Detailed content
🍕 The Tax OS pizza — ~34% context
Remember the pie chart metaphor from Track 1? Here it gets numbers. In the Tax & Bookkeeping OS, the biggest slice isn't “agents” or “skills” — it's context, at ~34%. Identity, which seemed like the natural starting point, becomes the smallest slice (8%). Knowing how to read this distribution is what keeps you from spending energy on the wrong layer.
🌱 New here?
Context is the knowledge the OS stores and injects into the conversation when needed—here, the laws and audit standards. PII (from the English Personally Identifiable Information) is sensitive personal information, such as a CPF/SIN or account number. Ledger is the ledger: the categorized list of all transactions.
How to read: each slice is one of the 6 layers; its size shows how much of your effort that domain calls for. In Tax OS, more than a third is context (in teal, the largest). Agents (in cyan) are a small slice—they come in only to judge and review.
✓ Where to invest here
- ✓Compendium of audit laws and standards (context).
- ✓Pipeline that converts PDFs and saves context window space.
- ✓Hooks for PII and account separation.
✗ Where NOT to waste time
- ✗A long, poetic identity (8%—cut to the bone).
- ✗A swarm of agents before you have the compendium.
- ✗Generic skills without tax context behind them.
Why learn
Because the pizza is your priority map. When you know context is worth 34%, you stop trying to “fix” the OS with more agents and dig where it really matters: tax knowledge. This reasoning—“which layer matters most here?”—comes up in every domain of the track.
Key concepts
📚 Context: audit pain points + compendium of laws
Tax OS context comes from reverse engineering the worst-case scenario. The author says he knew nothing about taxes and started by asking: "what were the worst audits ever conducted in my country?". Scan YouTube and legislation for these nightmares, then consolidate everything into a compendium, is what gives the OS a real understanding of your tax system.
🌱 New here?
One compendium (or playbook) is a distilled reference document that brings many sources together in one place. Harvesting is the act of gathering these sources (transcripts, laws, advice from lawyers) before distilling them.
🔎 What goes into the tax compendium
- Audit scenarios — the classification errors that trip people up most.
- Legislation by jurisdiction — the law in your province/state and country.
- Gray area vs. black and white — what’s safe and what’s risky.
- Can it run across the company? — when a personal expense becomes a business expense.
💡 The tax golden rule
Whenever there's doubt between the "gray area" and "black and white," lean toward black-and-white. The compendium exists precisely to steer you toward safe classification—so you don’t have to hope you won’t get audited.
Why learn
Because it’s the 34% slice. Without this context, the model will only give you Google’s generic answer—and with taxes, that’s dangerous. Reversing the worst case and distilling it into a compendium turns the vague fear of taxman into a concrete list of pitfalls to avoid.
Key concepts
📄 PDF-to-text pipeline: raw → converted → synthesized
Bank statements often come as PDFs — and raw PDFs are poison for the context window: they take up a lot of space and confuse the model. The solution is a three-folder pipeline: the PDF goes into raw/, becomes clean markdown in converted/ and, finally, is summarized in synthesized/. Only the summaries go into the context.
How to read: the data flows from left (raw, in teal) to right (synthesized, in cyan). It’s the same pattern raw → synthesized from Track 1, applied to taxes: you store everything in raw/, but only injects the nuggets from synthesized/.
Illustrative recreation — the pipeline folder
tax-os/substrate/ ├── raw/ # PDFs originais (nunca editados) │ └── extrato-2026-05.pdf ├── converted/ # markdown espelho, 1 por PDF │ └── extrato-2026-05.md └── synthesized/ # só os nuggets entram no contexto └── resumo-2026.md
Why learn
Because it’s what makes the monthly review inexpensive and reliable. Dumping the entire PDF into context exceeds the window and causes errors; converting and synthesizing it keeps the raw history safe in raw/ and let the model work with only the essentials. You’ll see this pipeline as copy-run in Topic 7.
Key concepts
🔁 Skill: monthly review (statement → ledger → automatic sending)
The domain’s most repeated task is the end-of-month review—and what gets repeated most is the best candidate to become a skill. This skill takes the already converted statement and categorizes each transaction into a ledger, prepares the submission to the accountant and triggers alerts for out-of-pattern spending.
The monthly review pipeline
Read the converted statement
The skill opens the markdown for converted/ from that month—not the raw PDF.
Categorize in the ledger
Each transaction gets a category and a personal/business flag, following the compendium.
Flag and prepare the submission
Unusual expenses become alerts; the summary is ready for the accountant, with the transactions you need to explain.
💡 Cadence tip
The author points out that, outside the end of the fiscal year, the Tax OS hardly needs any new context. Run the skill once a month and let the system sleep the rest of the time — there’s no need to keep it “on” all the time.
Why learn
Because it’s where the OS pays for itself. Turning hours of spreadsheet work into a one-command skill frees up your time and reduces human classification errors—the very errors that trigger audits. And, like every skill, it’s never "done": you refine it in the session post-mortem.
Key concepts
🚧 Personal≠Business, PII, Currency Rules & Hooks
Here money and sensitive data come into play—and that’s why the rule may not be enough. Remember the distinction from Track 1: rule is a strong suggestion (“do not enter” sign); hook is deterministic (locked door). In the Tax OS, the three critical guardrails are: never mix personal and business finances, block PII, and convert everything to one currency.
✓ Becomes a hook (deterministic)
- ✓Block pushes to GitHub if there’s PII (SIN, account, certificate).
- ✓Reject any write that mixes personal and business accounts.
- ✓Convert every value to a single currency before adding.
✗ Remains only a “soft” rule
- ✗"Try to favor conservative classification" (preference, not a hard stop).
- ✗"Write the summary in a clear tone" (style, not safety).
- ✗"Use bullets in the report" (format, reversible).
⚠️ The mistake to avoid
Leave PII protection as a "soft rule" in the CLAUDE.md. In a long context, the model may "forget" the suggestion and upload an account number to a repository. Sensitive data and money call for hook — something that fails closed, not a polite request.
Why learn
Because it’s the “deterministic where it hurts” principle in action. You decide exactly what’s nonnegotiable (PII, account separation, currency) and what can be a preference. Getting that boundary wrong is what turns convenience into a real risk of a leak or a tax error.
Key concepts
🤖 Agents: Tax Strategist + Skeptical Reviewer
Agents are only ~10% of the pie, but they have a precise role: judge. There are two. The tax strategist suggests the best classification within the law; the skeptical reviewer is the devil’s advocate who "roasts you before the accountant does"—it tries to find the flaw in your classification before a real auditor does.
Tax strategist
Read the compendium and propose the most advantageous classification that still fits the "black and white". Orchestrate the review skill.
Skeptical reviewer
Takes on the auditor’s voice: questions every questionable transaction and requires justification before the summary goes to the accountant.
💡 The review gate
The skeptical reviewer is a review gate: nothing reaches the counter without going through it. It’s the same pattern you’ll see in the Support OS and Consulting OS — an adversarial agent that protects the output.
Why learn
Because it shows agents in their proper place: few in number, at the end of the line, and there only to judge. Promoting an agent makes sense here because tax review is a routine you’d already do by hand—and the skeptical reviewer gives you a chance to make mistakes at home, for free, instead of making them in an audit.
Key concepts
⚙️ Practical: bank statement PDF → Markdown pipeline
Time to get hands-on. The prompt below is copy-run: paste it into Claude Code (with your Tax OS folder open), and it will build the pipeline raw → converted → synthesized of Topic 3. Replace the parts marked with <...> by your values.
🎯 Objective
Have Claude Code create the three folders and convert each PDF statement into clean markdown and a monthly summary — without mixing accounts or writing outside those folders.
Você é meu engenheiro de dados do Tax OS. Monte um pipeline que transforme extratos bancários em PDF em markdown limpo, em três pastas. Contexto: - Pasta dos PDFs crus: <./substrate/raw/> - Moeda de destino: <CAD> - Meu ano fiscal: de <janeiro> a <dezembro> Faça passo a passo, listando o plano e esperando meu "ok" antes de CADA escrita: 1. Crie a estrutura: substrate/raw/ substrate/converted/ substrate/synthesized/ 2. Para cada PDF em raw/, extraia o texto e salve um .md espelho em converted/ (mesmo nome do arquivo), com uma tabela: data | descrição | valor | moeda. 3. Converta todo valor para <CAD> e marque com [conv] quando houver conversão. 4. NUNCA misture conta pessoal e de negócio: separe em converted/pessoal/ e converted/negocio/ pela conta de origem. 5. Gere substrate/synthesized/resumo-<2026>.md com: total por categoria, os 3 maiores gastos e as transações que eu preciso explicar ao contador. 6. NÃO escreva nada fora dessas três pastas. Se encontrar PII (nº de conta, CPF/SIN), mascare antes de salvar e me avise.
✅ How to verify that it worked
- 1.converted/ has exactly one .md for each PDF of raw/.
- 2.No amount appears without a currency; converted amounts are marked with [conv].
- 3.The subfolders exist personal/ e negocio/ — nothing mixed together.
- 4.synthesized/resumo-2026.md shows the total by category and the list of explanations.
- 5.Nothing was saved outside the three folders, and no account number appeared in plain text.
💡 Next step
After the pipeline runs, ask Claude: "distill this conversation into a slash command /extrato that receives the month as an argument". É o reverse meta-prompting from Track 3, turning the step-by-step process into a reusable skill.
Why learn
Because theory without execution goes stale. Running this copy-run puts the 34% context slice into practice: you leave the module with the data pipeline up and running, ready for the monthly review skill—and with the habit of always defining objective, copyable block, and how to verify.
Key concepts
✅ Module summary
Next module:
5.2 — Sales OS 📈 (the pizza where tools and skills rise to the top)