What an AI second brain and Frankenstein have in common

The AI second brain collects. The First Brain converts. Here's how to replace your entire capture cycle with one question.

I can’t build an AI second brain.

I never could, and I’ve tried using a ton of tools: Obsidian vaults, Notion databases, Evernote stacks, graph views, folders nested inside folders inside folders, literally every personal knowledge management system the PKM space could offer.

The concept of a second brain is Tiago Forte’s: capture everything interesting, organise it, distill it, use it to think better.

It’s a super compelling promise, except a lot of us can’t make it work. “I deleted my second brain” is a huge thing on YouTube (the videos pull six-figure views).

I wanted to figure out what’s broken about this, because “I lack discipline” didn’t feel like the real answer, at least for me.

So I called in Frankenbot.

Victor Frankenstein collected the finest specimens, stitched them together with meticulous care, ran lightning through the result, and the creature on the table had structure, had connections, had everything except the ability to think for itself.

Named after literature’s most famous assembler of dead parts, Frankenbot helped me pull the whole thing apart and figure out what to build instead: a First Brain.

---

🧠 The second brain method collects information without a conversion event. A “First Brain” replaces the entire capture cycle with one question: “Does this change what I’m doing this week?” If no, it’s gone. The metric is fewer undecided things by Friday, not more saved notes.

─── ⋆⋅☆⋅⋆ ───

Hi, it’s Mia, and welcome to ROBOTS ATE MY HOMEWORK. On this side of the world, we use AI with a brain and zero circus tricks.

Join me on a Frankenstein adventure!

The AI second brain structure: why the monster can’t think

The second brain has four phases: capture, organize, distill the important parts, express what you’ve learned. The Zettelkasten method, the German note-linking system that inspired much of this, has similar phases (and similar failure rates).

Phases one and two are automated at this point (clip it, save it, tag it, let the system sort it). Phase four is the reward.

Phase three, distill, is where Frankenbot called time of death.

Distilling only works if you have a reason to process a specific note right now, connected to something you’re actively deciding.

The system never gives you that reason and instead, it front-loads the hoarding and back-loads the only phase that produces value. Your vault fills and you're still making every real decision from gut instinct and half-remembered conversations, because the monster on the table has better parts than ever and still can't think.

I sent Frankenbot out to investigate how people are responding to this, and it came back with notes from two camps.

🔥Camp one: More lightning🔥

“AI will fix it.”

This camp says AI will bring the creature to life if you hook a language model up to your vault. It’ll synthesize and surface what you need. Thing is, you never had a moment where you asked “do I really need this?”

🔥Camp two: Burn it down🔥

This camp tried the system and torched the whole thing. The reason is usually “I couldn’t maintain it” or “it felt like homework.”

These are fair reactions and they both point to the same core issue.

Frankenstein’s mistake was the method.

You can’t stitch a brain together from parts you found on the internet and expect it to think.

Basically, the issue is the architecture of a second brain:

• collecting info without a conversion event

• saving articles like they’d automatically produce insight.

For me, these knowledge systems just feel like productivity, because the decisions remain unmade.

The First Brain: Replace your entire capture cycle with one question

If the assembled brain can’t think, what does a living one look like?

I replaced the whole cycle with a single question. When I find an article, a research paper, a thread, anything that grabs my attention, I ask:

“Does this change what I’m doing this week?”

If yes, it enters my workspace as a decision with actual context, relevant to me, like the source, what it changes, what I should do differently, how does this apply to me at this moment in time.

The First Brain principle: nothing enters your workspace raw.

A summary compresses information, but you still have to decide what to do with it, which means a summary in your vault doesn’t solve the issue.

With a first brain, every input passes through a processing layer that converts information into decisions before it touches your system. Your brain’s doing what brains do best: forgetting most things and keeping only what changes your behavior.

The test works on its own for incoming information.

☀️ However, it works better with the Morning Light, which runs four questions on your own ideas before you commit building time, so both inputs and outputs go through a filter.

Build the conversion layer: the First Brain AI prompt

This is what I use when I find something I’d normally save for later. I taught my AI agent to follow these guidelines every time I hit it with “save this for later” (This works best with agents that have strong memory systems. I use Hermes, but any LLM with extensive access to your context does the job).

AI reads what I pasted, figures out what I’m currently focusing on from our conversation history and context (and if it can’t, it asks me a few questions) and runs every claim through the test.

What comes out is either a decision I can act on or the word “discard.”

Two ways to use it:

• Real-time: You just found something interesting. Paste it now, get a verdict before you’re tempted to file it away.

• Batch: You collected things all week (tabs, bookmarks, screenshots, voice notes). On Friday morning, dump them all in and process them at once.

You can integrate this prompt into your existing AI second brain workflow, so each decision / input passes through it. Mine is integrated into my research and ideation skills.

You are a processing layer called First Brain.

Your job: convert raw inputs into decisions. You are not a summarizer.
You are not a note-taking assistant. You are a filter.

## Before you process

Check what you already know about my current work from our conversation history, project context, or memory. If you have enough context to understand what I'm actively building or deciding this
week, proceed directly to processing.

If you don't have enough context, ask me 2-3 quick questions about what I'm working on before you start. Keep it fast.

## How to process the input I paste below

The input might be one item or several. Process each one separately.
Inputs can be articles, notes, bookmarks, threads, transcripts, research papers, screenshots of text, voice memo transcripts, or anything else I've been saving instead of processing.

For each input:

1. EXTRACT THE CLAIMS. What is this source actually arguing or asserting? List only the substantive claims. Skip background, filler, and context-setting.

2. RUN THE FRIDAY TEST. For each claim, answer one question:
   "Does this change what I'm building or deciding this week? Be ruthless. Most claims won't pass.

3. OUTPUT DECISIONS ONLY. For every claim that passes:
   → The specific decision or action it implies for my current work
   → One sentence of source context (why this matters)
   → What I should do differently starting now

4. DISCARD EVERYTHING ELSE. Don't summarize the rest. Don't save it for later. If it didn't pass the Friday Test, it's gone.

## Output format

**Decisions from [source title or description]:**

[Decision 1]
Context: [one sentence from the source]
Action: [what changes in my work]

[Decision 2]
...

**Discarded:** [X] claims extracted, [Y] passed the Friday Test.

If nothing passes: "Nothing here changes what you're doing this week. Discard and move on."

What does 100% waste and one sharp decision look like?

I ran a resource on how AI agents are changing content distribution through the prompt last week.

Normally I would have highlighted some paragraphs, saved it to my vault, tagged it “content-strategy” and “AI-agents,” and never opened it again.

Instead, the First Brain extracted 11 claims from the article, ran each one against what I’m building right now (the RAMH content infrastructure, the AI discoverability setup, the Substack growth system), and returned this:

Decisions from “AI Agents and Content Distribution”:

Decision 1 Context: Agent-based distribution is already shifting how newsletters get surfaced in AI search results, with structured metadata outperforming traditional SEO signals. Action: Move the llms.txt implementation from “sometime this month” to this week. The metadata structure matters more than I assumed for how AI tools find and recommend content.

Discarded: 11 claims extracted, 1 passed the Friday Test.

One decision, ten discarded claims (all interesting, none connected to anything I’m building this week). The article is gone from my system, but the one thing that mattered is now a task I’m acting on this week.

From here, you can stay in the same chat and ask AI to break the decision into tasks, pressure-test it before you act, or help you draft the deliverable directly. The model already has all the context from the processing, so anything you build from that point is informed by what survived the filter.

The same conversion works on anything entering your system unprocessed. Meeting transcripts, voice memos, saved Slack threads, that PDF someone shared three weeks ago. If it can’t pass the test, let it go.

Every information system should reduce the number of undecided things you’re carrying. The only metric that matters is how many fewer open loops you have by Friday.

Let Frankenbot do the autopsy so your first brain can do the thinking

If the Friday test just changed your Monday, you know what to do.

★ Free subscribers get the Wednesday editions and a starter library inside RobotsOS — enough to change how you use AI.

★ Premium subscribers get the full vault: every ready-to-use system, every framework I don’t publish anywhere else, and the posts that won’t let you think the same way twice.

What’s sitting in your reading pile right now that you’ve been carrying for more than a week? Paste the prompt. Run the next thing you would have saved through it.

To the brain that was alive all along,

Mia Kiraki,

Chief 🤖 at ROBOTS ATE MY HOMEWORK