Why I built a Claude idea machine
The hard part of posting regularly isn’t usually the writing. It’s the empty page before the writing starts.
That sounds almost too simple, but it’s been true for me over and over. If I already have a decent angle, I can sit down and draft. I can clean it up. I can trim the ugly bits, fix the rhythm, make the jokes less embarrassing, and get something publishable out the door. What slows me down is the question that shows up before any of that: what on earth am I going to write about today?
That’s the real enemy. Not the sentence craft. Not the editing. The blank page.
Most people I know who want to post more don’t fail because they can’t write. They stall because they’re trying to invent something worth saying from scratch, every single time. That gets old fast. It also gets weirdly stressful. A person can have a dozen half-finished drafts and still feel stuck, because none of them quite feel worth the effort.
So I started looking for a way to stop treating ideas like rare little museum pieces I had to discover through luck.
A reliable posting habit usually dies in the idea stage, long before anyone complains about the draft.
That’s where my Claude idea machine comes in. I didn’t build it as a shiny AI party trick. I built it because I wanted a practical answer to a very unglamorous problem: I needed a system that could keep surfacing things I might actually want to write about, without making me sit there and stare at a cursor like it owed me money.
This isn’t a generic AI content workflow tutorial, and it’s definitely not me pretending the model does all the work. It’s my behind-the-scenes setup for turning real work into usable material. The point is not to replace thinking. The point is to keep me from spending half my energy deciding what the next post should be.
That distinction matters. A lot of AI content talk gets sloppy right there. People jump straight to “write faster” when the bottleneck is earlier in the process. If the machine spits out words but the topic is thin, you still have a thin post. If the topic is strong, the writing part becomes much less dramatic.
I wanted something that could help me stay consistent without turning every session into a fresh brainstorming session from hell. Claude gave me a way to do that. Not magically. Not perfectly. Just practically, which is usually enough.
In the next section, I’ll show you where the raw material actually comes from, because it’s not random prompts or internet noise. It starts with my week, which turns out to be messier and more useful than I’d like to admit.
The idea engine starts with my actual week
After I figured out that the real problem wasn’t typing, it was choosing what to write about, I stopped looking for clever prompt magic and started feeding the system my actual life. That’s the part most people skip. They ask an AI for ideas in the abstract, then wonder why the results feel generic. I wanted the opposite.
So I built the first layer around a simple habit: I let the Oracle read about a week’s worth of my real work. That means Slack messages, Notion pages, Gmail threads, and meeting notes. The point isn’t to vacuum up everything I’ve touched forever. It’s to catch the little patterns that show up when you’re in the middle of shipping things, answering people, revising plans, and putting out fires. That’s where the decent material hides.
I keep the whole setup inside a Claude Project, which makes it much easier to keep the instructions and context in one place instead of rebuilding the same scaffolding every time. If you’re poking around the basics, Anthropic’s Claude docs are the obvious starting point.
The Oracle doesn’t stop at my own notes, either. It also pulls in outside signals from the X and LinkedIn accounts I follow. That matters more than it sounds like it should. A good idea often shows up when something I’ve been wrestling with internally bumps into a post, a comment, or a short thread from somebody else. Sometimes the outside piece gives me language. Sometimes it gives me a sharper angle. Sometimes it just tells me, “Yep, that thing you’ve been half-thinking about is real.”
Good idea generation usually looks less like invention and more like noticing what’s already been sitting in front of you.
Each day, the system hands me a ranked list of roughly fifteen content spikes. I’m talking about the things that feel worth a closer look because they pop against the rest of the week. A customer question that keeps repeating. A decision I had to explain twice. A meeting note with a weirdly useful phrase in it. A post from someone I follow that sets off an argument in my head. The ranking matters because I don’t want a pile of raw scraps. I want a short stack of things worth my time.
That ranking step is where this starts to feel less like generic idea generation and more like a filter. It doesn’t make up topics from scratch. It digs through what I’ve already said, read, answered, and argued about, then points at the bits that might turn into something publishable. That’s a lot better than waiting around for writer’s block to loosen its grip, which is usually a terrible strategy if you post with any regularity.
I also like that this approach keeps me honest. If the week was thin, the list is thin. If I spent three days in customer calls, the material reflects that. If I read a bunch of sharp posts from people in my feed, that shows up too. The system isn’t pretending to be smarter than my week. It’s paying attention to it.
And that’s really the trick. I’m not asking Claude to invent a personality or hallucinate a content calendar. I’m asking it to spot the scraps that already exist, rank them, and hand me the few that deserve a real look. In the next section, I’ll show you what I had to map out before I could make that work without turning the whole thing into a mess.
I mapped the old workflow before adding AI
Before I let Claude touch a single draft, I drew the whole content process the way it actually existed, not the way I wished it existed. That meant writing down every handoff, every check-in, every place where an idea got grabbed, polished, ignored, revived, or quietly dropped on the floor. I wanted the mess in plain view. If I was going to build a system around this stuff, I needed to know what the system was doing already.
So I started with the boring parts. Where ideas came from. Who noticed them. Where they sat after someone said, “We should write about this.” What happened between a rough thought and a published post. Which steps lived in Slack, which ones lived in Notion, which ones lived in my head, and which ones only existed because nobody had asked whether they were still necessary. I even used Claude’s Artifacts to keep the map visible while I was working through it, which helped because a process diagram in your head tends to grow extra limbs.
Then I tried the uncomfortable version: rebuild it from scratch with no guardrails and no old habits. That’s where the interesting stuff showed up. A few steps turned out to be doing nothing except making the workflow feel official. Others were solving problems that no longer existed. One part was basically me asking for the same idea in three different places, then acting surprised when the output got repetitive. Classic. The whole exercise made it obvious that a lot of the drag wasn’t AI at all. It was me, plus process clutter, plus a couple of decisions made long before I had a better way to work.
That’s the part people miss when they jump straight to automation. AI doesn’t just create bad output out of nowhere. It often puts a bad process under a brighter light. If the inputs are vague, the structure is fuzzy, or the workflow depends on tribal knowledge and heroic memory, the model will usually reflect that right back at you. Maybe with better grammar. Maybe with more speed. Still bad, though. Faster chaos is still chaos.
I’ve found this especially true in content strategy. If you’re trying to build something repeatable, whether it looks more like a Morning Brew-style editorial engine or just a sane way to keep shipping every week, the first job is not “add AI.” The first job is “what are we already doing, and why does half of it feel ridiculous?” Once I stripped the old process down, the weak links were easy to spot. They weren’t hiding. They were just wrapped in routine.
The fastest way to improve an AI workflow is to stop pretending the old workflow was fine.
That’s also why I’m suspicious of people who automate first and diagnose later. It usually means they’re asking a model to clean up a pile they never bothered to sort. A better move is to name the pile, empty it out, and decide what deserves to go back in. I used Anthropic’s prompt generator a few times while testing rebuild ideas, but the real win came from the reset itself. Once the old workflow was mapped cleanly, the next step was obvious: stop feeding the machine vague mush and start giving it a process worth automating.
The interview is where the good stuff comes out
After I mapped the old workflow, the next obvious problem was the part where my ideas usually went stale. I don’t usually run out of words. I run out of shape. A half-formed thought is easy. Turning it into something a reader can actually use is the annoying part.
So I built the interview step to do what a decent editor would do on a good day: keep asking until I stop being vague.
The setup is a panel of six personas. I’ve got Tim Ferriss in there, Joe Rogan, Larry King, Howard Stern, Barbara Walters, and Michael Barrow. They each push in a different direction. Tim wants the system, the process, the repeatable move. Joe keeps pulling for a real story, a specific example, the moment where the thing actually happened. Larry slows it down and asks for clarification. Howard is the one who won’t let me hide behind polished nonsense. Barbara goes after the clean, direct version. Michael Barrow is the steady follow-up machine, the guy who hears an answer and immediately wants the part I skipped.
That mix matters because one generic question rarely gets me anywhere useful. One decent question can still leave me with a sentence that sounds fine and says almost nothing. Six voices, each with a slightly different appetite for detail, make it much harder for me to get away with that.
If the answer still sounds slippery after the interview, the problem usually wasn’t the model. It was me being vague in the first place.
I keep the whole thing grounded with Wispr Flow voice-to-text, which is a very practical part of the process. I talk through the answer instead of trying to type something elegant from scratch. That sounds minor until you try it. Speaking keeps the draft tied to the actual words I’d use in conversation, including the little detours, the half-jokes, and the weirdly specific examples that never show up when I’m staring at a blank doc trying to sound smart.
It also means I can hear when I’m dodging the real answer. If I say, “Well, I guess the main thing is…” I know I’m probably about to ramble. If I can answer in one clean line, great. If not, that’s the cue to keep going until the thought has edges.
I borrowed some of the basic logic from the way Anthropic talks about prompt engineering for business performance, but I’m not trying to turn this into a lab exercise. The point isn’t to be clever with prompts. The point is to force a better conversation. Better questions get better material. Better material gets a draft that doesn’t need to be rescued later.
That’s also why I don’t blame the model first when output feels weak. Most of the time, the model is just faithfully serving up a weak interview. Garbage in, polished garbage out. Slightly better garbage in, slightly better garbage out. Not exactly a dream slogan, but it’s honest.
Once I stopped expecting AI to magically extract substance from a fuzzy thought, the whole thing got a lot calmer. I’m not asking it to invent my point of view. I’m asking it to interrogate me until I’ve actually said something worth writing. Then, in the next step, I can worry about voice and quality control.
My voice, the Writer’s Council, and the anti-slop system
Once I’ve got something worth saying, the next problem is simpler and more annoying: making it sound like me. I’ve never trusted a draft just because it’s clean. Clean can still feel airless. Clean can still sound like a committee wrote it after a long lunch. So I built a voice file in Markdown that acts like a running model of how I actually write.
It’s not fancy. It’s a pile of observations I’ve earned the hard way. My top-performing posts go in there. So do the hook patterns that seem to keep people reading, the structures I return to when I’m being honest instead of theatrical, and the little language habits that keep showing up in my work. That includes a kind of self-deprecating confidence, which is a very specific thing and also, apparently, very much my thing. If you’ve got a personal brand, you probably know the feeling. The point isn’t to manufacture a voice. It’s to stop accidentally sanding off the one you already have.
If a draft sounds polished but not like me, it’s still wrong.
The loop matters because I don’t want the system to freeze my style in amber. After a draft gets published, I compare the draft with the final version and look for the places where the machine guessed right and the places where it wandered off. That comparison gives me better material than any vague “how did this feel?” postmortem ever could. I can see which openings held up, which transitions dragged, which examples landed, and which lines read fine in a draft but died in public.
From there, the lessons file earns its keep. If a pattern keeps showing up, I don’t leave it in a notebook to rot. I fold it back into the system. Maybe a certain kind of opening works when I’m explaining a process. Maybe I keep using a specific sentence shape when I’m being blunt about a mistake. Maybe I have a tendency to over-explain whenever I get excited. That goes in too. The whole setup gets tighter because it keeps learning from my own output instead of pretending every draft is a fresh start.
The fun part, if that’s the word, is the Writer’s Council. I run each draft through six reviewer personas before anything ships. One of them is basically an AI slop allergist, which is exactly as charming and annoying as it sounds. That persona exists to catch the flat, overcooked, vaguely synthetic stuff that can sneak into a paragraph when nobody’s paying attention. The others are there to test different instincts: clarity, tone, pacing, usefulness, and whether the piece still sounds like a human wrote it after a long week and a few too many tabs.
I don’t let the draft out the door unless it clears about a nine out of ten. Not a polite nine. A real one. If it’s an eight and a half, it stays in the shop. That standard sounds fussy until you compare it with the cost of publishing something that reads like it was mass-produced in a hurry. The whole anti-slop system is really just me refusing to confuse output with quality.
If you’re building this kind of workflow in Claude, the Anthropic Build with Claude guide is a useful place to start. I’m using it as a tool, not a crutch, and that distinction matters more than it sounds. The model can help me move faster, sure. What I care about more is whether the draft survives contact with my own standards.
What this changes for distribution and team content
Once I zoom out from the Claude setup itself, the point gets a lot less nerdy and a lot more practical. I’m not building this because I enjoy staring at workflows for sport. I’m building it because distribution is the game, and content is one of the few ways a small team can keep showing up without begging for permission.
A lot of companies leave employee posting on the table, which is a weird choice when you think about it. You’ve already paid for smart people who know the product, the customers, and the weird edge cases. Then you let them sit there with all that context while the company account posts the same polished line for the ninth time. It’s a strange habit. Most teams don’t need more silence. They need more voices.
The best distribution channel for a company is often the people already inside it, if you let them speak in a way that sounds like themselves.
I keep thinking about Storyarb’s “Own the Internet” effort, which drove about two-fifths of inbound leads in a quarter. That number sticks with me because it isn’t about vanity metrics or empty reach. It’s about a team showing up in enough places, often enough, that buyers start finding them before sales ever gets involved. That’s the sort of outcome I care about. Not applause. Pipeline.
Tenex took a similar idea and made it feel less like homework by running the Creator Cup, with a $5,000 prize pool tied to posting. I like that a lot. It gives people a reason to try, a reason to keep score, and a reason to take the awkward first few posts seriously. Half the battle with internal content is getting people past the “who cares what I think?” stage. A little friendly competition helps. So does making it obvious that the company values the effort.
For a bootstrapped company, this matters even more. I can’t count on press to bail me out. I can’t assume investors will turn every announcement into a megaphone. If I want attention, I have to earn it the boring way, one post, one lesson, one opinion at a time. That means content isn’t a side project for me. It’s part of the distribution plan.
And if one strong hire can write, think, and ship in public, that person may pay for the whole experiment before long. That’s the math I keep coming back to.
