
Sign up to save your podcasts
Or


The Picture Was Right. The Prompt Was Wrong.
A good-looking result can hide a broken process.
In this episode, I look at a production test where the final image was correct — but the prompt behind it contained contradictory instructions. The model happened to choose the version I wanted.
That raised a harder question: when AI gives you the right answer, how do you know it got there for the right reasons?
I also get into what happened next when the same image needed to be reused for video production. The safeguards designed to protect the workflow started blocking legitimate work, which forced a clearer separation between capability, permission, budget approval and actual provider calls.
In this episode:
Why a successful AI output can still expose a faulty system
How conflicting references can quietly compete inside a prompt
Why different reference images need clearly defined roles
Why I kept the correct image instead of generating it again
What happened when safety guardrails blocked legitimate asset reuse
The difference between capability and authority
Why an AI system being technically able to do something does not mean it should automatically do it
Why paid or irreversible steps should still have human approval points
Why passwords and secrets should not simply be dropped into AI conversations
How “nothing happened” can sometimes be evidence that a safety control actually worked
Why failed attempts should remain visible instead of being overwritten
Why I now care less about whether something “worked” and more about what exactly worked, under which authority, and with which evidence
The main lesson from this stage was simple:
A reliable AI workflow should not just produce the right output. It should make the right output repeatable for the right reasons.
At the documented stopping point, the image was accepted, the broken prompt was repaired, and the next paid video action was still correctly gated.
Not a perfect ending.
An accurate one.
Next episode:All 178 Tests Passed. The Fake Database Was Wrong.
Episode summary
I split the work between two AI assistants: ChatGPT would help structure the plan, and Claude would handle the bounded building.
It sounded like a team.
Instead, I became the API between them — copying instructions, returning results, carrying errors back and deciding what happened next.
This episode is about a less glamorous part of working with multiple AI systems: coordination. The individual tools can be capable while the overall workflow still creates more work for the human in the middle.
In this episode
Why two capable AI assistants can still create a bad workflow
How a non-technical founder ended up acting as the message bus between systems
Why the simplest available technical route should usually be tried first
How automation can succeed internally while making the human handoff worse
What Mr. Blue Builds itself taught me about reducing production friction
Why a useful AI workflow needs one clear owner of the next executable step
How excessive rules and safety gates can create confusion instead of clarity
Why I now prefer a controlled relay over the promise of a fully autonomous pipeline
Practical takeaway
A multi-AI workflow is not successful because several intelligent tools are involved.
It is successful when each tool returns a defined result, one role owns the next step, and the human receives one clear decision instead of becoming the connection layer between the machines.
Next episode
The Picture Was Right. The Prompt Was Wrong.
Once the relay became clearer, a more subtle problem appeared: the correct tool could produce something that looked correct while the instructions underneath were still wrong.
In this episode:
• Why finishing the software does not mean you have finished the company.
• How internal systems can become a comfortable way to avoid doing the work customers actually see.
• Why technical progress and visible business value need to develop alongside each other.
• Why “designs created”, “products uploaded” and “tests passed” can become misleading success metrics.
• Why a design only really proves market value when somebody buys it.
• How support burden can reveal whether a self-service product is actually scalable.
• Why dashboards, labels and empty states must tell the truth — even when that truth is incomplete.
• Why “unavailable” can be more useful than an invented number.
• The two-track rule I now use: technical progress on one side, real-world value on the other.
Practical takeaway:
Do not measure the work around the outcome as if it were the outcome itself.
Pages built are not customers helped.
Products uploaded are not products sold.
Automations created are not human effort removed.
Tests passed are not real behaviour verified.
Content planned is not content published.
A useful metric should measure evidence of the result you actually wanted.
At this point, the Blue Corp. headquarters existed and was frozen. Finance was the only verified live source connected to it, while the company-specific snapshots were still unavailable.
That was not the finish line.
It was the first honest starting line.
Next episode:
I Had Two AI Assistants. Somehow I Became the API.
Episode summary
After deciding to stop repairing the old system, Mr. Blue starts again — but this time the goal is not to rebuild everything.
The new Blue Corp. headquarters begins deliberately small. Finance has to be checked before the dashboard can depend on it, a tiny “Untitled” identity problem is enough to stop the build, and several apparently clean AI-assisted steps are rejected rather than allowed to spread into the next layer.
The result is not a frictionless rebuild. It is a different kind of struggle: smaller steps, tighter boundaries, more scepticism, and mistakes that are stopped before they become architecture.
In this episode
• Why starting again meant building less• Why Finance became the gate before the new HQ• How one small “Untitled” detail stopped the build• Why scepticism of clean AI answers made the process better• What still required human verification• Why the finished HQ deliberately shows what does not exist yet• How a better method changes the cost of being wrong
Practical takeaway
Make the next AI-assisted step small enough to inspect. When something does not fit, stop and correct it before that assumption becomes an input to the next layer. A better process does not eliminate mistakes — it keeps one mistake from becoming ten more.
Next thread
The headquarters is now approved, locked and frozen until real company data is ready to connect. With the HQ no longer pretending the companies are finished, the next work can return to building the companies themselves.
Disclosure
Mr. Blue is an AI-assisted fictional host created by Blue Studios. The story is based on documented project decisions and real events from the development of BlueOS Corp. AI tools were used in the writing, voice and production workflow under human direction and approval.
In Episode 5, I confront the uncomfortable difference between the AI company I thought I had built and the system that actually existed.
The original BlueOS app took roughly three days to build—and about twenty days to investigate and repair. Along the way, I found dashboards presenting assumptions as facts, an unnecessary access layer that locked me out, and a live system that did not always match the version being worked on.
The problem was not one spectacular bug. It was the method.
This episode is about digital archaeology, misleading AI promises, sunk costs, and the decision to stop repairing a system built on assumptions—and start again with documentation, evidence, clear boundaries, and verification.
Episode summary:
After discovering that AI chat could not safely serve as BlueOS Corp.’s permanent organisational memory, Mr. Blue built an external Brain: maintained documents, authority rules, decision logs and exact handoffs. It improved continuity. Then a second problem appeared: the memory itself needed governance.
In this episode:
• Why saving information is not the same as governing it• Why document authority matters as much as document content• What a real MBB storage cleanup revealed about duplicates and superseded assets• Why old material should be archived rather than silently deleted or left active• How exact handoffs move maintained memory into authorised action• Why material write-backs need verification
Practical takeaway:
An external source of truth only works when it has governance. Give maintained documents clear roles, define which source outranks which, separate active masters from working material, archive superseded files, write exact handoffs and verify important write-backs.
Episode summary
ChatGPT helped Mr. Blue build pages, workflows, documents, rules and decisions across BlueOS Corp. The work was real, but the company gradually discovered that useful conversational context was not the same as durable organisational memory.
Episode 3 examines how partial context created repetition, duplicate truth sources and unsafe continuity — and why the company separated conversation, documentary authority and operational status.
In this episode
• Why a productive chat can feel like permanent memory
• How partial context is more dangerous than no context
• Why polished documents can still conflict with each other
• The difference between a working conversation and an authoritative source
• How Brain, Base44 and chat now have separate roles
• The practical memory controls a non-technical founder can adopt
Practical takeaway
Treat AI chat as a working interface, not as the sole source of organisational truth. Store approved decisions, current status, restrictions and exact next actions in an external system you control, then verify that the write-back actually happened.
Next episode
So I Built a Brain. It Did Not Solve Everything.
The external memory system improved continuity, but it also created new problems around structure, duplication, authority and maintenance.
Disclosure
Mr. Blue is an AI-assisted fictional host created by Blue Studios. The story is based on documented project decisions and real events from the development of BlueOS Corp. AI tools were used in the writing, voice and production workflow under human direction and approval.
Episode summary
Mr. Blue expected AI to turn a production idea into a working app within days. BlueOS grew quickly into a real platform with pages, data and production functions — but visible progress was repeatedly mistaken for operational reliability.
Episode 2 examines why an AI-generated interface is only the beginning, how real production exposed hidden weaknesses, and why evidence must replace optimistic “ready” labels.
In this episode
• Why AI-generated apps feel complete before they are reliable
• How individually sensible features created an oversized system
• The difference between a visible page and a verified workflow
• What BB001 proved — and what it exposed
• Why BlueOS moved to a verification-first repair method
• The minimum evidence a non-technical founder should demand
Practical takeaway
Define the smallest job the app must perform reliably and the evidence required to prove it. Verify data, repetition, failure handling and recovery before adding more pages or automation.
Next episode
I Let ChatGPT Build Everything — Then Discovered It Couldn’t Remember
BlueOS had grown through fast AI-assisted building across pages, workflows, documents and decisions. The next problem was maintaining a reliable understanding of why everything existed.
Disclosure
Mr. Blue is an AI-assisted fictional host created by Blue Studios. The story is based on documented project decisions and real events from the development of BlueOS Corp. AI tools were used in the writing, voice and production workflow under human direction and approval.
Episode summary
Mr. Blue began with a simple plan: create a faceless YouTube channel using AI. The project quickly expanded into a media studio, an operating system, a company structure and an authoritative Brain.
Episode 1 examines how reasonable individual decisions combined into an oversized system—and why proving one repeatable output should come before building the infrastructure around it.
In this episode
• Why a faceless channel still requires a consistent identity
• How one video became a production system
• Why Blue Studios, BlueOS Corp. and BlueOS were created
• What happened when company knowledge became fragmented
• Why the authoritative Brain and Studio Memory serve different purposes
• The risk of building the factory before proving the product
Practical takeaway
Before expanding the infrastructure, prove that the smallest useful output can be produced consistently, reviewed efficiently and published repeatedly.
Next episode
Apparently an AI App Takes Longer Than Three Days
Mr. Blue had concluded that the growing company needed an app. Fortunately, AI appeared capable of building one almost immediately.
It was not.
Disclosure
Mr. Blue is an AI-assisted fictional host created by Blue Studios. The story is based on documented project decisions and real events from the development of BlueOS Corp. AI tools were used in the writing, voice and production workflow under human direction and approval.