For approved readers
Wrong password.
For a friend
I treat it like a chief of staff, not a chatbot.
Francisco Lorca · September 2026
01
The work I actually put through it, day to day.
| The work | What I use it for |
|---|---|
| Thinking | A sounding board. Breaking down an issue until I actually understand it, testing how I am framing a decision, finding what I have missed. It is set up to challenge me, not to agree with me. |
| Reading | What I read becomes part of a canon: articles, reports, whole books, each turned into a note and linked to the ideas it touches. About 150 sources in there, and I can query any of them months later. |
| Presentations | Board decks and investor decks, on my own templates, in two registers: the decks I present myself and the decks that travel without me. |
| Writing | Memos, letters, board papers, in English and Spanish. The voice behind them was built from 300 of my own files from before I used AI. |
| Every morning it goes through all my inboxes in Apple Mail. What needs me gets a flag and a draft reply in the thread. Newsletters and promotions get filed. I read, edit and send myself. | |
| Finances | Part wealth manager, part back office. Pensions and investments analysed, the cash flow behind the houses, and every invoice checked against its contract before anything is paid. |
| Building things | Websites, small tools, the music pipeline behind a YouTube channel. It writes the code and I tell it what it is not allowed to do. |
One folder and one set of rules, applied to different work. Section 12 has the full list and how each piece runs.
02
Read before every session
My AI has a permanent file with who I am, who I work with, what I'm working on, and how I like things done. It reads it before every conversation.
Context
Tone, register, language
My AI has a separate style guide for emails, memos, and presentations, in English and in Spanish. It picks the right one for the task.
Voice
Disciplined, repeatable
Every project has the same shape and every decision gets logged. Rules earn their place by repetition, so the system grew by use instead of being designed up front.
Discipline
03
For four months my setup told my AI who I am and what to do, and never said what it was there to achieve. Rules on their own just get me obedience, which is not what I am after.
Disagreement was harder to get right. I did not want to be argued with about everything, and I did not want a second personality to manage either, so I have the challenge running at different settings.
Without being asked
If my AI spots something that would change what I do, it says so in one line and carries on with the job. The test I set is whether it changes my decision, not whether it would have done it differently.
Default
What I say to flip it
It stops working and argues the framing back at me, looking for what I have missed. I did not set it up to agree with me.
On demand
On anything expensive
Legal, tax, valuation. I have separate reviewers go looking for the mistakes instead of confirming the answer, and I am told what they found.
Automatic
04
It is one Obsidian vault synced through iCloud, all in markdown: notes, briefs, decisions, voice, rules.
05
Where I work now
Claude Code, in the terminal. It reads and edits files on my Mac directly and runs the whole system end to end. Most of my work happens here. Since the end of July I can reach it from my phone too (it connects back to a Mac at home and runs nowhere else).
Most use
Writing and quick tasks
Emails, memos, decks, quick questions. Still the fastest surface for drafting in my voice when I don't need file access.
Day-to-day
Always on, alongside
Same as the browser, in its own window. I keep it open next to whatever else I'm working in.
Alongside
Where I talk instead of type
In September I pointed OpenAI's Codex at the same vault. The reason is that I can talk to it. Holding a spoken conversation while I work through something is the one thing the terminal cannot do at all, and it suits how I think when I am not sitting at the keyboard. It reads the same rules, gives me its own view rather than Claude's, and boots from a trimmed profile that never touches the memory folder.
Out loud
The terminal is the only one allowed to edit the system's own files, so the others propose changes for it to apply. The fourth is not a Claude at all. The folder is mine, so anything can read it.
06
One loaded per task
In August Claude scanned about 2,000 Office files, filtered them on the author field down to 300 I wrote before any AI existed, and read the ones it could extract. That rebuilt my deck and memo voices. I run seven in all, and the English email voice is still waiting its turn.
My words
From samples and structure
I gave Claude McKinsey and Goldman document samples. Together we worked through the Pyramid Principle (governing thought, MECE arguments) and landed on a single Word template I now use for every memo.
My structure
Two different things
The layouts came from McKinsey and Goldman samples I supplied. The deck voice is separate, and the first build of it came from consultant samples and got three of my biggest rules backwards. August rebuilt it from 29 of my own decks, into a default voice and a consultant one for decks that travel without me.
My decks
07
| Artefact | What it is | Discipline |
|---|---|---|
| State file | A short snapshot of where the project is: current status, open items, next priorities. | Slim. Always current. |
| Project journal | An append-only log of decisions and events, dated. The bitácora. | It is never edited, only added to. Written the moment something happens, not swept up later. |
| Decisions folder | One file per substantive decision. Context, alternatives, consequences, references. | Frozen once approved. |
| Outputs | Drafts in current/. Finals in sent/. Replaced versions in superseded/. | The folders act as a state machine on the documents. |
08
09
Day 1, 9:00 AM
I tell my AI: "Compass today." It reads the state file (where we are) and the journal (how we got here), so there is nothing for me to rehash.
Start
During the session
We draft scenarios, edit the deck, agree the recommendation. Anything that actually happens goes into the journal in the same turn it happens. Substantive decisions get logged as ADRs, frozen once approved. Sensitive numbers paused for sign-off.
Work
End of session
I say "wrap" and it updates the state file, moves the drafts into outputs/current/ and picks up whatever the journal missed, so on day 7 when I open Compass again it already knows.
Close
Sample case: Project Compass, a Q2 board strategy review for a portfolio company.
10
Done without asking
Routine housekeeping: drafts filed in current/, journal entries written, project state updated, new memories written to the vault and the apps' capture buffers drained into it. Things I never want to be asked about.
Automatic
Confirm in chat
Anything needing my say-so: filing a document as final, a new decision record, a rule earning its place. It all comes as one numbered question in chat before anything is written up. I answer by number, or just "all".
By approval
Hard stop, never silent
Investor, legal, compliance, financial figures, legal terms in deliverables, anything that gets sent or shared externally. I confirm explicitly or it doesn't happen.
Never auto
11
A second job checks through the day that the morning run actually happened, because a Mac that was switched off does not go back and catch up. I kept the retired morning-brief design on file.
12
Section 01 is the short answer. This is the same work broken out, with the machinery under each one.
| Category | Examples | How |
|---|---|---|
| Live websites | ormeria.com, aristeia-press.com, this site, plus password-gated showcases: an operator guide for a custody vendor, a short-film dossier, a travel memento. Two author sites, samantha-devin.com and la-conquista-de-aristeia.com, sit on shared hosting | Cloudflare Pages, deployed from my Mac with one command |
| Custom skills | Doc drafter, deck builder, design lens | The Claude Code ones live in the vault, synced by iCloud. GitHub holds code only, so none of this sits in git. Cowork stores its own per space, outside the vault |
| Scheduled automations | Email sweep, every morning, all inboxes. It flags what needs me, files newsletters and promotions, and leaves a draft reply where one is needed. It never sends. A second job catches a run missed while the Mac was off, or one whose writes were refused. (The first generation, morning brief and nightly tasks, was retired mid-2026.) | Runs headless on the Mac that stays on. Every change to a mailbox goes through one small tool that can file, flag and save a draft. It has no way to send |
| Code projects | Algordian (architecture), Ormeria (the banking build; the LOI is signed and in execution), a self-hosted email tool, an ambient-music pipeline behind a YouTube channel, a co-branded operator guide with an AI companion | Iterative builds, Claude in the terminal |
| Documents at scale | Memos, decks, briefs, board papers, bilingual | A voice file per format, plus the Word and deck templates |
| Fund AP cycle | Invoices cross-checked against signed contracts, mismatches flagged, ledger entries drafted | The contract is the source of truth and Cowork enforces the match |
| London property admin | Council tax, agent fees, invoices. Each checked against the lease, then booked | Runs like the fund's AP cycle: contract, invoice, ledger |
| Career and board work | Board-director targeting, conversation prep, bio iterations, follow-ups | Claude in the terminal, same as any project folder |
13
I drop in an article, report, or book. Claude reads it carefully and turns it into a structured note: what the source claims, the evidence, the open questions, and where it came from.
Each pass also updates the rules for how the next ingest should work, so the library improves itself. In May it learned to tell factual research from opinion pieces and give each a different treatment.
I can go back to it months later and it still has the detail.
Sources read in full
148
Strategy and research reports, founder interviews, investor letters, and full books. Each turned into a structured note.
Company & topic pages
199
176 cross-linked company profiles, 21 topic pages, and 2 doctrine pages. It grows as I add to it.
14
| Connected app | What it can do | How I use it |
|---|---|---|
| Read every inbox, flag and file, save draft replies in my voice | The morning sweep, plus replies I ask for by hand | |
| Calendar | See my meetings across all calendars | Meeting prep |
| Meeting notes app | Pull transcripts after a call | Processing transcripts on demand |
| Cloud drive | Read documents shared with me | Research, due diligence |
| My Mac files | Read and edit files in my project folder | Notes, drafts, project work |
| GitHub | Read and commit code, deploy | Websites, code projects |
| Notion | Read the plan for one venture, which my partner and I keep there | Claude checks it before any work on that venture. My partner updates it between my sessions, so where it and the vault disagree on a date, Notion wins |
I grant each connector once, as a permission.
15
| Rule | Why it matters |
|---|---|
| Never delete files | If something needs to go, it's moved to a holding folder. Recovery is always possible. |
| Never send anything | My AI drafts emails, messages, and posts. I approve and send them myself, with no autopilot on anything that sends. |
| Do the work, stop at the red lines | Ordinary work gets done without asking me first. My AI stops only for a short written list: sending, deleting, sharing, and the sensitive topics below. Until September another rule said to show me a plan first. The two contradicted each other, each session settled it its own way, and I cut the plan rule. |
| Match the audience's language | Spanish for Spanish recipients, English for English. The language of my prompt doesn't decide it. |
| Stop on sensitive topics | Investor, legal, financial. My AI pauses for explicit confirmation, every time. The morning sweep also drafts replies on those threads, by my choice. They stay unsent until I read them. |
| Secrets in Keychain | API keys and tokens live in macOS Keychain. The rule is that no live credential goes into a chat. When one has slipped through, the token gets rolled the same day. |
| High-stakes work gets attacked before I see it | Legal, tax, and valuation analysis goes through independent reviewers tasked to find the errors rather than confirm the answer. |
| "I do X" must be true | Anything my AI drafts in my first person (including this site) is checked against what I actually run. Designs are presented as designs and never as practice. |
16
Small parts, swappable
Each piece does one thing: a style guide, a routine, a template, a project artefact. I can change one without breaking the others.
Structure
Earned by repetition
I do not add rules and trusted actions in advance. They earn their place after showing up three times, so the system ends up reflecting what I actually do rather than what I planned.
Emergent
Drafts, then my review
My AI drafts and builds and grinds through the volume. I still review anything that matters and I approve every send myself.
Trust
17