idea and it'll build a perfectly convincing case for yes. ask whether it's a bad idea instead**, and it builds an equally convincing case for no**. neither answer can really be trusted on its own, since the model's just agreeing with whatever framing it was handed.

the fix spins up several separate sessions at once, each given a distinct persona, a contrarian actively hunting for what goes wrong, an optimist, a handful of others. their answers get shuffled anonymously and handed to a fresh round of sessions acting as peer reviewers. finally, one last session, acting as a kind of chairman, weighs every reviewed answer and delivers a single final verdict. genuinely useful for real calls, whether to ship a feature, what to prioritize next, anything where a single model's first instinct isn't enough to trust on its own.

two layers, not just one

everything covered so far lives at the personal level, a founder's own operating system, the extension of one person working faster. think of it like a robot arm bolted onto someone already capable, now able to reach further and do more.

a second layer sits above that: actual business infrastructure meant to run independently of any one person's laptop. customer support shouldn't live on a founder's personal machine, it needs its own dedicated environment, running continuously, accessible to a team rather than one individual. this is where more advanced frameworks come in, running on a dedicated machine rather than a personal laptop, built to operate autonomously around the clock.

the good news, nothing about the underlying principles changes moving into that second layer. same context, same tools, same skills, just running on infrastructure built to outlast any single person's session.

the folder structure underneath everything

everything runs out of one holding company style folder, called an os folder. inside it, subfolders for each major pillar, a business, a brand, a personal category. inside those, more specific folders, content, marketing, finance, newsletter.

global context and skills, the stuff true no matter what's being worked on, a founder's core identity, standing principles, sit at the top level claude.md. project specific context sits inside individual subfolders, layered on top of the global file rather than replacing it. tag a specific folder directly with an @ symbol to point an agent at exactly the right context without leaving the main working folder.

skills follow the same split. some live globally, usable anywhere. others live inside one specific project folder because they'll genuinely never be needed elsewhere.

the real cost, and why it barely matters

running this entire setup costs roughly $1,000 a month across every tool combined, not a small number, but functionally the cost of a fraction of one real employee for something closer to the output of a hundred.

worth knowing honestly, every major ai lab is currently pricing these plans well below the real compute cost, subsidizing usage hard to capture market share before eventually needing to raise prices as they push toward profitability. today's pricing isn't guaranteed to hold forever, which is exactly why the actual skill of building this system matters more than which specific tool happens to be cheapest this month.

getting an entire company there, not just one founder

top down rollouts, bringing in specialists to interview every department and build agents on employees' behalf, sound efficient and usually aren't. no outside specialist can understand every nuance of every role deeply enough to automate it properly at scale.

bottom up wins instead. train the people already doing the work to build their own systems. the best media buyer teaches the agent media buying. the best copywriter teaches the agent copywriting. that turns into a hiring freeze with a dramatically more leveraged team**, not layoffs,** and it spreads through actual company culture, build weeks, small competitions, weekly show and tell sessions, rather than a mandate handed down from the top.

the actual first step

create one folder. build a claude.md with a short overview of who you are and what you do. connect two or three tools. spend a day trying to run as much of a real workday as possible from inside that single agent session. the folder structure, the depth, the library of skills, all of that compounds over months. none of it needs to be perfect on day one, it just needs to start.

a few quick answers worth keeping close

→ first skill to build: an inbox triage skill, run three times a day on a schedule, cuts through the noise before it ever hits you

→ the biggest beginner mistake: being too perfectionist and too scared to type the first prompt, treating the whole thing like a black box that'll break if handled wrong

→ the most overrated tools right now: dedicated ai app builders. most of what they promise is just as achievable inside a properly set up project folder, without adding another subscription

→ the most underrated tool: a bridge layer like composio, the unglamorous infrastructure piece that quietly saves the most time once it's actually connected

→ the one thing worth never fully automating: real human judgment calls and anything involving direct phone sales, the trust and read of a real conversation isn't there yet

→ the honest timeline to feel a real productivity gain: a matter of days, not months, once context and a couple of tools are actually connected properly

p.s. if you want the exact plugin that builds this entire system for you, interviews you, connects your email and calendar and docs automatically, that's what the AI course hands over. reserve your seat here → awr-gamut-exploration.vercel.app/self-paced#reserve

running a team of 20 or more and need this rolled out company wide instead of just for yourself? that's a different scope entirely → awr-gamut-exploration.vercel.app/enterprise

remy