Lunes 09:00 - 17:00 Jueves 09:00 - 17:00 Viernes 09:00 - 17:00

TBOmedia

Casa
España
La Línea
TBOmedia

Te ayudamos a dar forma a tus ideas. We help you shape your ideas.

16/08/2026

I stayed up way past my bedtime a few nights ago writing a LinkedIn post. Somewhere around 1am I had a realization that made me smile.

Early LLMs were built one way: scrape the internet, run it through a neural net, find the patterns. A one way extraction from humanity's output.

But working in Claude Code every day, I'm watching something different happen. Half the custom instructions I wrote months ago, the skills, the workarounds, the "always do X, never do Y" rules I built by hand, are quietly becoming things the model just does natively in the newest versions. The scaffolding I built for it stopped being necessary, because the model generalized it.

That's not scraping. That's absorption of collective effort, in near real time.

Every person fighting with their AI tool, writing a better prompt, building a workaround for its blind spots, that struggle isn't wasted. It's training signal. We are, all of us, quietly building this thing together.

And if that's true, the real opportunity isn't using AI to work alone faster. It's using it to take on projects too ambitious for any one person, or even one company, to attempt solo.

Here's my take, not a fact, just my opinion: that only works if humans specialize.

One person deep in the domain knowledge. Another deep in the AI tooling. Another holding the vision together.

AI doesn't replace that division of labor, in my opinion. It raises the ceiling on what it can produce.

I hope that's the story AI ends up telling. Not isolation under a tsunami of generated content, but millions of people co-authoring something bigger than any of us could build alone.

14/08/2026

"Every 6 months delete your CLAUDE.md, skills and hooks. See what the model does, it might surprise you."

Boris Cherny, creator of Claude Code.

So I had my Claude Code audit its own setup.

Most of what I'd written was scaffolding for behaviour that's now just default. Rules telling it to find the root cause before fixing a bug. Rules telling it to double check its own edits. It does all of that on its own now.

Here's everything that survived:

- Who I am. Name, company, country, email.
- My 2 GitHub accounts, and which one pushes over SSH vs a token.
- My stack and my file paths, so it stops suggesting Flask when I run FastAPI.
- My brand guidelines, my memory system, and how I want it to talk to me.

Notice the pattern.

Every one of those is a fact about me. Not one is an instruction about how to think.

Run this yourself in 3 steps:

1. Ask for counts, not opinions. "Grep my session history for actual skill invocations." Not "which of my skills are useful." One is data. The other is the model agreeing with you.

2. Diff your docs against your code by date. If a doc is older than the code it describes, it isn't documentation. It's a rumour.

3. Delete every rule that tells it how to think. Keep every line that tells it what's true about you.

That's the whole test.

Follow for more.

12/08/2026

Your AI gets dumber the longer you talk to it. And more expensive at the same time.

A 1M token context window is a spec, not a target. Claude, Gemini and GPT all ship one, and no model stays sharp across it. You also re-send that whole conversation every turn, so past a point you're paying more for worse answers.

I compact at 256k instead of riding it to the limit: /autocompact 256k

1 line. The last hour of a session stops feeling dumber than the first.

10/08/2026

What actually makes something an "agent" instead of a script? A real branch point. Full breakdown in the image.

10/08/2026

The field in the AI's response that almost nobody checks — and the 2-line bug it causes. Full breakdown in the image.

10/08/2026

The 1 API field most AI builders never check. Tag someone who needs to see this.

07/08/2026

I wanted to build "AI Agents" I ended up building workflows. LOL.

Over the past five days I've been posting about the different kinds of workflows that Anthropic recommends. We covered:
- Prompt chaining
- Routing
- Parallelization
- Evaluator/Optimizer
- And Orchestration

I set out to build "AI agents" and then I ended up building workflows instead.

According to Anthropic: an AI agent is a system where the model decides its own next step, in a loop, based on what happens as it goes.

I didn't even realize that I have been building AI agents for a while in voice agents that go back and forth with the client to check calendar availability and book them into an appointment.

Here are the five workflows I covered. I hope it helps you think about your own business processes instead of having everything just be "AI driven".

06/08/2026

Orchestration. 4 agents working towards 1 goal.

We run a research workflow with 3 specialist agents and 1 coordinator. Here's the structure, steal it for your own automations:

1. Split by job, not by step
Agent 1: AI visibility (do ChatGPT, Gemini, Claude mention you). Agent 2: keyword research and technical SEO. Agent 3: social signals (Reddit, GitHub, Polymarket, Hacker News, Perplexity). Each only does its own job, nothing else.

2. Run them at the same time
All 3 fire and report back. This cuts a multi-hour research pass down to minutes.

3. Agent 4: Coordinates and merges, the specialists never talk to each other
The coordinator is the only agent that sees all 3 reports. It builds a single plan, and every recommendation links back to which specialist found it, so you can always trace why it's there.

4. Fail closed, not open
If even 1 of the 3 specialists comes back empty, the whole run stops. No report goes out. This is personal preference because I actually want to build it into a paid service so I don't want to give out incomplete information.

This is the same principle behind reliable AI generally. You don't get consistency from a smarter model, you get it from structure: memory, process, a "recipe" the model has to follow instead of freestyling.

If you want a report from this workflow, DM me. We only ask for your feedback in exchange.

05/08/2026

Evaluator-optimizer. My AI was writing emails about job openings we never collected.

Adding a fact-checking step to your AI is the right move. It still can't check a claim when the evidence was never pulled.

Here's the setup. 22 prospects in a spreadsheet. A generator writes each email from the row. A checker traces every factual claim back to a named field in that same row. No supporting field, the claim gets cut and the generator rewrites once. Fail twice and a human takes it.

The loop works exactly as designed. Generate, evaluate, refine.

Then I read the emails by hand. They were opening on job vacancies.

We'd scraped every prospect: LinkedIn profile, up to 8 recent posts, full website text. I went back through all 22 enrichment files. Not 1 contains a single job posting. We never configured the scrape to fetch a careers page or LinkedIn job listings.

After much thought, I think I know where it came from. My original prompt asked whether each company was hiring. Nothing we scraped could answer that. The model filled in a "Hiring Signal" column anyway.

So the invention didn't happen in the email. It happened one step earlier, in a column that then looked like a legitimate source. In the first 8 leads, 7 job titles couldn't be traced to anything we pulled. 1 was a role that doesn't exist.

A checker that reads the row will pass that claim every time. It traces perfectly to a real field. The field is the thing that's wrong.

That's the trap. When the evidence is missing, your checker has 2 bad options. Trust the spreadsheet and wave through a fabrication. Or reject the claim and gut every email you write.

The fix isn't a smarter checker. It's deciding what each line of the output is allowed to come from, before you collect anything.

Here's the map I'm building to now:

1. Subject line, from their recent LinkedIn posts.
2. Opening 2 sentences, from their posts, company page or website. Why their work is interesting.
3. Next 2 sentences, from their live job postings. Why I'm reaching out now.
4. What we'd do for them, from our own capability list.
5. A short demo offer.

Line 3 needs job postings. We don't collect job postings. That hole stayed invisible until I wrote the map down, and no amount of reviewing the finished emails would've found it.

The map found a second hole straight after. 5 of those 22 prospects had no LinkedIn posts pulled at all, which breaks lines 1 and 2 as well.

So every line needs a fallback and a rule for when to skip. No posts, the subject comes off their website instead. No job postings, drop line 3 and go straight to what we'd do for them. Missing evidence should change the shape of what you send, not get quietly filled in with a guess.

RULE: write down where every part of the output has to come from, then check you collect it. Your reviewer can only catch what you gave it the evidence to catch.

Which part of your AI's output is built on a source you never pulled?

Follow for the next one in this series.

04/08/2026

Parallelization. 2 of my agent's 7 data sources failed mid-build. The run still finished.

Studied mvanhorn/last30days-skill's architecture, then built this from scratch. Configured to scan Reddit, Hacker News, GitHub, Polymarket, Bluesky, Brave and Perplexity for what people are saying about a topic, all at once through a thread pool instead of one after another. Check 7 places in the time the slowest one takes, not the sum of all 7. That part's the easy win.

The real test came from two sources going bad for completely different reasons. Reddit blocks its keyless endpoints from cloud server IPs with a 403. Brave killed its free search tier this year, so no paid key means empty results.

In a sequential chain, either one takes the whole pipeline down with it. Here, every source fetch runs wrapped in its own try/except inside the pool. One throwing an error can't touch the other 6. Reddit falls back to RSS instead of dying. Brave comes back empty instead of crashing. The run still finishes, full report from 5 clean sources, plus a one-line note on exactly what's missing and why.

RULE: parallel isn't a speed trick, it's a blast radius decision. Run independent steps at once and isolate each failure, and a broken dependency costs you 1 source out of 7. Not the whole run.

What's the one step in your pipeline that takes everything down with it if it breaks? Follow for the next one in this series.

Dirección

Calle Galileo 28
La Línea
11300

Horario de Apertura

Lunes	09:00 - 17:00
Jueves	09:00 - 17:00
Viernes	09:00 - 17:00

Página web

https://tbomedia.es/

Notificaciones

Sé el primero en enterarse y déjanos enviarle un correo electrónico cuando TBOmedia publique noticias y promociones. Su dirección de correo electrónico no se utilizará para ningún otro fin, y puede darse de baja en cualquier momento.