An AI manager that guards my deep work

I built my own manager in Claude Code: a skill, one context file, and evidence from the tools I already use.

Problem

Small work arrives without limit: support, prospect emails, analytics requests. Together it fills the day and the deep work slips.

Protect two blocks of deep work a day, and know what actually moved each week.

My own manager in Claude Code: a skill with one context file that checks my real work across our tools.

Evidence it reads Git commits GitHub PRs Linear tickets Calendar Notion meeting notes

→

Claude Code The manager

Skill · how a manager works: sizing, check-ins, planning, 1:1s

Context · one ledger file: goals, committed outcomes, queues, who I'm waiting on

Tools · reads and writes the calendar, Linear and git

→

What I get 2-minute morning check-in Deep blocks on the calendar End-of-day review Weekly plan Fortnightly 1:1

Fig. 1 — Evidence in, the manager in the middle, what I get out.

Small · 30 min or lessBatched

Emails, replies, reviews. One window a day, and what doesn't fit waits.

Deep · 30 min to 3 hoursTwo a day

Specs, prototypes, scoping. Each one on the calendar and naming what it produces.

Outcome · over 3 hoursBroken down

Not a task yet. Split into sized tasks before it enters a queue.

Fig. 2 — Every task gets a size when it arrives.

a
Keep skill, context and tools apart. Instructions, memory and evidence each live in their own place.
b
Evidence, not self-report. It checks git, Linear and the calendar before it asks me anything.
c
Size work when it arrives. Unclear work is deep work, even when it looks small.