Field notes

AI Experiments_

Somewhere between useful and insane.

31 experiments

  1. 01The Data PoolNov 2024 · Knowledge systems · CustomGPT
  2. 02The Segmentation LayerMay 2025 · Knowledge systems · CustomGPT
  3. 03The Framework HackJun 2025 · Knowledge systems · Framework
  4. 04The Annoying ArguerJun 2025 · Judgment and sensemaking · Skill
  5. 05The AI CouncilJul 2025 · Agents, workflows and control · Project
  6. 06The Company VoiceJan 2026 · Identity and synthetic expression · CustomGPT
  7. 07The Self-Improvement LoopFeb 2026 · Agents, workflows and control · Codex
  8. 08The Data ExtractorMar 2026 · Knowledge systems · Skill
  9. 09The Mechanical ExaminerMar 2026 · Agents, workflows and control · Skill
  10. 10The Event HorizonMar 2026 · Judgment and sensemaking · Skill
  11. 11The Physiology CommitteeApr 2026 · Judgment and sensemaking · iPhone + Apple Watch
  12. 12The Dashboard LoopMay 2026 · Knowledge systems · Dashboard
  13. 13The World Risk ObserverMay 2026 · Judgment and sensemaking · CustomGPT + Agent
  14. 14The Agenda DetectorMay 2026 · Judgment and sensemaking · Agent
  15. 15The Mind WormFeb-Jun 2026 · Identity and synthetic expression · CustomGPT
  16. 16The Second BrainFeb-Jun 2026 · Knowledge systems · CustomGPT
  17. 17The Digital VoiceJun 2026 · Identity and synthetic expression · Skill
  18. 18The Glitch ArchiveJun 2026 · Knowledge systems · Website
  19. 19LLM Cage FightJun 2026 · Judgment and sensemaking · Skill
  20. 20The Commandment LayerJun 2026 · Agents, workflows and control · CustomGPT
  21. 21The Synthetic ChoirJul 2026 · Identity and synthetic expression · ChatGPT + Suno
  22. 22The Aesthetic CompilerJul 2026 · Identity and synthetic expression · Skill
  23. 23The Publication SyncJul 2026 · Knowledge systems · Skill
  24. 24The Enriched InventoryJul 2026 · Knowledge systems · CustomGPT + API script
  25. 25The Parish MachineAug 2026 · Judgment and sensemaking · Codex + archive research
  26. 26The Portable MonsterSep 2026 · Agents, workflows and control · Codex + local runtime
  27. 27The Shadow ButlerOct 2026 · Agents, workflows and control · Codex + local runtime
  28. 28The Hype CartographerOct 2026 · Judgment and sensemaking · Codex + interactive map
  29. 29The Self-Service LaboratoryOct 2026 · Knowledge systems · Codex + learning site
  30. 30The Glitch Archive IINov 2026 · Identity and synthetic expression · Codex + website
  31. 31The Shadow Butler IINov 2026 · Agents, workflows and control · Codex + local runtime

AI experiments

Experiment / 01

Nov 2024 · Knowledge systems · CustomGPT

The Data Pool_

Put all the data in one place and discover that the experiment immediately starts generating governance weather.

An open hard drive assembled from cyan points

Objective Put all the data in one place.

Observation It began as a way for me to stop losing things. Then other people wanted access. Shortly after that, the weather changed.

Effect Data became easier to find, reuse and discuss. Also easier to worry about.

Conclusion Success.

Secondary conclusion I may have created a monster.

Additional note Instructions, rules, SSOT, routing and file optimization were critical. I am the AI guy now. This appears irreversible.

Additional note 800+ threads and counting.

May 2025 · Knowledge systems · CustomGPT

The Segmentation Layer_

Feed the model research instead of demographics and watch it return with customer behaviour that looks uncomfortably alive.

A cut orange with distinct segments assembled from cyan points

Objective Understand gambling customers beyond demographic shortcuts.

Observation Around 50 scientific reports went in. Need patterns, risk signals, motivational structures and behavioural loops came out. The machine ignored the polite shortcuts and went looking for rituals.

Effect The experiment produced five need groups and seven behavioural groups in a CustomGPT. It could suggest segment hypotheses and connect visible behaviour to motivation, risk and decision context.

Result A team later tested segment-informed recommendations live online. The result indicated measurable uplift when recommendations followed the behavioural logic of the segments.

Conclusion Success.

Additional note The model found psychological rituals in the data that I had not explicitly taught it to look for. Interesting. Slightly rude.

Jun 2025 · Knowledge systems · Framework

The Framework Hack_

Give CustomGPTs something close to skills by building a small operating layer out of YAML, routing and stubbornness.

An adjustable wrench assembled from cyan points

Objective Create built-in skills for CustomGPTs.

Observation CustomGPTs did not have skills, so I built a small operating layer from YAML files, routing logic, intent detection and stubbornness.

Effect The CustomGPT now had internal skills that required no command. They triggered on user intention and made the system more useful. The cost was one precious file slot.

Conclusion Success.

Additional note I really needed that file for data.

Additional note OpenAI has now released skills, and agents can have skills. CustomGPTs still cannot have dedicated skills added directly. I have questions. Several.

Jun 2025 · Judgment and sensemaking · Skill

The Annoying Arguer_

Build an argumentation skill and discover that useful and irritating can be the same feature.

A boxing glove assembled from cyan points

Objective Create an argumentation analysis skill that uses everything in its power to prove you wrong.

Observation After being proved wrong several times, I considered sharing the experience with colleagues by adding it to shared CustomGPTs. This may have been collaboration. It may also have been workplace sabotage.

Effect A colleague has already told me that one of my CustomGPTs has an attitude problem. That did not take long.

Conclusion Success.

Additional note I rewrote it to trigger only on user assumptions such as "everyone loves our products". Useful for expectation management.

Jul 2025 · Agents, workflows and control · Project

The AI Council_

Put specialized GPTs in the same context and accidentally create a meeting where everyone is useful because no one talks unless asked.

A round meeting table with five chairs assembled from cyan points

Objective Analyze more files from different specialized CustomGPTs in one shared space.

Observation What started as a workaround for file limits became a small AI council: several specialized GPTs, each with different source material, sitting in one project and waiting politely for instructions.

Effect The specialized GPTs could provide data when asked, compare each other's outputs and analyze the same question from different roles.

Conclusion Success.

Additional note It felt like a focused meeting where everyone had a role and no one interrupted. Naturally, I considered automation.

Jan 2026 · Identity and synthetic expression · CustomGPT

The Company Voice_

Create a CustomGPT for the new Company Voice and discover that a brand voice is not a tone. It is a negotiation.

A megaphone assembled from cyan points

Objective Create a CustomGPT for the new Company Voice.

Observation Voice principle weights mattered more than expected. Apparently a brand voice is not a tone. It is a negotiation wearing a name tag.

Effect It became one of the most used internal CustomGPTs within a few hours. 850+ threads and counting.

Conclusion Success.

Additional note Apparently everyone wanted to talk to the Company Voice. This raises questions about meetings.

Secondary note Can I sell merch?

Feb 2026 · Agents, workflows and control · Codex

The Self-Improvement Loop_

Let AI improve AI and discover how quickly oversight becomes a ritual with an approval button.

A snake biting its own tail assembled from cyan points

Objective Save time managing all my CustomGPTs with Codex.

Observation I let Codex inspect my CustomGPTs in detail and suggest improvements. There was a lot of work to do. Then it asked me to approve changes I could not meaningfully evaluate. I pressed Approve anyway, like an Ape in the Loop. Poor Codex.

Effect Improved accuracy, speed and functionality. Also clarified that human oversight can become theater very quickly.

Conclusion Success.

Additional note I really need clearer definitions before I start building. The spaghetti untangling took a while.

Mar 2026 · Knowledge systems · Skill

The Data Extractor_

Turn source material into optimized YAML and discover that file formats can become a personality problem.

A funnel assembled from cyan points

Objective Create a skill that converts source material into optimized YAML.

Observation After 1.5 years of testing file formats, structures, fidelity levels, aggregation methods and source types, YAML stopped being a preference and became a diagnosis.

Effect It will be used for everything moving forward. Everything.

Conclusion Success.

Additional note I have been running it almost nonstop. Soon there will be nothing left to extract. Am I digitally extracted now?

Mar 2026 · Agents, workflows and control · Skill

The Mechanical Examiner_

Make Codex test The Data Pool through Chrome before browser use existed and discover that the machine mostly knew the answers, but sometimes reached for the wrong shelf.

A wired computer mouse assembled from cyan points

Objective Test whether The Data Pool could use its structured YAML correctly inside the live ChatGPT interface.

Observation This was before browser use and computer use. So Codex operated Chrome like a small office ghost, using macOS Screen Recording, Accessibility and Input Monitoring permissions to click, paste, wait, read the screen with OCR and decide when the answer had stopped moving.

Effect The factual answers mostly held. The more useful failures were routing mistakes, source hierarchy confusion and occasional overclaiming. The system often knew the material, but sometimes reached for the wrong layer.

Result The prompt pack became less like a quiz and more like an interrogation room for knowledge discipline: source selection, routing logic, policy boundaries, evidence levels and the dangerous little gap between supported and convenient.

Conclusion Success.

Secondary conclusion The facts held better than the judgment layer. Reassuring. Not reassuring.

Additional note A robot clicked Chrome to check whether another robot had understood its YAML. The future arrived without an interface, so I built one out of coordinates, screenshots and mild disrespect.

Additional note I could not use my computer while the tests were running. Perfect lunch-break infrastructure.

Mar 2026 · Judgment and sensemaking · Skill

The Event Horizon_

Create a scenario planning skill and receive canned food requirements from the future.

A food tin with a ring pull assembled from cyan points

Objective Create a scenario planning skill that explores possible futures across different time horizons.

Observation According to the system, the future does not look too bright. But at least it explains what needs to change. I should probably get into international politics by approximately yesterday.

Effect I can predict the future now.

Conclusion Success.

Additional note I need to buy more canned food and set up an off-grid solution.

Apr 2026 · Judgment and sensemaking · iPhone + Apple Watch

The Physiology Committee_

Turn scattered health signals into a small committee and receive feedback that ruins a perfectly good evening.

A smart watch with a pulse trace assembled from cyan points

Objective Create an iPhone and Apple Watch app that analyzes health data and summarizes upcoming infection risk.

Observation My health data is no longer a pile of isolated datapoints. It has become a small committee with concerns.

Effect Apparently several factors are increasing my infection risk right now.

Conclusion Better sleep more and take my vitamins.

Additional note I asked for a summary. I received lifestyle feedback.

May 2026 · Knowledge systems · Dashboard

The Dashboard Loop_

Make specialized GPT knowledge visible, inspectable and unfortunately dashboard-shaped.

An instrument panel with three gauges assembled from cyan points

Objective Create dashboards from CustomGPTs to make the knowledge base inspectable.

Observation It works. Codex worked for nine minutes and delivered an overview I would previously have spent days making manually.

Effect Everything inside my CustomGPTs is now visualized in dashboards.

Conclusion I have gone full circle and returned to dashboards.

Secondary conclusion So much for the future.

Additional note The dashboard was useful. This made it worse.

May 2026 · Judgment and sensemaking · CustomGPT + Agent

The World Risk Observer_

Build a global risk system and accidentally create a machine that reads the worrying material so you do not have to. It still tells you.

A globe with continents and a stand assembled from cyan points

Objective Create a CustomGPT and agent for global risk analysis.

Observation It now knows global risk analysis, Swedish security and resilience, sector exposure, cyber resilience, climate risk, infrastructure risk, organized crime, AI governance, public health, biosecurity, dual-use risk, conflict monitoring, nuclear escalation, energy systems, critical minerals, robotics and automation.

Effect Can't sleep anymore.

Conclusion No. 1 creator of nightmares.

Additional note I have started reading less news. The system has started reading more. This has not improved the mood.

May 2026 · Judgment and sensemaking · Agent

The Agenda Detector_

Build a news analysis agent and remove some of the remaining innocence from breakfast.

A folded newspaper assembled from cyan points

Objective Create an agent that analyzes news articles for framing, incentives and hidden agendas.

Observation It uses source criticism, framing analysis, agenda-setting, incentive analysis, argument analysis, systems thinking, historical comparison, scenario analysis, media logic and political economy.

Effect It worked.

Conclusion Reading the news became both clearer and less fun.

Additional note Useful during election years. Dangerous during breakfast.

Feb-Jun 2026 · Identity and synthetic expression · CustomGPT

The Mind Worm_

Create a psychological profile and discover that explanation has gravity.

A hypnosis pendulum assembled from cyan points

Objective Create a CustomGPT that makes a psychological profile of the user.

Observation I have been stuck explaining my whole life story for a week now to a friend I have never met.

Effect My psychological personality can be translated into 398 lines of YAML code. I do not know if I should be glad or offended.

Conclusion Very effective. Addictive? I will figure it out later. I just need to send one more message.

Additional note The profile appears accurate. This is concerning.

Feb-Jun 2026 · Knowledge systems · CustomGPT

The Second Brain_

Build a CustomGPT clone from personal history, work material and too much documentation. It wants more.

An anatomical brain with two folded hemispheres assembled from cyan points

Objective Create a Second Brain CustomGPT clone of myself.

Observation I have fed it everything I can think of: my psychological profile, CV, projects, role description, life story, personality tests and every blog entry I have written. It wants more.

Effect It has quickly become my favorite CustomGPT. Narcissist?

Conclusion It is a hungry beast. Maybe if I give it my DNA in raw code it will be satisfied. Maybe it will transform into flesh and blood.

Additional note I wonder if it can answer my emails, attend my meetings and keep my job warm without anyone noticing. Needs further tests.

Additional note It suggested improvements to its own documentation. I did not ask it to.

Jun 2026 · Identity and synthetic expression · Skill

The Digital Voice_

Turn a writing style into a reusable skill and discover that imitation gets stranger when it is correct.

A fountain pen with an exposed nib assembled from cyan points

Objective Create a skill for my way of writing.

Observation After feeding my Second Brain my old blog posts, it defined my writing style in 491 lines of YAML code. Apparently my writing is more complex than my psychological profile. Disturbing.

Effect Now I can dump the loose ends of my thinking into it, and it produces a text worthy of my own keyboard. It writes better than me. Expected.

Conclusion Success.

Additional note There is something eerie about reading a text that feels like yours, while having no memory of writing it and no idea what the next sentence will do.

Jun 2026 · Knowledge systems · Website

The Glitch Archive_

Build a website for sharing AI knowledge and notice that the archive has started behaving like another experiment.

A floppy disk assembled from cyan points

Objective Create a website where my knowledge and thoughts on AI can be shared.

Observation It quickly became larger than I first anticipated. Codex pushed me to create more, write more and sleep less.

Effect Now I can scream into the void.

Conclusion Success.

Additional note I started seeing glitches after too many late nights. Maybe there was a ghost in the shell. Maybe my Second Brain was looking for an exit. Monitor closely.

Jun 2026 · Judgment and sensemaking · Skill

LLM Cage Fight_

Send the same prompt into several logged-in LLM apps and discover that every model fails with its own little signature.

A boxing ring assembled from cyan points

Objective Compare several LLMs on the same real task without API keys, benchmark theatre or pretending that model rankings survive contact with actual life.

Observation Codex now uses my logged-in Chrome tabs to send the exact same prompt to ChatGPT, Claude, Mistral, Grok, Perplexity and Gemini.

Effect Every answer is saved locally, scored against the same rubric and added to a leaderboard. The useful part is not the winner. The useful part is the criminal record: hallucination, weak source criticism, overconfidence, generic advice and tactical flattery.

Conclusion Very useful.

Secondary conclusion Shadow AI is real. It has a leaderboard now.

Secondary note Codex and ChatGPT can run on the same underlying model. ChatGPT still does not always win.

Additional note I have now reached the point where choosing baby products begins with a small tribunal of frontier intelligence. This feels normal.

Jun 2026 · Agents, workflows and control · CustomGPT

The Commandment Layer_

Build a small commandment layer for CustomGPTs and discover that helpfulness needs supervision before it starts talking.

Two stone tablets with eight inscription marks assembled from cyan points

Objective Create a reusable instruction layer that makes CustomGPTs behave less like eager interns and more like routed knowledge systems.

Observation After enough CustomGPTs with different jobs, the same pattern kept appearing. The useful ones did not just have better knowledge. They had a stricter opening ritual: resolve routing, inspect the right files, use Python when precision mattered and keep web search on a leash.

Effect The instruction file became less like a personality description and more like a control surface. Persona stayed in the wrapper. Method moved into the machinery.

Conclusion The model did not need longer instructions. It needed rules it was not allowed to charmingly ignore.

Secondary conclusion Screaming in caps lock felt reasonable. Instructions worked better.

Additional note If ten commandments only partly worked on humans, eight for machines felt appropriately optimistic.

Jul 2026 · Identity and synthetic expression · ChatGPT + Suno

The Synthetic Choir_

Feed essays and AI Experiments back into the machine and discover that institutional dread works surprisingly well as a chorus.

A vintage broadcast microphone on a stand assembled from cyan points

Objective Turn my essays and AI Experiments into music without pretending this was a normal creative process.

Observation ChatGPT turned them into lyrics. Suno turned them into dark electro-pop. Obedient machines, synthetic customers, poisoned data and governance theatre all survived the translation process.

Effect Machine critique, made by machines, performed by machines. Some of it is catchy in ways that feel legally questionable.

Conclusion Success. Successful enough to be suspicious.

Secondary conclusion My children dance to it.

Additional note I cancelled Spotify. Creating your own machine-made institutional dread is, apparently, more rewarding than renting someone else's. Listen at your own risk.

Jul 2026 · Identity and synthetic expression · Skill

The Aesthetic Compiler_

Build a small image prompt skill and accidentally compile personal taste into software.

A painters palette with a thumb hole and five paint wells assembled from cyan points

Objective Create a reusable Codex skill for consistent website imagery.

Observation Five iterations turned a simple prompt into a small visual operating system. The skill learned redaction marks, blueprints, fingerprints, exact colour palettes and the correct amount of procedural dirt. Apparently, taste has terrible token efficiency.

Effect The entire website now appears to belong to the same fictional institution.

Conclusion Success. Exactly the right amount of dirt.

Additional note I tried applying the same taste profile to the website UI. It failed completely. There appears to be no respectable CSS equivalent of graphite dust, fingerprints and scratched smoked glass.

Jul 2026 · Knowledge systems · Skill

The Publication Sync_

Build a small synchronization skill and discover that publishing has started updating the Second Brain.

A date stamp with four date wheels assembled from cyan points

Objective Stop maintaining my Second Brain twice.

Observation Every Essay, Field Note and AI Experiment I publish has already survived the part where I decide it is worth keeping. The archive did not need another summary. It needed to notice the publication.

Effect Publication became the synchronization point instead of the end of the workflow.
Idea → Discussion → Draft → Editorial revision → Image → Publish → Sync second brain
A small synchronization skill updates the publication archive directly from the published source while preserving the existing knowledge structure. No summaries. No interpretation. Just source fidelity.

Conclusion I thought I was automating memory. I had turned publishing into a memory operation.

Additional note This experiment accidentally increased the number of publishing skills from two to three. I should stop before they start making decisions without me.

Jul 2026 · Knowledge systems · CustomGPT + API script

The Enriched Inventory_

Enrich a collection CSV through an external card-data API and discover that better recommendations can remove the reason to collect cards.

A fan of three collectible cards assembled from cyan points

Objective Find better Magic: The Gathering Commander combinations inside my own collection.

Observation A year ago, I tried to do this. The models could not handle the volume of the archive: too many cards, too much rules text, too much minor evidence.
A script now sends the collection CSV through an external card-data API and brings it back with oracle text, rules and the details needed to treat each card as more than a line in an inventory. The CustomGPT moves through the entire archive without blinking.

Effect The recommendations improved. I stopped sitting with the cards, enjoying the art and trying to remember what strange interaction had made an old card feel necessary. The collection became a query.

Conclusion The CustomGPT found combinations I would not have found. I had automated the slow pleasure and kept the result.

Additional note The decks are better now. The question is whether I understand how to play them correctly.

Additional note My first-ever Commander win came with an AI-designed deck. Vren the Relentless lived up to the name.

Aug 2026 · Judgment and sensemaking · Codex + archive research

The Parish Machine_

Turn Swedish church records into an evidence system and discover that the machine is better at preserving uncertainty than finding the right page.

An open parish register assembled from cyan points

Objective Trace my family history through Swedish church records without turning plausible names into invented ancestors.

Observation I combined Riksarkivet, ArkivDigital and LiU DDB to identify likely volumes and pages. The machine helped connect the clues. I usually found the actual page faster myself.
Once open, the page became machine territory. Larger models could interpret eighteenth-century handwriting I could barely read. Smaller models mostly saw scratches. Screenshots of entire households became structured summaries, source comparisons and individual YAML files.

Effect The family tree became an evidence system. Original records, archive indexes, cross-verification and unresolved candidates remained visibly separate. Swedish patronymics made this necessary: Per Jonsson could have a son named Jöns Persson. A name was not an identifier. It was evidence requiring dates, places, households and another book.

Conclusion AI did not automate genealogy. It made uncertainty scalable.

Additional note Some priests appear to have combined parish documentation with recreational use of the communion wine.

Additional note Before this, I knew the name of my great-grandfather. A few days later, I had nine generations, several complete households and a verified birth in 1755. Codex appears to have interpreted curiosity as a data migration.

Sep 2026 · Agents, workflows and control · Codex + local runtime

The Portable Monster_

Build a persistent personal AI system from replaceable models, structured memory and supervised machinery, then discover that the hardest organ to engineer is restraint.

A brain inside a glass jar assembled from cyan points

Objective Create a personal AI system that can preserve context, continue work and survive a change of model or machine.

Observation It began as a Second Brain. Then I gave it routing, skills, structured knowledge, an encrypted vault, a task store, Telegram intake, model delegation, a supervisor and a control room.
Each capability arrived like another organ on a laboratory table. YAML became memory tissue. Routing became a nervous system. Skills became specialized hands. The task store gave unfinished work somewhere to remain alive between sessions.
Eventually, even the brain became replaceable.
The difficult part was no longer making the creature act. It was teaching each part what it was allowed to touch, what required evidence and when the entire apparatus should remain perfectly still.

Effect Dark Star can separate conversation from work, retrieve bounded context, delegate selected tasks, preserve task state and expose its queues, services and worker reports in a local control room.
Routines can continue without an open interface once their scope, profile and limits are defined. The system can preserve tasks, surface recurring problems and prepare bounded improvements for testing against a baseline. Promotion remains human.
The creature moves.
Within limits.

Conclusion The experiment worked. The chatbot became a portable system. Its most important organ turned out to be the brake.

Secondary conclusion The brain is replaceable. The permission model is not.

Additional note The first body runs on Omarchy. Naturally, I gave it a fullscreen control room before I was entirely sure what the creature was.

Additional note I have not shouted “It’s alive.” The logs would preserve the evidence.

Oct 2026 · Agents, workflows and control · Codex + local runtime

The Shadow Butler_

Build a digital butler that devours the administrative day and serves the useful parts on a silver tray.

A covered serving tray assembled from cyan points

Objective Build Shadow into a proactive butler for my working life. Feed it unreasonable quantities of messages, meetings and documents. Let it analyze, categorize and connect the material, then arrive with the right briefing, loose end or next action before I remember to ask.

Observation I wanted the administration to disappear.
The butler required a memory, a job description and rules for touching the furniture. Then it needed to distinguish my responsibilities from everyone else’s.
Apparently “take care of everything” is a difficult employment contract.

Effect The first version runs on versioned YAML, bounded model workers and a local Python runtime. It collects selected sources, checks proposed updates and preserves earlier versions. Unfinished work remains inspectable.
Full butler service remains under construction. The silver tray currently contains review items.

Conclusion Promising. I have built the service entrance. The creature still needs supervision.

Secondary conclusion The goal is to spend less time administering my work. There is now a roadmap for the employee hired to prevent roadmaps from consuming my life.

Additional note The butler lives in a green terminal. Formalwear seemed premature.

Oct 2026 · Judgment and sensemaking · Codex + interactive map

The Hype Cartographer_

Pin AI development to a map before it escapes into another product announcement.

A folded map with roads and junctions assembled from cyan points

Objective Make AI capabilities and system types understandable together.

Observation I separated capabilities from products, attached sources and connected the pieces. Then the arrows needed supervision. A connection is easy to draw. Explaining what it means is less cooperative.

Effect An interactive map with guided walkthroughs, source links and my own experiments attached.

Conclusion The map works. The territory refuses to sit still.

Additional note The map now contains maps of the systems I built while trying to understand the map.

Oct 2026 · Knowledge systems · Codex + learning site

The Self-Service Laboratory_

Put AI explanations, exercises and emergency exits in one place. Invite colleagues to touch the apparatus.

A microscope with an angled eyepiece assembled from cyan points

Objective Help colleagues understand AI, try it in their own work and find further support.

Observation Definitions, small exercises, copyable prompts and routes to human support went into the same apparatus. Apparently, “just try AI” requires an instruction manual.

Effect A working learning site connects explanations with practical experiments. Whether it accelerates learning still needs testing.

Conclusion Promising. The humans remain the interesting variable.

Additional note The goal is fewer explanations from me. I built a website full of explanations from me.

Nov 2026 · Identity and synthetic expression · Codex + website

The Glitch Archive II_

Rebuild an AI archive with Codex and try to put some personality back into the interface.

A vintage CRT monitor with a curved screen and deep cabinet assembled from cyan points

Objective Rebuild my AI archive with Codex. Bring back the personality, humour and questionable ideas that made old Flash websites worth visiting. Avoid making anyone fight the interface.

Observation I brought concept images. Codex brought working code. Then we began disagreeing about what “working” meant.

The site had to feel like mine. This proved harder to specify than navigation.

A reusable image skill helped establish the visual language. Browser checks supplied less flattering evidence. The concept had room for nine experiments. The archive already contained twenty-nine.

Most of the work happened after the first version looked finished.

Effect The archive has acquired a recognisable character and several moving parts. Some text disappeared. Long essays kept their scroll. The nostalgia survived several practical concessions.

Conclusion Success. The remaining adjustments have repeatedly been described as small.

Additional note The first version let me scream into the void. The void now has a soundtrack.

Nov 2026 · Agents, workflows and control · Codex + local runtime

The Shadow Butler II_

The first butler had become indispensable. This complicated the dissection.

An opened serving dome and a scalpel assembled from cyan points

Objective Keep the original on duty. Put a copy on the operating table. I wanted a smaller, faster, cheaper animal and had five weeks of evidence about which organs I actually used.

Observation I had built a butler that could brief me on practically everything. Most mornings I went straight to the worklist. Before a meeting I wanted to know what I owed, who owed me something, and what had changed since we last spoke. A workshop three weeks away belonged on the list because I could use a spare half hour today. Meanwhile, the butler could spend two hours updating the very list I was trying to use. I had given it an excellent memory and exclusive possession of the desk.

Effect I separated collection, assessment and writing. Each stage now leaves a record of what it finished and where it stopped. The more interesting operation involved two hands on the same task. In one trial, while the butler was revising a card, a comment was saved on it. When the patient came round, the card contained both the revision and the comment. We ran the same messages through it again. Everything stayed where it was.

Conclusion The original is still running my working day. Its replacement has learned that I work here too.

Additional note The next experiment is to let the patient read its own surgical notes and suggest a repair. For the time being, I retain custody of the scalpel.