🗑️🔥 Substrate & Context — the flaming garbage truck

O Substrate is your domain’s raw knowledge—and you almost always start with what the author calls "a dumpster fire". Especially in a business, they're 500, 5,000 files of various types and structures: PDFs, spreadsheets, zips, even unusual files like Parquet. Chaos. The Substrate is the layer that turns this chaos into something the OS can use.

Why learn

Because most people try to connect AI on top of the chaos and gets frustrated with bad answers. Accepting the flaming dumpster truck as a starting point—and treating it methodically—is what separates an OS that knows your domain from one that guesses. In context-heavy domains (Tax OS ~34%), this layer is the most important of all.

Key concepts

Before asking Claude Code for anything, do a mini data engineering: organize. The author describes two simple cuts — first by file type (PDFs with PDFs, spreadsheets with spreadsheets), then by age (last 6 months vs. last 2 years). This isolates parts of the material and gives the OS clean ground to work with.

The cleanup flow in 3 cuts

Why learn

Because the quality of the OS is limited by the quality of the Substrate. Dumping 5.000 disorganized files into AI produces vague and expensive answers (lots of tokens). Organizing first is the low-cost step that multiplies everything that follows.

Key concepts

After organizing things comes the harvesting (harvest). The guiding question is: beyond what I already have, what's missing to create the “expert of my dreams” in this domain? You’ll gather knowledge from where it lives — YouTube, Instagram, experts, laws, lawyers — and bring it into Substrato.

How to read: the chaos (orange) is organized, then gathered from external sources, then a workhorse model distills everything. Two folders remain: raw/ (kept as a backup, gray) and synthesized/ (the nuggets, blue)—only this one enters the context.

Why learn

Because what you already have is rarely enough. A true specialist knows things that aren’t in your files—they’re in experts’ heads and public sources. Harvesting deliberately closes that gap instead of hoping the model “knows.”

Key concepts

The target of harvesting is the tacit knowledge: what a generic model no knows why it isn't well documented—it's in the know-how of the people doing the work. In Freedom OS, for example, these are details like which translation agencies are reliable, what needs to be notarized (officially signed) and which documents to translate into Arabic or French depending on the country.

Why learn

Because it’s exactly the tacit knowledge that makes the OS feel like a expert and not a chatbot. Without it, the OS repeats what any model would say. Capturing tacit knowledge is the difference between “generic” and “no one does it like my OS”—the test that comes back in the Trilha 4 audit.

Key concepts

Scattered collected material still isn't useful. The next step is to aggregate sources into a compendium (a compendium.md) — a distilled reference document — and in playbooks: playbooks for “how to do X in this domain.” It’s what the OS consults to respond with an expert’s depth.

Why learn

Because compendiums and playbooks are the format that makes knowledge usable. They’re also the bridge to the next track: a good playbook is almost a Skill waiting to be packaged.

Key concepts

The Substrate’s core pattern is separation raw vs synthesized. You keep a folder raw/ with the raw transcripts and documents (kept just in case) and a folder adjacent and complementary with the distilled versions — the nuggets. The golden rule: inject only the nuggets in context.

Why learn

Because it’s the pattern that repeats in every domain — pantry vs. finished dish. Confusing the two (injecting the raw material) blows up the context window, raises costs, and worsens the answers. Separating raw from synthesized content is what keeps the OS lean and the knowledge retrievable.

Key concepts

The person doing the distillation doesn’t need to use the most expensive model. The author uses a workhorse model ("workhorse" model)—cheap and efficient, like Gemini Flash — running in a loop for all files: "use this model, go through each file, and create a synthesis of the central nuggets from these transcripts".

Why learn

Because distilling by hand doesn’t scale, and using a top-of-the-line model for it is wasteful. The “cheap workhorse + loop” pair makes synthesis viable at scale—and when you generalize the script, you get a Substrate engine that runs every time new material arrives.

Key concepts

Let’s build the synthesis engine. The copy-run prompt below tells Claude Code to run a workhorse model in a loop about everything in raw/ and write one nugget per file in synthesized/. Replace only the sections between <brackets>.

Você é meu pipeline de Substrato. Use o modelo workhorse
<modelo barato: ex. gemini-flash> para destilar meu material bruto.

Passos:
1. Liste todos os arquivos em ./raw/ (transcrições, PDFs já em texto,
   posts). NÃO modifique nada em raw/ — é a minha despensa.
2. Para CADA arquivo, chame o modelo workhorse e gere um nugget:
   - no máximo <N: ex. 10> bullets com o essencial;
   - capture o conhecimento tácito (exceções, "o que dá errado",
     atalhos), não só o resumo óbvio;
   - foque no que responde à minha pergunta-objetivo:
     "<sua pergunta-objetivo do domínio>".
3. Salve cada nugget em ./synthesized/ com o MESMO nome do
   arquivo de origem + sufixo ".nugget.md".
4. Ao final, gere ./synthesized/compendium.md agregando os
   nuggets por tema, e me diga quantos arquivos processou.

Regra: só os arquivos de synthesized/ entram no contexto depois.
Nunca injete o conteúdo de raw/ diretamente.

Why learn

Because this is the operational heart of the Substrate layer. Running it once gives you the synthesized/; generalizing the prompt gives you the “knowledge harvester” skill that you reuse whenever new material arrives — the self-improving system with you in the loop.

Key concepts