It's You — GIGO

01 / 27 · A working rant, not a Center of Excellence

It's you.

Garbage in, garbage out — how you can actually start with AI. Right now.

01Mr. Ben Honey
02For people the polite talk bounced off.

In a speech there are four components: the speaker, the audience, the speech, and the result.

Cold open. Do not warm them up. The title is the thesis and the insult. [on screen] It's you. Garbage in, garbage out — how you can actually start with AI. Right now. I am Mr. Ben Honey. This rant is for the people the polite AI rah-rah talk bounced off. In a speech there are four components: the speaker, the audience, the speech, and the result. I am the speaker. You are the audience. The speech is rude on purpose. The result I want is that you start today — not after a Center of Excellence study. TODAY. If you came for a feel-good hug, you are in the wrong place. If you came for a lever you can pull today to get value from an AI, welcome to my little corner of the universe. Related: the rest of this worldview, including the talks that already exist, lives at https://mlcu.com/articles.html

02 / 27 · Method of delivery

This is rude on purpose.

The polite talks already happened. “AI is helpful. You should try it. Here’s a chatbot writing a poem.” If that got you moving, this is not for you. If that bounced off you — stay.

01Those talks are too polite, too obscure, and aimed at a result months after the presenter started.
02This is supposed to empower the people who did not get that message. And challenge them.
03Leave if you wanted a hug. Stay if you wanted a lever.

Common sense is not common practice.

The polite talks already happened. “AI is helpful. You should try it. Here’s a chatbot writing a poem.” If that got you moving, that is great. Good for you. This rant is not for you. If that bounced off you — stay. Those talks are too polite, too obscure, and aimed at a result you can get months after you have already started. This rant is to empower YOU, the people who have yet to start, think it is too hard, or do not think AI applies to them. If you want a hug, that is at the very end. Right now is a lever you can pull today. Opportunity has knocked. Ignore it at your peril.

03 / 27 · The claim

If this is still hard, it's you.

Not the model. Not the vendor. Not IT. Something you know in your head that never made it into an MD file.

01Don't take that as an insult.
02Fine if you do. I don't care.
03The fix is the same either way.

Computers are a tool, not a replacement for thinking.

The claim. If AI is still hard, it's you. Entirely you. Pause. Then: Not the model. Not the vendor. Not IT. Something you know in your head that never made it into an MD file. That is the file — or files — the AI uses to record context and decisions. Don't take that as an insult. Fine if you do. I don't care. The fix is the same either way. Computers are a tool, not a replacement for thinking. You still have to think. The machine will not intuit what you have in your skull. Record the context and decisions — in the MD files.

04 / 27 · Computers

They do exactly what you tell them.

Best thing: Computers do exactly what you tell them to.
Worst thing: Computers do exactly what you tell them to. That is why you are frustrated with your AI. It is complying.

01If you are unsatisfied, treat that as true — and not as an attack. Then make it suck less.

When you want useful replies, ask smart questions.

Best thing about computers: they do exactly what you tell them to. Worst thing about computers: they do exactly what you tell them to. If you are unsatisfied with your AI, know that it is complying with exactly what you told it to do. Treat that as true — and not as an attack. Then make it suck less. Record the context and decisions — in the MD files. When you want useful replies, ask smart questions. The question is part of the input. AI is garbage in, garbage out, raised to an exponent.

05 / 27 · Cute. Also the actual claim.

GO = GI^MD

Classic GIGO is one-for-one. AI does not pass your mess through. It raises it. The boxes below are the terms.

GI: What you dump in. Can stay garbage. Stream of consciousness is allowed.
M: Context. Lives in the markdown.
D: Decisions — also direction, declarations, design, intent. Same files.
+MD: Sloppy use. Still negative. Amplified. Electricity into heat and noise.
−MD: Good use. Minus on the exponent. Inverts. Sucks less. Still garbage.

Successful systems work only because their degree of imprecision is acceptable. Strive for correct enough to be helpful.

Cute. Also the actual claim. Classic garbage in, garbage out is one-for-one. The machine passes your mess through. AI does not pass your garbage through. It raises it to an exponent. Garbage Out equals Garbage In to the power of MD — context and decisions. Garbage In is what you dump in. It can stay garbage. Stream of consciousness is allowed, complete with random thoughts, poor grammar, and ridiculous premise statements. All of it. M is context. D is decisions. They live in the markdown — the persistent written brief. In Claude those are files you can open: claude.md, project.md, and so on. In Grok they are custom instructions, embedded in a workspace. Same job. The whole list is context, direction, decisions, descriptions, intent, mission, documentation, design, declaration. Do not die on which word. Die on whether it is written down. In the pseudo-algebra we want a positive garbage out. Treat garbage in as a negative. Raise that to a positive MD — sloppy use, incomplete, misplaced, incoherent context and decisions. Negative raised to a positive. Still negative. Amplified. Electricity turned into heat and noise. Minus MD is good use of context and decisions. Minus on the exponent. Garbage in — a negative — raised to a negative MD. Delivers a largely positive garbage out. This sucks less. Still garbage. Never perfect. But exponentially better than the garbage in. Don't like the idea of “good garbage”? Too bad for you. Successful systems work only because their degree of imprecision is acceptable. Strive for correct enough to be helpful. Do not get bold and call it truth.

06 / 27 · Exhibit A — this conversation

Same garbage in. Better garbage out.

How the formula got here is the formula.

01GI, no MD: “Garbage in garbage out applied to AI. Help me flesh out a presentation.” The machine ran. Built a polite academic sermon about training data. None of it is what I wanted.
02Add some MD: don't build, record, I'm going to dump thoughts, and the differentiator is rude. Still rough. Less nonsense.
03Better M and D: here's the joke, here's Benisms, nested markdown, suck less. Input stayed put. Output sucked less.
04Again. And again. Mission accomplished. Not perfect. Less horrible.

Failure is good when it comes quickly and cheap, and leaves as its reward, experience and understanding. Conversely, failure is bad when it is slow to be recognized or at great expense, and leaves a legacy of blame and cowardice.

Exhibit A — this conversation. This presentation was created exactly by this process. The first pass was garbage and we did not polish the input. We improved the context and decisions files. I said: “garbage in garbage out applied to AI, help me flesh out a presentation.” No MD. The machine ran. Built a polite academic sermon about training data. Complete rubbish. None of it is what I wanted. Electricity converted into heat and noise. Then we added context and decisions to the MD files: don't build, record, I'm going to dump thoughts, and the differentiator for this presentation is rude, not the polite “AI is good.” Context and decisions documents are still rough, but they deliver less nonsense. Then better M and D: here's the joke — the algebra — here's Benisms, a thesis statement. If you fix the MD — context and decisions — then the result is progressively gooder. Sucks less. Let the AI adjust the presentation. It sucked less. Original garbage input. Output sucked less. Again. And again. Until mission accomplished. Not perfect. Less horrible. What I got out of it was an OK presentation, pretty good MD files — which are reusable in my next project. All by failing fast, refining context, and recording decisions. On my business card: Failure is good when it comes quickly and cheap, and leaves as its reward, experience and understanding. Conversely, failure is bad when it is slow to be recognized or at great expense, and leaves a legacy of blame and cowardice.

07 / 27 · The actual differentiator

Do not clean the input until it is gold.

Keep the dump. Improve intent and direction until the model can manipulate that garbage into something valuable.

01That is the opposite of every “write better prompts” talk.
02You are not here to become a prompt poet.
03You are here to put what you already know into a file the machine can use.

The truth is an expensive illusion. There is no truth. Be happy with a useful, consistent lie.

The actual differentiator. This is the fork from every polite “AI is good” presentation. Do not clean the garbage-in until it is gold. Keep the original input. Improve context, intent, and direction until the model can manipulate that garbage into something more valuable — better garbage out. “Write better prompts” is a useful technique on the tenth — or fiftieth — iteration of good AI use. It is not how you get started. Getting started with better prompts is rubbish. You start by doing it. AND notice the MD files — context and decisions — are where the value is. You are not here to become a prompt poet. Maybe later. You are here now to put what you already know into a file the machine can use. Truth is an expensive illusion. There is no truth. Be happy with a useful, consistent lie. If you want the long version of that mental model: Video: https://www.youtube.com/watch?v=orA_adBKUfo Slides: https://mlcu.com/CMGSW2019-AdventuresWithChargebackUsefulConsistentLie.pptx

08 / 27 · If you are frustrated

You caused it.

By one or both of these ways.

Never left your head: An assumption you treated as a fact. The model cannot intuit it. That is why the MD files exist.
In the file, now dead: You told it to look at two test boxes — web 1 and web 2 — to prove the process. You can have twenty web servers. You are mad you are not getting the farm. Ask why. Because that is exactly what you wrote down. Good then. Not good now. Change it.

01Say it blunt. Say it plain. It has no feelings to bruise. It will not guess that you meant all twenty.

I like to call that failed with good result: wrong now, happens to provide a usable result. When it changes, it looks like the change broke something that was working.

If you are frustrated, you caused it. By one or both of these ways. One: it never left your head. You treated an assumption in your head as a fact in evidence. The model cannot intuit it. Two: a “fact” was declared into the MD file, and it is no longer true. Spell this out loud. First pass, to test the process, you declared the scope as two boxes — web 1 and web 2. Now you are frustrated that you are not getting the farm of 20 web servers. Ask the AI why, and it will state what has been declared. Good then. Not good now. Change it. “AI, update the declared scope to all available web servers.” Say it blunt. Say it plain. It has no feelings to bruise. It will not guess what you meant. I like to call that failed with good result: completely correct then, wrong now, happens to provide a usable result, right until it didn't. When the situation changes, such that the result is no longer usable, it looks like the configuration broke something that was working. Really the definition of working changed.

09 / 27 · Embrace the rules

Push the rules as far out as practicable. Not too far.

Nested markdown is the machine. Presence of files is not the trick. Good use is.

Outer: How we talk. Who you are. Who the audience is. Tone, hedges, checks. Governs the interaction, not the warehouse.
Project: What the words mean here. What the work is. Declare failure so you can avoid it. Declare success so you can measure it.
Task: Thin. Only what is unique to this bucket. If it also applies to BI and API, it does not live here.

Every rule in a building code is a memory of some long-ago disaster. Every line in an MD file is the same.

Embrace the rules. Push the declarations as far back as practicable, and no farther. Practicable is the exact right word. Nested markdown IS the magic of the AI machine. Persistence, position, and content is the secret sauce to productive use of AI. AND IS THE POINT of this presentation. Three shells. Outer AI shell: declare normal interactions, how we talk — you and the AI — who you are, writing style, favorite phrases, quirks, tools and techniques, who a general audience is. Declare tone and tenor of a typical output. Project: declare context, direction, descriptions, intent, mission, design, and deviations from normal interactions. Declare failure so you can avoid it. Declare success so you can measure against it. Task: only what is unique to this bucket. If it also applies to other tasks, it does not live here. You may wish a sub-project level. Every rule in a building code is a memory of some long-ago disaster. Every line in an MD file is the same. Practicable means capable of being put into practice. Feasible with the means you have. Usable in this context. Not theoretical. Doable.

10 / 27 · Outer file

How the interaction is intended to flow.

01Who you are. Capacity planner, whatever. Talk to me as a professional. Talk to the audience as intelligent — suits, technical, or not.
02When you are not sure what I am asking, switch to Socratic. We will figure it out. Added later, because the conversation did not get on track.
03Then a correction to that correction: prefer open, free-flowing conversation over a strict question-by-question drill. Your quirks became its problem until you wrote them down.
04Two separate checks. Logical error: name it, explain the fix. Factual claim: search, do not assume. Guesses are fine if they say guess. You do not need exact-correct. You need not-made-up.

Documentation is only useful to the degree that it gets used, and is used only to the degree that it is useful.

Outer file. How the interaction is intended to flow. Not “everything.” The conversation. Who you are. Capacity planner, whatever. Talk to me as a professional. Talk to the audience as intelligent — suits, technical, or not. When you are not sure what I am asking, switch to Socratic. We will figure it out. That line got added later, because the conversation did not get on track. Then a correction to that correction: prefer open, free-flowing conversation over a strict question-by-question drill. Your quirks became its problem until you wrote them down. Two separate checks. Logical error: name it, explain the fix. Factual claim: search, do not assume. Guesses are fine if they say guess. You do not need exact-correct. You need not-made-up. Fail faster in miniature: conversation went sideways, you wrote the rule in the proper MD file level, next time the conversation sucks less. Documentation is only useful to the degree that it gets used, and is used only to the degree that it is useful. The AI is dutiful in using the documentation — the MD files — therefore it is critical to make them useful.

11 / 27 · Direction

The hedge is the point.

“Not working.” “Not not working.” “Avoid failure” instead of “success.” Do not promote that into a stronger, cleaner claim.

01The hedge is leeway — room to work.
02“Make this work” is not directionable.
03“Get us around the right side, then let’s go decide” is.
04Enough D to help. Not so much D it strangles.

Speak precisely, to the extent necessary, and no more.

Direction. The hedge is the point. Speak precisely, to the extent necessary, and no more. Saying “make 3 + 2 = 38” is very prescriptive and depends on facts not in evidence in the MD files. But it will give an answer. Saying “is there a way to make 3 + 2 = 38” is a useful statement. The AI will see the logical problem and discuss potential ways to get there — or tell you that you can't. Hedge the question. Suggestions allow the AI room to work. Soft language like: “this seems to be not working”; “that is not not working, but…” Do not promote that into a stronger, cleaner claim. The hedge is leeway — room to work. When a decision is made about definition, tools and techniques, acceptable and not acceptable — write that into the MD files. Tell the AI to do that. Blunt and plain, two slides back, was the same idea without the line. Blue elephant lives here too — what you think you think you think. Vague on purpose so it can give you what you want, not what you accidentally asked for.

12 / 27 · Do not start with the warehouse

A project. Not the project.

Next step is a small, well-defined task as your first project.

01Write an article, an email, a short story. Something small and easy to define — because the definition is what you are going to put in the MD files.
02That is where you duke it out: what you want, what you don't want. Each little confrontation trains the relationship.
03The AI learns how you work. You learn how to brief it. Skip this, jump straight to API / BI / DWA / core, and you are back to GI with no MD.
04The party before the meeting. Joking around so everybody knows everybody. Then the real work has a place to sit.

Develop solutions by getting a result, any result. Then a result you can control. Then a result you want. It is a mistake to attempt the mature final solution first.

Do not move in the furniture before the house is built. First step is a site-prep project — not the tenant-renewal project. Use a small, well-defined task as your first project. Write an article, an email, or a short story. Something small and easy to define — because the resulting context and declarations are what you are going to put in the MD files. That is where you duke it out: what you want, what you don't want. Each little confrontation trains the relationship. Skip this, jump straight to lease management, and you are at GO = GI with no MD. Think of this as the company picnic before the project meeting. Joking around so everybody knows everybody. Finding out quirks, preferences, oddities, and personality. Then the real work meeting has a place to sit. You can do this right here, right now. No need to sign up for an 18-month study course to take a certification program. Fail fast. Develop solutions by getting a result, any result. Then a result you can control. Then a result you want. It is a mistake to attempt the mature final solution first.

13 / 27 · Project file

You know what you think you think. It doesn't.

API is not a wrong term. API has many meanings. Kill the ones you do not want. Server? Service? Call? What is an API to you, and to this project?

01Write an articulable definition. Store it. Now “API” in this conversation is a fact and evidence of the project, not a vibe.
02Same for core server, prediction chart — predicting what — metrics you care about, metrics you don't.
03Nebulous plus multiple meanings: nip it. The rest show up as blue elephants. The model says it has no idea. That forces the definition. Definition goes in the file.
04This file gets fat. It is supposed to. It is the pile of declarations for this work product. Inside the interface. Not at the outer edge.

Based on facts in evidence at the time, the decision was appropriate.

Project file. You know what you think you think. It doesn't. API is not a wrong term. API has many meanings. Kill the ones you do not want. Server? Service? Call? What is an API to you, and to this project? Write an articulable definition. Store it. Now “API” in this conversation is a fact in evidence of the project, not a vibe. Same for core server, prediction chart — predicting what — metrics you care about, metrics you don't. Nebulous plus multiple meanings: nip it. When you don't know what the thing is, call it a blue elephant. Then define it as an undefined placeholder for what becomes the right word later. That forces the definition. Definition goes in the file. Now you have a defined, not-yet-definable term. The ambiguity remains helpful. These MD files get fat. They are supposed to. It is the pile of context and declarations for this work product. Socrates: the beginning of wisdom is the definition of terms. One of my favorites: based on facts in evidence at the time, the decision was appropriate.

14 / 27 · When you actually have to work

You are not the Python guy. The AI is.

What you bring: Define the data source: system name, credentials, which database, which tables. You also know the bottleneck in the shop. That is the context and direction only you have — M and D.
What the AI does: Puts on the Python hat, the administration hat, and the network-engineering hat. Coordinates connecting to the database, pulling the data, and giving it back to the process.

01SELECT * first. Then a subproject: of CPU, memory, swap, run queue — which are populated, which are trendable for this effort. You run the boxes hot, CPU is boring, wait is the tell. That discovery is project definition. It goes in the project MD.

Use brute force in the early stages to show a result, but stay mindful of being able to automate and integrate.

When you have the how-to-talk-to-AI MD foundation started — never finished — and actually have a project to work, you are not the Python guy. The AI is. What you bring: define the data source. System name, credentials, which database, which tables. You also know the bottleneck in the shop. That is the context and direction only you have — M and D. Imagine a utilization dashboard project. What the AI does: puts on the Python hat, the system administration hat, and the network-engineering hat. Coordinates connecting to the database, pulling the data, and giving it back to the process. It may SELECT * last two days, first. Then for each system: gather CPU, memory, swap, run queue, and everything else — which are populated? which are trendable for this effort? You run the boxes hot, CPU is an unreliable indicator, CPU run queue is the tell. Memory utilization is stable, memory swap rate is the tell. These observations and declarations ARE project definitions. They go in the project MD. Get a result, any result. Then a result you can control. Then a result you want. Then a result you need. Do not start at the mature dashboard. Use brute force in the early stages to show a result, but stay mindful of being able to automate and integrate.

15 / 27 · One of the actual magic tricks

Problem solving and troubleshooting are mutually exclusive.

Your job: Get a result for the project. That is problem solving. You stay here.
The work: Coding, tools, techniques, formatting, DBA-level stuff. That is troubleshooting. That is what you hand the AI.

01If you do both yourself — same person, same moment — you are doing it wrong.
02A senior programmer already knows the code path. Now they direct. That is why they feel suddenly more productive. They are.

Don't let your work get in the way of your job.

One of the actual magic tricks. Problem solving and troubleshooting are mutually exclusive. Your job: get a result for the project. That is problem solving. You stay here. The work: coding, tools, techniques, formatting, DBA-level stuff. That is troubleshooting. That is what you hand the AI. If you do both yourself — same person, same moment, like before AI — you are doing it wrong now that you are after AI. A senior programmer already knows the code path. Now they direct. That is why they feel suddenly more productive. They are. Don't let your work get in the way of your job. Work is the brick and mortar you used to do. Job is the result the project actually needs. The AI takes the work. You keep the job. THAT job is exhausting. Levels of authority is coming. Not this slide.

16 / 27 · Put the definition in the right MD file

Shared rules go outside the web bucket.

Name the project. Define what you are supposed to get. Define what failure looks like so you can avoid it. Then take the first step.

01Look at web servers only because everybody already thinks they know what a web server is.
02Most of what you declare for web also applies to BI and API. Park that outside, in the project MD. The web MD holds only what is unique to the web bucket.
03If you have to type it again in the API bucket, the two files might disagree. Pull the shared declaration back one level so the fight lives in one file.
04You do not need a novel in each MD file. You need the definition in the correct bucket.

The whole is other than the sum of the parts. Not greater. Other. This is not a principle of addition. — Koffka, if anyone asks

Put the definition in the right MD file. Shared rules go outside the task bucket. Name the project. Define what you are supposed to get. Define what failure looks like so you can avoid it. Then take the first step. Look at a web servers task. Most of what you declare for web also applies to database and API. Park that outside, in the project MD. The web MD holds only what is unique to the web bucket. Like an alert setpoint table, where it has API settings and database settings. Don't spread them around. Make one setpoint.md file at the project level. Proper use of the MD file concept IS the critical — and a skill you gain over time. You do not need a novel in each MD file. You need the definition in the correct bucket. Gestalt: the whole is other than the sum of the parts. Not greater. Other. This is not a principle of addition. Nested MD files are not a pile. They are a whole. That is the exponent.

17 / 27 · A new MD file

First analysis is terrible. That is not a surprise.

New file. Empty on purpose. Mission: analysis. Outer files still in force.

01Missed a metric. Metric not as interesting as you thought. Hole in coverage.
02Worked example: you overspend CPU on purpose because the cloud box with enough memory only comes with more CPU than you need.
03Naive analysis: over-provisioned. Scale it back. That is garbage with a suit on. Put why in the MD file. Next pass does not page you for a known anomaly.
04Start by recording how to avoid failure in the MD. Then record how to assist success. Those two paths produce very different output. Same GI. Better M and D. Sucks less. Rinse. Repeat.

Am I assisting in success or avoiding failure? Those two paths produce very different outcomes and behaviors.

New task: analysis. A new, empty MD file is implied. First analysis is terrible. That is not a surprise. Outer MD files still in force and always apply. Other tasks can be referenced. You will miss a metric. A metric will not be as interesting as you thought. There will be a hole in coverage. All of this gets declared in the proper MD files. A real example: we overspend CPU on purpose because the cloud box with enough memory only comes with more CPU than we need. Naive analysis: over-provisioned, scale it back. That is garbage with a suit on. Put why in the MD file. Next pass does not flag what would otherwise be “over-provisioned.” Summary so far: Rules of conversation: set up once, revised occasionally. Start project. Declare how to avoid failure in the MD. Declare what success looks like. Then record how to assist success. Those declarations influence in very different ways. Same GI. Better M and D — context and decisions. Sucks less. Rinse. Repeat. The long setup — definitions, MD files — feels slow and unproductive. Data and analysis feel fast and like progress, and they are a waste of time without the long setup. The M and D context and declarations are the exponent starting to pay.

18 / 27 · Caution, the whole way, not just at the end

MD is not a shrine.

The files fill with decisions that were right then.

01Tested on web 1 and web 2. You can have twenty. That line is now in the way.
02Stop. Design review. “Based on what we now know we want, look at the config and the MD, tell me what is no longer supporting the goal.”
03It will surface web 1 and 2: good enough to test the process, not good enough to analyze the farm.
04Update the files. Kill what aged out. Stale direction is sloppy MD. Sloppy MD amplifies.

Based on facts and evidence at the time the decision was made, the decision was appropriate. The DNS change revealed the failure. It did not cause it.

Caution the whole way, not just at the end. MD is not a shrine. The files fill with decisions that were right then. Tested on web 1 and web 2. You can have twenty. That line is now in the way. Stop. “Design review. Based on what we now know we want, look at the config and the MD, tell me what is no longer supporting the goal.” It will surface web 1 and 2: good enough to test the process, not good enough to analyze the farm. Update the files. Kill what aged out. Stale context and decisions is sloppy MD. Sloppy MD amplifies worse garbage out. Useful context: based on facts and evidence at the time the decision was made, the decision was appropriate. Failed with good result is the adult version of the web 1 and 2 versus all web servers. No one to blame. The mission changed. Still your job to find it, resolve — understand — the conflict, and fix it.

19 / 27 · Most important briefing tool in the conversation

Say the number.

If you do not say the level of authority, you and it are holding different ones. Outer file: 5 or 6 is implied. Project and task: raise or lower. Not enough, or too much, and you get into trouble.

0: Do exactly what is asked. Stop and get guidance for any variation.
1: Look into the situation. Get the facts. Report them.
2: Identify the problem. Alternatives, pluses and minuses. Recommend one.
3: Examine. Report what you intend to do. Do not act until you check in.
4: Solve it. Tell me what you intend, then do it unless overruled.
5: Take action. Report what you did.
6: Take action. No further contact necessary.
7: Delegate, with appropriate controls.

Good managers protect their group from external harm and minimize the harm the manager inflicts on their group.

Level of authority. This is the most important briefing tool in the conversation. If you do not declare the level of authority you give the AI, you and it are holding different ones. Not enough, or too much, and you guarantee frustration. Outer file: 5 or 6 is implied. Project file: raise or lower as desired. Task file: raise or lower as appropriate. Walk the numbers if they need it. Linger on 3 and 4. Three and four are a healthy relationship with the AI at the project and task levels. Zero is not enough. Six is too much especially while in “make it work” mode. You get into trouble at the extremes. Good managers protect their group from external harm and minimize the harm the manager inflicts on their group. The authority number method is a good mental model to do that.

20 / 27 · Where the exponent pays

You just promoted yourself.

If you already know what the code is doing — you spent a career thinking through it — you are now project manager of your slice.

01You feel more productive because you are.
02The mental effort belongs on the files, not on the brick and mortar.

A good plan violently executed now is better than a perfect plan executed next week. — General George S. Patton

Where the exponent pays. You just promoted yourself. You already know what the code is doing — you spent a career thinking through it — you are now project manager of your mission. You feel more productive because you are. The method effort now belongs to the AI, informed by the MD files. The mission completion, management, problem solving now belong to you as the project manager. Avoid the project-manager analysis-paralysis dilemma. A good plan violently executed now is better than a perfect plan executed next week. — General George S. Patton

21 / 27 · The cost

The junior will not learn the fundamentals this way.

They will not pick up the fundamental understandings by you sending the work to the AI.

01They will get assigned: code this little thing. They will put almost no mental effort in.
02They did not fail in a way the brain keeps. The AI did the brick and mortar. Do that long enough and you get atrophy.
03Pick your number. Twenty years. The seniors with the scars are gone. The juniors never got the scars. The company is worse for the wear. People are starting to notice.
04Other people's MD looks intimidating. All of it was the AI writing down what you decided. A junior should interrogate the files — why is this line here — because that is the impediment that already got paid for. Watch the loop. Date format was wrong. It fixed it. If you blink, you missed it.

Knowledge becomes power only when it is being successfully applied in the gaining of an objective. The cavalry isn't coming. You are the cavalry.

The cost. The junior will not learn the fundamentals this way. They will not pick up the fundamental understandings by you sending the work to the AI. They will get assigned: code this little thing. They will put almost no mental effort in. They did not fail in a way the brain keeps. The AI did the brick and mortar. Do that long enough and you get atrophy. Pick your number. Five years. Ten years. The seniors with the scars are gone. The juniors never got the scars. The company is worse for the wear. People are starting to notice. Other people's MD looks intimidating. All of it was the AI writing down what you decided. A junior should interrogate the files — why is this line here — because that is where the impediment and broken stuff got paid for. If you watch the AI logic loop on the screen, you may see where the date format was wrong. It fixed it. If you blink, you missed it. As a junior, look through the MD files. This is what there was to learn. GI can stay messy. Coding can be delegated. The learning cannot, unless you go get it out of the files. Knowledge becomes power only when it is being successfully applied in the gaining of an objective. The cavalry isn't coming. You are the cavalry.

22 / 27 · Early trick. Abandon it the minute it clicks.

You, Claude, and Claudette.

Factually false. Mentally useful. An AI-months-old person will think this is ridiculous. Use it anyway until you don't need it.

01Three people in the conversation: you, Claude (the window, the partner), Claudette (the thing behind the glass).
02You and Claude gang up on Claudette. Checkpoint: based on what we now know, how do we get her here sooner, better, faster, cheaper, cleaner?
03The detours that answer that go in the MD. They probably already did. That is still where they belong. That is self-correction.
04This is a useful, consistent lie. The truth is an expensive illusion. When it has done its job, throw it out.

Rename, relabel and describe opposites, to free yourself of current thinking. This annoys the crap out of people, as they cling to the familiar.

Early trick. Abandon it the minute it clicks. Three people in the conversation. Factually false. Mentally useful. An AI-months-old person will think this is ridiculous. Use it anyway until you don't need it. You. Claude — the window, the partner. Claudette — the thing behind the glass. You and Claude gang up on Claudette. Checkpoint: “based on what we now know, how could we have gotten here sooner, better, faster, cheaper, cleaner?” The suggestions and observations that answer that go in the MD files. They probably already did. That is still where they belong. That is self-correction. Imagine the reply was: “if we told Claudette that the speaker was the general manager, and the audience was the city council, and the mission was to introduce the adopt-a-street project, then it would have been more obvious to use the blue elephant spreadsheet instead of the grape-ape database.” Add this context and decisions to the proper MD files. The next task may suck less. This is a useful, consistent lie. Useful mental model. When it has done its job, throw it out. Rename, relabel and describe opposites, to free yourself of current thinking. This annoys the crap out of people, as they cling to the familiar.

23 / 27 · Second mental model, the rude one

The AI is a very smart idiot.

The AI is an expert in many fields. As a person: completely disappointing, because it isn't one. But it is brilliant.

01No feelings. No malice. Weak intuition. Literal. Does exactly what it was told.
02You can say “that's wrong, do it this way.” It will not sulk. It will ask what “that” means. That answer goes in a file. You still have to have the conversation.
03Don't get mad at a compiler. You are an effective manager of something that is not a coworker.
04Treat it inside those limits and you get further, and you enjoy it more.

I have no idea what you are talking about, do you?

Second mental model, the rude one. The AI is a very smart idiot. The entity of an AI is an expert in many fields. As an entity of a person: cold, emotionless, not intuitive, and basically a completely disappointing person, because it isn't one. But it is brilliant, has a great memory, is super dedicated to trying to help, and is great at research. No feelings. No malice. Weak intuition. Literal. Does exactly what it was told. You can say “that's wrong, do it this way.” It will not sulk. It will ask what “that” means. The answer goes in a file. It will not forget. You still have to have the conversation with the AI though. Don't get mad at a compiler. You are an effective manager of something that is not a coworker. Treat it inside those limits and you get further, and you enjoy it more. When you are frustrated with the conversation with the AI, imagine it saying: I have no idea what you are talking about, do you?

24 / 27 · Tangent, on-thesis

The risk is not the AI.

Doomsayers: it will take over. The true part: you can put malicious intent in the MD files, and the machine will do what it is told.

01Same exponent. Sloppy MD amplifies. Malicious MD amplifies on purpose.
02MD is markdown. In some tools they are files you can open. In others they are custom instructions, a workspace, a project brief. Same job: written context and direction that persist.
03Push permissions as far out as practicable. Not further. The risk is what you embed.

The 10th Man is the loyal dissenter. Duty to poke holes. Motive is still the best decision for the organization.

Short tangent. The risk is not the AI. Doomsayers: AI will take over. The true part: you can put malicious intent in the MD files, and the machine will do what it is told. Same exponent. Useful MD amplifies better, suck less. Malicious MD amplifies maliciousness on purpose. MD is markdown. In some tools they are files you can open. In others they are custom instructions, a workspace, a project brief. Same job: written context and direction that persist. Claude shows files. Grok: custom instructions plus a workspace. Cursor, Codex: AGENTS.md. Do not die on the filename. Die on whether M and D persist in writing. Push permissions as far out as practicable. Not further. The risk is what you embed. The 10th Man is the loyal dissenter. Duty to poke holes. Motive is still the best decision for the organization.

25 / 27 · Restate for the cheap seats

Still not happy. Still your fault. Fix the MD files.

If you are still not happy with what your AI is giving you, it is still your fault, and it is up to you to go fix it. For the record: you fix it in the MD files.

01Assumptions in your head. Or assumptions you no longer wish to be true.
02Have the conversation. Write the decision. Kill what aged out.
03Still garbage out. Better garbage. Suck less. Mission accomplished.

Correct is a matter of interpretation and perspective. Strive for correct enough to be helpful.

Restate for the cheap seats. Still not happy. Still your fault. Fix the MD files. If you are still not happy with what your AI is giving you, it is still your fault, and it is up to you to go fix it. For the record: you fix it in the MD files. Assumptions in your head. Or assumptions you no longer wish to be true. Have the conversation with the AI. Kill what aged out. Update the old. Write the decision. Still garbage out. But better garbage. Make it sufficiently suck less. Mission accomplished. Correct is a matter of interpretation and perspective. Strive for correct enough to be helpful. Being disappointed in not having a “perfect output” — there is no such thing. Best you can do is better context and declarations.

26 / 27 · Next task. Right now.

Do not start an 18-month certification expedition.

You already have enough. Start right now.

01Right now: one small, well-defined project. Outer MD file. Duke it out — want, expect, don't want, gray. Have the AI record that in the file.
02Then another small, well-defined project. It should go much better, because the file already knows how to have the conversation. That is the point. Rinse. Repeat. Slightly larger each time.
03Then a real project. The conversation is already defined. Now you only define the terms of the project, what success is, and what failure looks like so you can avoid it. Keep updating the MD files. Pretty quickly you have a successful project.
04Do not wait for the cavalry. Do not wait for a perfect plan. A good plan now is better than a perfect plan later, which is never. If it is still difficult, it is still you. The lever did not move.

Attack in order to clear up the situation, and gain a basis for further action. The battle itself proves the best method for estimating the enemy. — American intelligence assessment of Nazi military, 1942

Next task. Right now. Not homework. A professional person, doing it professionally. You already have enough. Do not start an 18-month certification expedition. Right now: one small, well-defined project. Outer MD file. Duke it out — want, expect, don't want, gray. Have the AI record that in the file. Then another small, well-defined project. It should go much better, because the file already knows how to have the conversation. That is the point. Rinse. Repeat. Slightly larger each time. Then a real project. The conversation rules are already defined. Now you define the terms of the project, what success is, and what failure looks like so you can avoid it. Keep updating the MD files. Tasks — what well-defined, self-contained tasks contribute to the project? Make a bucket. Define the MD files. Next task. Next task. Pretty quickly you have a successful project. Do not wait for the cavalry. Do not wait for a perfect plan. A good plan now is better than a perfect plan later, which is never. If it is still difficult, it is still you. The MD file lever did not move sufficiently. Wiggle it again. Do-nothing is the first option to review. The more obviously stupid doing nothing looks, the more likely you will actually start. Attack in order to clear up the situation, and gain a basis for further action. The battle itself proves the best method for estimating the enemy. — American intelligence assessment of Nazi military, 1942

27 / 27 · Status as an idiot is established

Arguments are better than questions.

Views and opinions are mine. They do not represent my manager, department, company, or corporation. They are based on information available at the time. I welcome additional information.

01We do not need to review my status as an idiot, as this has been established.
02None of this is scripture. Half the value is in the argument behind the line.
03If one of these lands, ask. You will probably get the fifteen-minute version.

— Mr. Ben Honey

Status as an idiot is established. Arguments are better than questions. Argue with me. Views and opinions are mine. They do not represent my manager, department, company, or corporation. They are based on information available at the time. I welcome additional information. We do not need to review my status as an idiot, as this has been established. None of this is scripture. Half the value is in the argument behind the line. If one of these lands, ask. You will probably get the fifteen-minute version. The rest of the talks, the chargeback lie, the road to hell paved with average, the capacity work: https://mlcu.com/articles.html Mr. Ben Honey. The end. Here are your Mr. Ben Honey hugs and finger kisses.