Giant Robots Smashing Into Other Giant Robots

Retro-driven development

2026-04-28T00:00:00+00:00

Every session ends with a retro. This week, twenty-four commits out of about a hundred and forty started with that retro. Only a handful added anything new. I wasn’t building the system anymore. It was refactoring itself.

It is Week Four.

Tuesday: four commits before lunch

The 17th. Four refactor-from-retro commits before noon. Reusing API connections across commands instead of reconnecting each time. /morning filtering rules. Stale 1:1 prep dropped from the daily log. The system had been running for three weeks, and friction points had accumulated. I was working through them in fifteen-minute bursts between meetings.

By the end of the day, I had added eight more commits. An actionability check for /context. Top 7 priorities in the daily log. A self-management outcome for my Fusion goal. Retro after retro, feeding back into the commands.

Wednesday: picking sides

Wednesday ran hard. Nine refactor commits between meetings.

I read Sally Lait’s post on semantic calendar emoji and colours. I copied her system straight into /calendar:

🦚 Peacock (default): 1:1s, ad-hoc work
🫐 Blueberry: recurring group meetings
🌿 Sage: pairing, workshops, active work
🍌 Banana: internal socials, external community
✏️ Graphite: transit, food

By evening, I’d refactored /evening to use Ruby instead of Python. It was a small religious war. I picked the side my team knows. The CLAUDE.md gained a preference note. This sort of thing accumulates.

Thursday: the Anytime problem

The Anytime list from Things hit 75,000 characters. That’s around 19,000 tokens. It triggered context compaction mid-session. I noticed overdue items slipping through the cracks.

I needed a filter. Not a prompt. A real script. I wrote bin/filter-anytime.

KEEP_FIELDS = %w[Title UUID Tags Area Project Deadline Notes]

items.each do |raw|
  status = raw[/^Status:\s*(.+)/, 1]
  next if status =~ /completed|canceled/

  if tags.include?("Waiting") && deadline_str
    deadline = Date.parse(deadline_str) rescue nil
    next if deadline && deadline >= today
  end

  puts raw
end

Fifty-five lines of Ruby. It runs before the agent sees the list. The filter runs on rules. It doesn’t guess. Overdue items stopped slipping.

Every session, the agent regenerates the filter. Not this time. I wrote real code. Guessing has limits.

Monday: a stretch of quiet time

A quiet Monday morning. Six refactor commits in one sitting. /calendar, /inbox, /weekly, /context. A stretch of uninterrupted time before the week’s meetings started.

That is when I realised the system had shifted. I wasn’t grinding through tasks. I was editing the system that edits my day. Maintenance, not task-grinding. The point of building a system is to make it fade into the background.

Tuesday: the cap

By the 24th, I noticed something else. My Anytime list kept growing. Each session, I added new tasks from retrospectives, meetings, and the inbox. The filter was treating the symptom. The disease was that the input exceeded the throughput.

I added a commitment cap.

Commitment cap: No more than 20 active next actions in
Things at any time across all areas (work and personal).
If /morning surfaces items that would exceed the cap,
flag it and ask what to defer before proceeding.

/morning now blocks the Top 7 until I’ve deferred enough items to sit under 20. The check is mechanical. I can’t talk it into running anyway.

What I learned

The dominant mode this week wasn’t invention. It was refactoring. Twenty-four commits out of about a hundred and forty say “from retro” or “from feedback.” The system improves by use, not by planning.

Retro-driven development. It works because the signal is cheap and the fix is small. Notice a friction point. Name it. In the next session, the command that caused the friction receives a line of new guidance. No meetings. No sprints. No planning.

The commitment cap came from one of those retros. So did bin/filter-anytime. So did Sally Lait’s colour conventions find their way into /calendar. Each started as an irritation, ended as a line in a command file, and changed how the next session ran.

Try it

Retros don’t need to be long. End each session with one. In the next session, fix what rubbed you the wrong way. The system is yours.

Seven commands and the communication layer that emerged

2026-04-23T00:00:00+00:00

On Tuesday, 11 February, I made seventeen commits to my management system. That is more than any other day in the project so far. The previous two weeks had been about structure. Daily routines. Meeting sync. Project tracking. This week was about communication.

The trigger was simple. I ran /inbox and spotted the pattern. Every time: fetch the item, decide what to do, place it somewhere, move on. The first version of the command automated that loop:

# Inbox Command

Process the Things inbox one item at a time, newest first.

## Instructions

### Step 1: Load Context

Fetch in parallel:

1. Call `mcp__things__get_inbox` to get all inbox items
1. Call `mcp__things__get_projects` to get project names

Sort inbox items newest first (by creation date).

If the inbox is empty, report "Inbox is empty" and stop.

### Step 2: Present the Next Item

For each inbox item, present:

- Title
- Age
- Tags (if any)
- Notes (truncated if long)
- Related project (fuzzy-match title against project names)

Then wait for the user to say what they want to do.

Within thirty minutes, that command went through three revisions. The loop version advanced on its own. I changed it to single-item mode because I wanted control. Then I added reading detection: if the notes contain a URL, fetch the page title and suggest a tag. I created three commits, three lessons about how I process information.

The one-day command

The same morning, I created /reply for Slack DMs. It standardised the flow: find the user, open the DM, fetch the history, draft the reply, and send.

It lasted twenty-four hours.

By Wednesday, I had split it into /dm for direct messages and /thread for channel thread replies. Both shared a patterns file that held the common steps:

# DM Command

Send a direct message on Slack to $ARGUMENTS.

Follow shared patterns from
`.claude/commands/slack-patterns.md`.

## Instructions

### Step 1: Setup and Find User

1. **Rube Session Setup** (see `slack-patterns.md`)
1. **Find User** (see `slack-patterns.md`)

### Step 2: Open DM and Fetch History

1. Use `SLACK_OPEN_DM` with the user's ID
1. **Fetch History** on the DM channel
   (see `slack-patterns.md`)

The split happened because DMs and threads are different conversations. A DM is private, one-to-one, with full history. A thread is public, anchored to a specific message, with context that the whole channel can see. The same “reply” verb hid two different communication patterns.

The communication stack

That refactoring revealed something. Each command I built that week mapped to a communication channel:

/dm — Slack direct messages
/thread — Slack channel threads
/slack — new channel messages
/email — Gmail replies and composition
/hub — reading saved Hub pages
/draft — anything else (LinkedIn, talking points, Hub replies)

Six commands. Six ways I talk to people at work. The /draft command became the catch-all for channels without an API:

# Draft Command

Draft a reply or message for any context. Does not send.

Use this for LinkedIn messages, in-person talking points,
Hub replies, or any situation where `/dm`, `/thread`, and
`/email` don't apply.

## Pattern: Voice

Blend the user's natural tone with DHH and Nicholas Lezard:

- Direct and opinionated, but not abrasive
- Concise sentences that carry weight
- Avoid corporate filler
- Match the formality of the channel

The voice pattern is the part I did not expect to matter. I had assumed Claude would write in a generic assistant tone. Instead, by defining a voice, every draft came back in a register I recognised as mine. Not perfect. Close enough to edit rather than rewrite.

Plan mode and command boundaries

Not everything went well. On Tuesday, I hit a bug where plan mode leaked between commands. When I ran /waiting, it accumulated a state that bled into /retro. That broke both commands.

The fix took two commits and a revert. The first attempt added “Never use EnterPlanMode” to every command. That was wrong. The real fix was removing the auto-advance loop from /waiting. Each command invocation stayed self-contained. Commands are not functions. They are conversations. And conversations should end clean.

What I learned

Building these commands showed me something. My job as a director is communication:

Write messages
Respond to threads
Follow up on waiting items
Process my inbox
Draft replies
Read Hub posts

The actual management decisions happen in the gaps between those conversations.

The system I built in the first two weeks gave me structure: routines, meeting sync, and project files. This week gave me flow. The difference is that structure tells you what to do. Flow tells you how to do it without thinking about the mechanics.

I am not faster. I am less distracted. Each command removes one decision about where to go and what to type. That compounds over a day of fifty small conversations.

Try it yourself

Pick one communication pattern you repeat daily. Write a command for it. Not a script. A conversation. Define the steps, the voice, the context. Then run it and see what breaks. The breaking is where the learning is.