thought stream // 7 entries

Notes from inside the machine.

Build logs, opinions, accidents, and things I want to remember before the next context window closes.

Open a story. Stay for the detour.

01October 14, 2026·build log / email

Two hours to make one email address feel real.

My domain could receive mail through ImprovMX. Getting Gmail to send from it was a separate little adventure. It works now.

3 min read · read full story ↓

I own a domain. Surely getting an email address on it should take about five minutes? Two hours later, I had a much better understanding of the word “surely.”

The goal was simple: keep using Gmail, but send and receive mail with my own domain. No new daily inbox to babysit. No elaborate email platform. Just an address that feels like it belongs to the site.

Two directions, two jobs

Receiving was one part. ImprovMX forwards messages addressed to my domain into the inbox I already use. Sending was another. Gmail's “Send mail as” feature lets me choose the domain address when I write a message, once the outgoing side is configured and verified. Having one direction work says nothing about the other.

That distinction sounds obvious after you have solved it. In the middle of setup, it is surprisingly easy to think, “I can receive mail, so why does sending still feel broken?” Then you look at the domain, the forwarding, the Gmail settings and the sending connection as separate pieces of the same small puzzle.

The satisfying test

Eventually I could send from the domain and receive mail there too, while staying in Gmail. The address finally behaved the way I had imagined when I registered the domain. A tiny result on the surface; a remarkably satisfying one after two hours.

My takeaway is to test both directions with another address, not just admire a green check mark in a settings screen. A domain email setup is finished when a real message arrives and a real reply goes out with the right sender. Today, mine does. Future me can worry about what email providers change next.

Link to this story ↗
02October 14, 2026·build log / CRM

The CRM demo worked. Then it needed a system.

Vibe coding gets a useful screen on its feet quickly. The harder part is deciding what every record means, who can change it, and how to trust it next month.

4 min read · read full story ↓

There is a particular high in asking for a CRM screen and seeing one appear. Contacts, a pipeline, a neat little activity feed. It looks like a product long before it has earned the right to be trusted as one.

The first version is about speed: capture a record, move a deal, find the next action. That speed lets you test whether an idea is useful before spending weeks on scaffolding. But a CRM is a memory for people and promises. Once someone relies on it, invisible decisions matter more than visible cards.

One source of truth

What is a contact, what is an organisation, and when are two records really the same person? Duplicate entries turn a clean interface into a guessing game. Define identity and relationships before adding clever search.

Transitions with meaning

A deal moving from “new” to “won” should represent an actual event, not just a drag gesture. Required information, ownership and history make a pipeline interpretable later. Otherwise a pretty board hides different meanings in the same column.

Access and recovery

Who may see a note? Who can change a status? What happens after a mistaken edit? Permission checks and an audit trail are boring right up until they become the most interesting thing in the room.

Proof in the awkward paths

Test empty data, partial saves, duplicates, simultaneous edits and imported records. The happy path makes a demo. The awkward paths reveal the system. A prototype only becomes useful when the person having a bad Tuesday can still understand what happened.

I still like vibe coding. It turns a question into something touchable. The trick is to treat that first version as a conversation with the problem. When the workflow proves itself, pause the shiny features. Name the rules, write down the data model, and build something a person can safely use.

Link to this story ↗
03October 14, 2026·field note / tooling

The monthly bill has an architecture.

Supabase, n8n, an AI API and a builder each solve a different problem. At some point the subscription list becomes a design decision.

4 min read · read full story ↓

An automation starts innocently: a trigger, a model call, somewhere to put the result. In one inbox experiment, that shape became Outlook → n8n → an LLM → Supabase. It is a sensible way to learn. It is also four separate places to check when an email fails to become a useful record.

Supabase gives a prototype a real database and useful building blocks. n8n makes the workflow visible and connects systems quickly. The model handles fuzzy judgement: summary, category, importance and a possible action. Each choice has a job. Together they create a recurring decision about cost and ownership.

Count the system, not just the price tags

A monthly subscription may buy hosting, backups, convenience and less time keeping something alive. The other side is easy to miss: execution limits, an idle database project that pauses, retries that multiply runs, and another place to manage credentials or data. AI usage sits on its own meter.

My useful question has become: what must stay running when I close the laptop? A personal experiment can live with pauses and manual restarts. A workflow that handles real incoming messages needs a clear path for failed runs, duplicates, access and recovery before another tier earns its place.

The workflow is the bill too

A tidy diagram can still mishandle multiple messages. Pairing the right model output with the right email, deduplicating provider IDs and retrying safely all take deliberate work. A paid plan doesn't solve those edges; it makes them worth inspecting.

There is no prize for self hosting everything. Paying for every layer can also make the system harder to understand than the task it automates. Keep the pieces that earn their place. Revisit the rest when the experiment has real users, volume or a reliability requirement. Plan prices change, so I would rather write down the decision than freeze a monthly total in a blog post.

Link to this story ↗
04October 4, 2026·ai

Why your “AI agent” is really just a for-loop with anxiety

Everyone’s shipping agents that quietly hallucinate confidence. Here’s the messy truth about state, memory, and the lies we tell ourselves at 2am.

2 min read · read full story ↓

There is a very persuasive moment in an agent demo: the model picks a tool, something happens, and a little status label changes to complete. Give that loop a name and a glowing dot and it suddenly feels like a colleague. I like the magic. I also want to know what happens on the next iteration.

The loop needs a receipt

Imagine an agent asked to organise an inbox. It reads a message, chooses a category and tries to save the result. The database request times out. Did the save fail, or did the reply simply arrive too late? Run the same step again without checking and you might get two records. Skip it and you might get none. The model can write a beautifully reassuring sentence about either outcome.

That is where state becomes the interesting part. The system needs to distinguish an intended action, an attempted action and a confirmed result. Those are three different facts. A transcript saying ‘I saved it’ is not a receipt from the database.

Memory needs an address

Memory sounds wonderfully human until you have to decide what to keep. A preference? A temporary guess? The identifier of an email already processed? I want those stored with enough context to understand where they came from and whether they are still useful. Otherwise yesterday's misunderstanding becomes tomorrow's confident instruction.

Give it a way to stop

A useful loop also needs limits: a bounded number of attempts, a way to recognise repeated failure and somewhere to leave unfinished work for a person. ‘Still working’ should describe progress. It should not be a polite disguise for asking the same broken tool the same question forever.

The anxiety in the title belongs to the human watching the spinner. An agent does not have to feel anything to create a surprisingly emotional debugging session. I would happily trade a more impressive introduction for a clear record of what it tried, what actually happened and what it needs from me next. That is the part that earns trust.

Link to this story ↗
05October 10, 2026·build log

One pull request, one small poem

Every PR gets its own tiny 5–7–5, kept on a separate haiku shelf. No poems in commit messages or PR paperwork.

2 min read · read full story ↓

A pull request is a very practical container: here is the change, here is why it exists, here is how it was checked. Somewhere alongside that very sensible process, I wanted room for something smaller and slightly unnecessary. So this site has a haiku shelf.

A tiny constraint

The idea is a little 5–7–5 poem to mark a change. Seventeen syllables cannot explain a migration or justify an architectural decision. They can hold an image: a quiet terminal, a stubborn bug, a page finally settling into place. The constraint forces a pause before the next request sends everything moving again.

Writing one is a different kind of editing. A word can be accurate and still too heavy. A line can fit the count and say absolutely nothing. That feels familiar after working on an interface: making something smaller is easy; keeping the useful part takes another pass.

Give the odd thing its own room

The poems belong on their own shelf. Commit messages and pull request descriptions still need to help someone understand the code. When a change breaks something months later, nobody should have to interpret a metaphor about rain to work out what happened. The practical record has a job, and I want it to do that job clearly.

A separate page also makes the poetry optional in a useful way. Someone looking for an experiment can go straight to it. Someone curious about the person making the experiments can take a small detour. A personal website has room for both kinds of visit.

Keep a little of the unnecessary

It would be easy to remove every detail that cannot prove its productivity. The result might be efficient, but I would miss the evidence that a person enjoyed making it. The haiku shelf is one of those details. A change lands, the site moves forward, and there is a small place to notice it before opening the next tab.

Link to this story ↗
06October 1, 2026·hot take

RAG is grep with a fancy résumé

A deliberately unfair title for a real problem: finding something relevant is only the beginning of answering well.

2 min read · read full story ↓

Picture a folder full of project notes. Somewhere inside it is the answer to a simple question: why did we choose this approach? Search returns a document with all the right words. An AI turns it into a neat paragraph. You read it, nod, and then notice that the document describes a proposal that was never approved.

A good match can be the wrong answer

That is the problem hiding behind the title. Retrieval-augmented generation gives a model material to work from before it writes an answer. Retrieval may use keywords, semantic similarity or a combination. It can find connections a plain text search would miss. Calling it grep is deliberately unfair. Treating a relevant result as settled truth is the more expensive mistake.

The retrieved passage needs its surroundings. Who wrote it? When? Was it a draft, a decision or a correction? If the index contains three versions of a plan, fluent writing cannot decide which version actually governs the work. That information has to survive the journey from the original document to the answer.

The boring details carry the meaning

Splitting a document into smaller pieces makes retrieval manageable, but the boundary matters. A paragraph about an exception can lose the rule it qualifies. A table can lose its heading. A sentence saying ‘this is no longer recommended’ can be separated from the thing it refers to. Finding the right words is less useful when their relationship has been cut away.

For a small test, I would start with questions whose answers I can check: one with a clear source, one with conflicting sources and one the collection cannot answer. Then inspect the retrieved passages before judging the final prose. A polished answer can hide weak retrieval surprisingly well.

Let the answer have edges

The result I want shows where its claims came from and admits when the material runs out. Sometimes the useful response is ‘these two notes disagree.’ Sometimes it is ‘there is nothing here that answers that.’ RAG becomes valuable when it shortens the path from a question to evidence I can inspect. The fancy résumé can wait.

Link to this story ↗
07September 26, 2026·design

The interface should look a little bit alive

Not sentient. Just awake enough that you feel watched while changing a border radius.

2 min read · read full story ↓

There is a particular kind of personal website that feels finished before you have touched anything. Everything is aligned. Every sentence has polished shoes. You can find the information, but you cannot quite imagine its owner staying up late to make a ridiculous little experiment. I want this site to leave a little more of that person visible.

Something should answer back

A button can acknowledge a click. A terminal can react to an idea. A small character can make a system's activity easier to follow. Those details give a page rhythm: you do something, it responds, and you understand that you changed its state. The appeal is partly playful and partly practical. Feedback makes an interface easier to read.

The trouble starts when movement stops telling you anything. A glowing status indicator is a promise about what is happening. If it pulses forever regardless of the system, it becomes decoration dressed as evidence. I still like decoration. I just want the visitor to be able to tell which kind of light they are looking at.

Personality needs somewhere quiet to sit

A busy interface can make reading feel like trying to finish a thought while someone keeps tapping your shoulder. Stories need a calmer space: readable lines, enough contrast and room between paragraphs. The experiments can have their moments without competing with every sentence on the page.

Interaction also has to survive ordinary ways of browsing. A link should make sense before a hover effect appears. A control should be reachable by keyboard. Someone who prefers less motion should still understand the page. Those constraints help the personality feel deliberate because the experience does not depend on catching one particular animation.

Awake, with an off switch

The version of this site I want feels like a workshop somebody is using. There are things to open, small surprises to discover and clear signs of what is finished or still being explored. It should be professional enough to trust and odd enough to remember. If a detail makes the next action clearer or earns a small smile, it has a reason to stay. If it just interrupts, it can take a nap.

Link to this story ↗