Case study · personal ops · 2026

fredis: the assistant that runs my working day.

A private AI second brain: 24 skills, persistent memory, and a team of agents wired into Slack, Gmail and HubSpot that draft everything and send nothing. I hold the only send key, and I use it every day.

A live product of my own, built outside this concept. Not work for Fin, and not affiliated with them.

skills, loaded per task: 24
repos on one board: 9
advisors: researcher, analyst, PA, CEO, builder: 5
send key, held by me: 1

The context

Five products in a year, solo, means the operations of a small studio land on one person: research, notes, follow-ups, repo issues, client threads. fredis carries that load. It holds research, notes and decisions across sessions, so nothing resets between working days.

The bet

An assistant earns trust through what it is not allowed to do. Every agent in fredis drafts; none of them sends. Ownership of every outbound word, commit and deal stays with me.

The fredis board: 342 items across nine repos, with planned, in-progress, done and blocked counts per repo and each repo’s next item — The board: 342 items across nine repos, each with its status counts and next item. One person, a portfolio of products, one view of what is planned, in progress and blocked. Tap a capture to open it full size.

The fredis review queue: drafted tickets waiting in columns for review and send, with the note that sending anything external happens in your own mail or HubSpot, never from this page — The board: 342 items across nine repos, each with its status counts and next item. One person, a portfolio of products, one view of what is planned, in progress and blocked. Tap a capture to open it full size.

The design decisions

Agents draft everything, send nothing.

The PA, sales and marketing agents prepare replies, follow-ups and outreach inside Slack, Gmail and HubSpot, and stop there. The send is mine. This is the rule every client product in this portfolio inherited: the AI advises, a person decides.

Nothing runs without an approved PRD.

Builds fire through a gated dispatch path: an approved PRD first, one headless session at a time, and the work arrives as a reviewable draft rather than an applied change. A diff is a proposal; a deploy is a decision.

Advisors carry rungs.

Each advisor on the roster has a rung that sets how much it may do alone: per-action approval for the researcher, packet-writes only for the PA. The run history keeps every run with its cost, including the failed ones.

Skills, not one long prompt.

Twenty-four skills load by task, so the assistant works from written-down process. When a way of working improves, the skill is edited once and every later session inherits it.

Memory is the product.

Research, notes and decisions persist across sessions. An assistant that forgets yesterday is a search box; one that remembers is an advisor.

The test bench for everything else.

Every pattern the other four products ship was proven here first, on my own work, where a failure costs me an afternoon instead of costing a client.

Where it stands

Private, and in daily use across nine repos. What can be shown is the owner console above; the contents (my inbox, my clients, my notes) stay out of frame. Ask me for the walkthrough and I will run it live.

What I learned

The rules matter more than the model. Nearly everything on these screens is a limit: an approval gate, a rung, one session lane, a PRD before any build. Those limits are what let one person hand real work to a team of agents and still own every outcome.