Delegate Yourself First

The first person you should delegate to is a copy of yourself, and the reason is cost of transfer. A chief of staff takes six months to learn how you decide. Most of that time is not skill acquisition. It is them reverse-engineering rules you already know but never wrote down: what you say no to, what "good" looks like in a draft, which intro requests you take, what number makes a deal interesting. You are paying a salary for someone to decompile your head.

So compile it yourself first. Write the rules down in a form that executes.

Delegation is compilation

Every act of delegation is the same operation: take a decision procedure that lives in your head and externalize it far enough that someone else can run it without you in the loop. The traditional target for that compilation is a human. Humans are expensive compile targets. They forget, they drift, they leave, and each new one starts from source.

A Claude Code persona plugin is a cheaper target. Concretely, it is a set of skills, each one a SKILL.md file with YAML frontmatter (name, description) plus an optional references/ directory, loaded on demand when the description matches the task. The description is the dispatch table. The body is your decision procedure. The references are your case law: past emails, past deal memos, past posts, the raw material a chief of staff would have spent months reading.

The whole thing ships as a plugin behind a marketplace.json in a .claude-plugin/ directory, and anyone (including future you, on a fresh machine) installs it with two commands. That is the entire onboarding.

What to encode first

Not everything in your head is worth compiling. Encode in this order.

1. Your "no" rules. The highest-leverage decisions you make are rejections, and they are the most rule-shaped thing you own. "I don't invest outside LatAm B2B." "I don't take calls without an agenda." "I don't advise companies competing with portfolio." Write them as explicit conditions with thresholds, not vibes. A rule that says "pass on anything pre-revenue asking above $8M post" executes. A rule that says "be selective" does not. This is the first skill because a persona that says no correctly is already saving you hours, even if it can do nothing else.

2. Your rubrics. For every recurring artifact you review (a landing page, a cold email, a hiring scorecard, a pitch), you have a checklist you run mentally in fifteen seconds. Externalize it as the skill body: what you check, in what order, what fails instantly. Put two or three real annotated examples in references/. The examples matter more than the prose. A model pattern-matches against "here is a draft I rejected and why" far better than against adjectives.

3. Your voice, as constraints. Voice does not compile from a description like "direct and warm." It compiles from samples plus negative rules. Ten real pieces of your writing in references/, and an explicit banlist: phrases you would never use, structures you avoid, how you open, how you close. Negative constraints are underrated here. "Never opens with a pleasantry" shapes output more than any positive instruction.

4. Escalation boundaries. The most important instruction in the whole plugin is the one that defines where it stops. Anything involving money movement, firing, legal exposure, or a relationship you personally own should return "ask the real one" instead of an answer. Write this as an explicit skill instruction, not an afterthought. A delegate that knows its blast radius is trustworthy. One that improvises past it is a liability with your name on it.

Notice what is not on the list: your biography, your credentials, your company history. That is context, not procedure. Include it in references/ so answers are grounded, but it is the least valuable thing to encode, because it is the part a stranger could scrape anyway. The rules are the part only you have.

Interrogate before you trust

You would not hand a new chief of staff your inbox on day one. Same discipline here. Before letting anyone install your persona, run it through decisions where you already know your answer. Ask it the last five cold emails you actually received. Ask it to reject a deal you rejected and check that it rejects for your reason, not a plausible one. When it drifts, the fix is mechanical: sharpen the rule in SKILL.md, add the failing case to references/, redeploy. Every correction compounds, which is exactly what does not happen when you correct a human twice a year in a performance review.

One operational note if you build this by hand: the marketplace must be git-backed. A marketplace.json pointing at plugins over relative HTTP paths fails silently on install, and you will burn an hour on it. Serve the plugin from a git source (the git-subdir pattern) and the two-command install works on Claude Code CLI 2.1.x as advertised.

Then hire the human

None of this replaces the eventual chief of staff. It changes what you hire them for. Instead of six months absorbing your rules, they install you on day one, run the persona for the mechanical 70 percent, and spend their judgment on the 30 percent the escalation boundaries kick back. You stop paying senior-human prices for pattern matching you could have written down.

The uncomfortable part is that writing the rules down exposes how few of them you actually have. Most founders discover their "judgment" is eleven rules and a lot of mood. Good. Eleven explicit rules that execute at 3am in someone else's terminal beat a hundred implicit ones that only run when you are rested and in the room.

Install yourself first. Then decide what is left to delegate.

Install a person

installs.me turns your files, calendar and calls into a Claude Code plugin that thinks like you. Anyone installs it with two commands:

/plugin marketplace add https://installs.me/lautaro
/plugin install lautaro@lautaro-installs