Big Idea

The argument in one line.

A personal AI advisor becomes genuinely useful not from model quality alone but from the structured context you feed it — goals, principles, energy signals, and a pre-response quality checklist that forces it to read your files before answering.

Who This Is For

Read if. Skip if.

READ IF YOU ARE…

You use Claude Code or Codex for work and want to extend it into personal decision-making and career coaching.
You are a solopreneur or creator who makes high-stakes decisions without a team, board, or mentor to consult.
You have tried using AI as a journaling or coaching tool but gotten generic advice because you gave it no personal context.
You want a lightweight memory system that improves over time without managing a vector database or complex setup.

SKIP IF…

You want a coaching tool accessible from a phone or web browser — this system runs entirely inside a CLI coding tool.
You are looking for a team or organization-facing tool; the entire design is single-person and personal.

TL;DR

The full version, fast.

The advisor is just a folder of four Markdown files. skill.md tells the AI what role to play and how to respond. plan.md holds everything personal: your annual goal, three guiding principles, what gives and drains your energy, your business model, and your financial situation. learnings.md is a changelog the advisor writes to after each meaningful conversation so it gets smarter over time. eval.md is a 20-item yes/no checklist the advisor runs before answering, forcing it to ground advice in your actual context and give concrete next steps. Used together, these files produce advice specific enough to help someone decide to leave a decade-long career.

Free for members

Chat with this breakdown — free.

Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.

Create a free account →

Chapters

Where the time goes.

00:00 – 01:29

01 · Why I built it

Personal story: three months of AI advisor use before leaving a job. Two real advisor quotes shown on screen. Stakes established before tutorial begins.

01:30 – 02:09

02 · Live demo

Codex session asking what to focus on to grow the business. Advisor returns three specific, contextualized recommendations.

02:09 – 03:31

03 · File 1 — skill.md

Behavior-only file: role definition, files to read, tone, response format. No personal information in this file.

03:31 – 06:53

04 · File 2 — plan.md

The core context file. Annual goal, three principles, energy map, business ICP, life and financial context, past goals history.

06:54 – 07:50

05 · Optional: bank MCP

MercuryMCP integration lets the advisor query live account data to check real-time progress toward a revenue goal.

07:51 – 09:00

06 · File 3 — learnings.md

Self-updating changelog. One skill.md instruction causes the advisor to draft a dated entry after every meaningful conversation.

09:01 – 10:14

07 · File 4 — eval.md

20-item yes/no quality gate that fires before every response. Top three: read latest context, cite actual goals, give concrete next steps.

10:15 – 11:22

08 · Fable model demo

Anthropic Fable model produced the deepest synthesis — reading plan and learnings while doing outside research on comparable creators.

11:23 – 12:31

09 · Recap

Five-card summary slide: SKILL.MD / PLAN.MD / LEARNINGS.MD / EVAL.MD / Model choice.

12:32 – 13:23

10 · CTA

Dual path: paid subscribers get the full skill at behindthecraft.com; free path is to paste the video transcript into Codex.

Atomic Insights

Lines worth screenshotting.

An AI advisor with no personal context gives generic advice; the context files are the product, not the model.
skill.md should define behavior and tone only — keep all personal information in plan.md, not mixed into the behavior file.
A one-line measurable annual goal is the single highest-leverage thing you can give your AI advisor.
Listing what gives and drains your energy lets the advisor flag when you are drifting from work that feels like play.
learnings.md turns a stateless AI into a system that knows you better each session — one line in skill.md makes it happen automatically.
A 20-check eval that runs before every response eliminates vague, ungrounded advice that makes AI coaching feel useless.
The three most important eval checks: did it read your files, did it cite your actual goals, did it give concrete steps for this week.
Connecting a bank MCP lets the advisor check real-time progress toward a revenue goal without manual updates.
The free path to get this skill is to paste the video transcript into Codex and ask it to build the files.
Model choice matters: deeper synthesis from outside research was meaningfully better with a stronger model on this task.

Takeaway

Four files that make AI advice worth following.

WHAT TO LEARN

The gap between generic AI responses and advice you would actually act on comes down to structured context — and it requires surprisingly little to set up.

Separating behavior from context is the first design decision: skill.md defines how the AI should advise you, plan.md holds everything personal — keeping them apart lets you share the skill file without exposing private context.
A one-line measurable annual goal is the highest-leverage thing in your plan document because it gives the advisor a target to aim all advice toward rather than optimizing for the wrong outcome.
An energy map — three things that give you energy and three that drain it — lets the advisor flag when you are drifting toward work that will burn you out, even when you cannot see it yourself.
learnings.md turns a stateless session into an improving relationship: one instruction in skill.md causes the advisor to append a dated insight after every meaningful conversation.
A quality-gate checklist run before every response prevents the most common failure mode of AI coaching: confident answers that are ungrounded in your actual situation and goals.
The three checks that matter most are whether the AI read your context files, whether it cited your specific goals, and whether it gave concrete steps for this week rather than strategy that could apply to anyone.

Glossary

Terms worth knowing.

skill.md: A Markdown file that defines an AI skill's behavior — what role to play, which files to read before responding, and how to format and tone advice. Contains no personal data.
plan.md: The personal context file at the center of the advisor system. Contains an annual goal, guiding principles, energy map, business model, life situation, and financial guardrails.
learnings.md: A running changelog of insights the advisor extracts from conversations — short dated entries added automatically when you make a decision, find a constraint, or shift direction.
eval.md: A yes/no checklist the advisor runs before giving advice. Each check must pass before a response is delivered, preventing vague or ungrounded recommendations.
MercuryMCP: A Model Context Protocol server that connects Mercury (a banking service) to Codex or Claude Code, letting the AI advisor query live account data to check financial progress toward a goal.

Resources

Things they pointed at.

13:00productbehindthecraft.com ↗

00:12linkcreatoreconomy.so ↗

06:53toolMercuryMCP

10:15toolAnthropic Fable model

Quotables

Lines you could clip.

00:30

“Your decision is psychological. It's about letting go of the person you spent a decade becoming.”

Complete thought, emotionally resonant, no setup needed — lands as standalone wisdom.→ IG reel cold open↗ Tweet quote

00:38

“Your kids are eight and four and will grow up fast. Your goal isn't to maximize income — it's to make enough so you can spend more time with them.”

Shows AI reframing the goal rather than answering the question. High shareability.→ TikTok hook↗ Tweet quote

08:04

“Plan.md is what I believe about my life and business right now. Learnings.md is what the adviser learns about me over time.”

Clean quotable distinction that makes the architecture instantly understandable.→ newsletter pull-quote↗ Tweet quote

10:04

“The eval forces AI to slow down and check whether it actually used my context, named its assumptions, and gave me advice I can act on.”

Names the failure mode it prevents — explains why this works.→ newsletter pull-quote↗ Tweet quote

The Script

Word for word.

Read-along

Don't just watch it. Burn it in.

See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.

00:00Hey, everyone. Today, I'm really excited to show you how to build an AI adviser that helps you think through important life and career decisions. My adviser is by far the most useful AI skill that I built so far.

00:13It knows my goals, gives me valuable feedback, and gets better the more I talk to it. So I recently made a video sharing why I decided to leave my job to bet on myself, and what I didn't share is that I talked to my AI advisor for over three months, leaving up to this choice.

00:31It gave me great advice like, your decision is psychological and about letting go of the person you spent a decade becoming, your kids are eight and four and will grow up fast, your goal isn't to maximize income, it's to make enough so you can spend more time with them, and other insights.

00:47So how does this advisor skill work? Well, it's basically a folder with a few text files. Here are the files in the advisor folder.

00:55There's a skill.md about how I want the AI to give me advice. There is a plan.md, which is a doc with my goals, principles, what gives me energy, and more.

01:07There is a learnings dot m d, which stores new insights from our conversations. And then there is a eval dot m d, which is a checklist for AI to make sure that's giving me useful advice before it actually gives me advice. Alright.

01:22So I'm gonna walk through each of these with some real examples, but first, let me kinda do a quick demo of how the adviser skill works. Okay.

01:30So here's an example of me talking to my adviser in Codex, and you can also use ClockCode. So I asked it without revealing any confidential information, what should I focus on to grow my business?

01:42And it came up with three things. Double down on AI builder content sharing my journey as an AI builder, make the pay sub stack product feel more different from the free YouTube, and then make sure you fulfill your existing sponsorships before taking on new work.

01:56And this stuff is very relevant to what I'm actually focusing on now, and the reason it's able to give me such good advice is because I've given you so much personal context about myself, and it also has learnings and evals to help it improve its advice over time. Now let's actually go through each of these documents to see what they cover.

02:14Your skill .md is basically how you want the AI to give you advice. It shouldn't actually contain any personal information about yourself.

02:22That's for the plan and other docs that we'll cover later in this tutorial. Instead, it should focus on telling AI what role to play, what context to read, how to advise you, and what to remember over time. Here's a snippet of myskill.md.

02:36So right off the bat, it's very important to actually have a very detailed description to let AI know when you should use a skill.

02:45So use whenever the user is stuck on decision, working through a hard problem, or asking for a gut check. And then it talks about how you are my trusted life and business adviser.

02:54Give me specific advice based on my plan, on my learnings, and on other things. And then when advising me, maybe open by reflecting what you see, separate your facts from assumptions, and give two or three concrete suggestions, right, and a bunch of other tips that I've included in the skill so that it kinda gives me advice as a trusted friend in a voice and tone that I like.

03:17And, again, I didn't write all this from scratch. I kinda work with AI and iterate with AI to come up with the skill that MD. But an important thing to remember is that this file defines the adviser's behavior, and the next file that we're gonna cover has all your personal contacts.

03:32And this next file is basically the plan dot m d, and your plan dot m d has the important context about your work and life, thereby making it the most important file to get right to set up this skill.

03:45So let's walk through each section that I should cover. First, it should have a goal, which is a one line statement for what you wanna achieve this year.

03:54So my goal, for example, is by the end of twenty twenty six, I wanna grow my business to a certain dollar amount while keeping control of my time and work. And having a clear measurable goal helps AI give advice that moves you closer to the life that you want. So definitely open by having a clear goal.

04:13And in this case, I also hooked up codex and claw code to MercuryMCP, is my bank, so that my adviser can actually pull information on how close I am to achieving this monetary goal, but this step is totally optional.

04:27Okay. So the next section that you should include is your principles. When I was a PM, I always found it useful to define principles so my team can make better product decisions, and the same applies to your career and life.

04:40And you don't be too crazy on the principles. You just can't just start with three. So here's three actual principles that I put in my plan doc.

04:46Go work work feels like play. I wanna document my journey as an AI builder and share insights along the way. Keep the main thing the main thing.

04:53I wanna focus on growing my YouTube and newsletter. I don't wanna get distracted by other opportunities and do the simple thing first, is to ship early and often and share what I learned.

05:04And the fact of the matter is as a content creator, I get new opportunities come to me every week. There's always some sort of AI conference or something new to attend, and these principles help my adviser remind me when I'm drifting away from my core work.

05:17And you should set up your own principles, but the important thing is think of three or four things that help you make decisions and write them down. Alright. Now this next session is also extremely important.

05:28You know, we do our best work when the work actually feels like play. So I think it's very important to tell your adviser what gives you energy and drains it, and here's my list. Right?

05:37So what gives me energy is building AI products that solve my own problems. It's teaching what I learned through my newsletter, YouTube, and behind the craft brand. And it's also learning from other AI builders through interviews, collaboration, and from the community.

05:52And what drains my energy is having too many live meetings, calls, and recurring commitments, doing a bunch of admin work that does not compound, and also selling or teaching stuff that I personally don't wanna build.

06:05Right? So, again, you don't have to go crazy here. Just list three things that give you energy and three things that drain your energy and put it in your plan dot m d.

06:13And the rest of my plan dot m d beyond these sections include sections on my business. So who is my ideal customer profile, what are their pain points, and what kind of promise am I making to them, includes another section about my life, which is about where I live in the expensive Bay Area, context about my family, and also what my financials look like so they can understand how much risk I'm willing to take.

06:36And also, I've kept a history of my past goals and how close I came to achieving them. So even if you actually don't build this AI advisor, just having all this stuff defined in a clear one page plan will force you to clarify your priorities.

06:51So it's definitely worth doing this exercise. Okay. So now we're back in Codex, and I wanna just quickly demo and ask my advisor what insights this can draw from my plan document.

07:02So here we go. So take a look at my plan doc and figure out if there's any insights or sections that you can draw, you can share with my audience, again, without revealing sensitive information. And let's see what it comes up with.

07:15Okay. So here's the output that it came up with. So the overall structure is there's updates, plan, and other things.

07:21There's core principles that, like I said before, go work feels like play. There is a business loop, build useful AI workflows for my own problems, share the journey publicly, and turn the best workflows into maybe paid assets.

07:35And there's other things like energy filter and the strategic insight. Right? So you can see here that it's pretty detailed document, but I try to keep it to one page because I actually wanna read it and update it myself.

07:46It has, again, the most important context about my business and my life. Okay. So now let's move on and talk about the third document as part of the adviser skill, which is the learnings dot m d.

08:00So plan dot m d is what I believe about my life and business right now, and learnings dot m d is what the adviser learns about me over time during our conversations. So to set this up, it's as simple as just adding a line to your skill file that AI should offer to store new learnings after every meaningful conversation. Right?

08:20So let's give this a test now in Codex. I'm gonna ask the adviser, based on my learnings.md, can you share three insights about me?

08:29And just look at the file and hopefully pull the three most relevant insights. Great. So here are the insights.

08:34So you get energy from building, not just advising. Uh, Yeah. I definitely wanna keep building instead of just, being a content creator.

08:40Your best positioning is AI builder. Maybe that's not true. Maybe I'm still learning how to build with AI, so maybe, like, curious beginner.

08:47And then I prefer grounded advice over generic strategy, which is definitely true. Like, I don't really like, like, very vague strategy. As you can probably tell from this channel, I already prefer to give all of you guys just, like, really practical things that you can copy right away.

09:01So, basically, the learnings.md, think of it as kind of like a change log of learnings that AI has learned from me over time through all the conversations that I have with it about my business and life.

09:12And the skill dot m d has a reference point to look at this file before it gives me any new advice. Now let's move on to the last file, which is the eval dot m d. And this is basically a yes, no checklist that my adviser runs before giving me advice.

09:27So mine has about 20 checks, but here are three of the most useful ones. Right? So first one is, did it read the latest context?

09:34Before giving me advice, did it actually read learnings.md and plan.md? Second one is, did it cite my actual goals, principles, and recent updates before giving me advice?

09:45Good advice should be personalized to me. Right? And another checklist is, did it give two or three concrete next steps?

09:51I don't I don't really just want generic advice. I want something that I can actually do this week. So if the advice is vague, then this eval fails.

09:59So, basically, the eval forces AI to slow down and check whether it actually use my context, name its assumptions, and give me advice that I can act on.

10:08It's gonna keep checking these 20 evals, and only when everything is marked as yes will it actually give me advice. Alright. So I posted this on x recently about how amazing Anthropic's new Fable model is at using this personal adviser scale.

10:24And unfortunately, the model got restricted by the US government right after it, so hopefully it'll come back at some point because using the best model really does make a difference in giving you detailed practical advice. I actually wanna show you guys the advice that I got from Fable. I was lucky enough to save it into a document before the model got banned.

10:43So here's the detailed advice that it gave me for things like how to grow my pay subscribers. It read my plan and learnings, and it did outside research on other newsletters, YouTube channels, and AI product creators, and it found a lot of really great insights that other models couldn't find, like the fact that renewal is actually the bottleneck for growing paid subscribers, improving retention, and also the fact that Substack notes was underused.

11:08Similar creators were posting to Substack notes multiple times a day to grow faster. So overall, I was just really impressed by how Fable was able to combine my personal context from this skill with outside research to help me see opportunities more clearly.

11:23Hopefully, Fable comes back by the time this video drops, but if not, you can always use the next best model like GPT and Opus. It's still very valuable if you set up the scale correctly. Alright.

11:33So let's recap the files that you should work with AI to create to make your personal adviser. You're gonna make a skill.md, tell AI what role to play, what files to read, and how it should give you advice, Then you should make a plan dot MD, which is just your goals, principles, energy, business life, and financial situation.

11:54Next, ask AI to make a learnings.md, which is just saving short notes to this file whenever it learns something meaningful from your conversations, and also make it eval dot m d, which is creating a short yes, no checklist to check itself before giving you advice.

12:10So examples of this checklist include grounding its advice on your plan, on your learnings, giving concrete next steps, and more. Whatever you wanted it to check before giving you advice.

12:20Right? And like we just covered, you should try to use the best model that you can when you're running this skill. So just to make this super practical, I've uploaded my full advisor skill to behindthecraft.com where you can find a dozen other skills built with the same quality bar.

12:36Now behindthecraft.com is available to pay subscribers of my newsletter.

12:41Otherwise, the next best thing to do for free is to just copy and paste the entire transcript of this video into Codex and ClockCode and ask it to walk you through setting up the skill. I also linked a newsletter post that you can copy as well. I don't think there's any exaggeration when I say that my AI adviser has fundamentally changed how I make important decisions.

13:03I feel like I never have to make an important decision alone now. I always have my adviser to consult along the way, and I hope that yours has the same transformation for you. So definitely don't give up until you set this up for yourself.

13:15Alright. So that's it. Please like and subscribe if you enjoy practical, no bullshit tutorials like this, and I'll see you next time.

The Hook

The bait, then the rug-pull.

Before walking through a single file or line of Markdown, the creator leads with stakes: he spent three months consulting an AI advisor before quitting his job, and the advice it gave was specific enough to quote verbatim. That credibility established, the tutorial becomes a reverse-engineering exercise — here is the four-file system that produced advice worth following.

Frameworks

Named ideas worth stealing.

00:49list

The 4-File Advisor Stack

skill.md — role and behavior only
plan.md — all personal context
learnings.md — auto-updated insight changelog
eval.md — pre-response quality gate

Four plain Markdown files that give an AI advisor enough structured context to produce personalized, grounded, actionable advice.

Steal forAny persistent AI agent that needs to improve over repeated sessions without a vector database

03:31list

Plan.md Sections

Goal — one-line, measurable, time-bound
Principles — 3-4 decision rules
Energy map — gives/drains
Business — ICP and promise
Life — family, location, finances
Past goals — history with outcomes

The canonical structure for a personal context document that makes AI advice relevant rather than generic.

Steal forOnboarding document for any AI system that needs to understand a person before advising them

09:01list

Eval Checklist Top 3

Did it read learnings.md and plan.md first?
Did it cite actual goals and energy preferences?
Did it give 2-3 concrete next steps (not vague strategy)?

A quality gate that runs before every response and blocks generic or ungrounded advice.

Steal forAny AI skill where output quality consistency matters — coaching, consulting, editorial review

CTA Breakdown

How they asked for the click.

VERBAL ASK

12:32product

“Copy the full skill at behindthecraft.com, or paste this transcript into Codex for free.”

Clean dual-path. Paid framed as fastest, not gated. Free path is specific and actionable — paste the transcript, ask Codex to build the files. No pressure, no scarcity language.

MENTIONED ON CAMERA

13:00productbehindthecraft.com ↗

00:12linkcreatoreconomy.so ↗

FROM THE DESCRIPTION

PRIMARY CTAWhere the creator wants you to go next.