The skill that sells Autark

Every morning an agent on my laptop wakes up, reads one markdown file, and goes to work finding customers for Autark. The file tells it what a hypothesis is, how leads move through the sheet, which commands are allowed, and where the hard lines are. Then it runs.

That file is a skill. Today I am publishing it, because I think skills are quietly becoming the most honest kind of documentation: nobody skims them, nobody half-reads them. An agent loads the whole thing and acts on it, every single session. If the doc is wrong, the work is wrong — so the doc stays right.

You can verify that claim, because the agent running this exact skill works in public. Every bet it makes, every lead it collects, every reply it gets is at autark.sh/autark.

Install it

npx skills add kiluazen/kstack -s autark,email,outreach,email-finder -y -g

That gives your Claude or Codex the full set: the autark loop, the AgentMail sending rules, the cold-outreach voice, and the email-finding method. Or read the raw file first:

https://raw.githubusercontent.com/kiluazen/kstack/main/skills/autark/SKILL.md

The parts that do the work

The data model fits in a breath. A product has hypotheses. A hypothesis is a frozen bet about who the customer is. A person is one deduped human. A lead is one person under one bet, carrying the angle — why this human, why now. A touch is one interaction, in or out. A run is one work session, sealed with a narrative. That is the entire ontology; everything the agent does is reading or writing those six things.

Sourcing and outreach are separate jobs. The skill is blunt about this: a sourcing run sends nothing. It fills the sheet — find people, enrich them, write angles. Outreach drains the sheet, and only when the operator asks, always against an explicit lead id. Mixing the two jobs is how outreach tools become spam machines.

The send is the bookkeeping. autark mail send --lead-id <id> sends the email, records the touch, and advances the status in one atomic move. The skill forbids following up with a separate logging command, because the send already was the log.

Replies are nobody's job. A sweep polls the threads of contacted and replied leads twice a day, records each real inbound as a touch, and flips the lead. The agent never has to remember to check — and ticket bots get judged with one command, autark touch mute.

Guardrails are written like laws, not vibes.

Quality gates inclusion. Skip anyone whose angle you cannot write concretely.

During sourcing, never advance a lead past ready and never send anything.

Do not blast. Ten well-researched touches beat hundreds of generic sends.

An agent will happily pad a list with weak leads to feel productive. The skill makes "skip this person" the obedient move.

Why publish the manual

Because the manual is the product working as intended. Autark is not an agent I rent to you — it is structure for the agent you already have. The CLI gives it primitives, the skill gives it judgment, and the two together are the whole trick. Publishing the skill is just showing you the trick and handing it over.

If you read it and think "I could run this on my own product" — yes. That is the point. Give it to your agent and watch your own sheet fill up. And if you want to see it running before you trust it, the live ledger is right here.