Ari’s Space*   est. 2026
Skip to content
ari*s space[ a. ]
← Libraryworkflows / 2026-05-10

Your Claude experience has two surfaces. Customise both.

Shape the replies for the way you read, then make session state visible in a local statusline. This is the May 2026 setup behind my voice rules and six-field display.

5 parts / Hover to explore. Click to read.

Shape the replies for the way you read, then make session state visible in a local statusline. This is the May 2026 setup behind my voice rules and six-field display.

VOICE SURFACE
Reply rules -> readable responses

STATE SURFACE
Session JSON + local state
             |
        local script
             |
      visible statusline
Reply rules shape the language, while a local statusline exposes session state. These are complementary surfaces; the essay's examples are not a supplied installer or complete script.
47% · opus-4.7 · WORKSPACE · main* · 2🛠 · 02:41
The original six-field display: local session state, colour-coded for a quick read.

I need two things from an AI workspace: language I can skim and a clear view of what is happening. Changing the prompt alone doesn't settle either for long.

I'm dyslexic. A wall of text can contain the answer and still be invisible to me. The interface has a similar problem: if I have to ask for the session state every few minutes, it isn't doing enough of the work.

This is the setup I described in May 2026. One surface shapes the replies. The other keeps the useful state in view.

Surface 1

Voice

~/.claude/CLAUDE.md

~/.codex/AGENTS.md

How Claude talks to you. Markers, header tiers, mono signals, paragraph length, where decisions live.

Surface 2

Interface

~/.claude/scripts/statusline.py

~/.claude/settings.json

How you read what's happening. Persistent, glanceable, colour-coded by zone.

Two surfaces. Both configurable. Both compound.

Make the voice readable for you

I put my response rules in ~/.claude/CLAUDE.md. They tell the agent how to make a reply usable for my brain:

  1. End with a bold marker such as → DECISION, → QUESTION or → NO DECISION NEEDED. I can skim for what is required of me in under three seconds.
  2. Use heading size to show importance. One main headline, then sections and small labels at the scale they deserve.
  3. Reserve monospace for paths, values and commands. The font change is a signal, so don't spend it on emphasis.
  4. Put a horizontal rule at a major pivot, such as the move from status to a decision.
  5. Bold the words that carry the point, so my eyes have anchors.
  6. Use absolute paths that I can open from the terminal.
#THE headline of the reply (max 1)
##Main sections (status, decision, design, question)
###Sub-sections inside one of those
####Inline labels and micro-groupings
Header tier = importance. The size matches how load-bearing the section is to read.

I mirrored the same rules in ~/.codex/AGENTS.md for Codex. The aim was one readable voice across two harnesses, with each harness finding the instructions where it expects them.

These are my choices. Yours should serve the way you read, rather than copying a format that happens to work for me.

Put session state where you can see it

The statusline example has six fields: context percentage, model, workspace, branch and dirty flag, dispatch-worker count, and elapsed session time.

 47% · opus-4.7 · WORKSPACE · main* · 2🛠 · 02:41

The opus-4.7 label belongs to the original illustration. It is not a recommendation for a model to choose now.

A local script reads session JSON on standard input and prints the line. Claude Code's statusline documentation confirms that this local display consumes no API tokens. Updates are event-driven, including assistant messages and other session changes, with an optional refresh interval. They aren't limited to exactly one update per turn.

Your script can also read local state for fields such as dispatch-worker count. Those additions need their own data source; a drawing of the field doesn't supply one.

47% · opus-4.7 · WORKSPACE · main* · 2🛠 · 02:41
47%  context: green <50, yellow 50 to 75, orange 75 to 90, red over 90
opus-4.7  model: opus magenta, sonnet cyan, haiku grey
WORKSPACE  per-prefix brand colour
main*  white clean, yellow dirty, red on conflict
2🛠  dispatch worker count: yellow when 4 or more
02:41  session clock, always dim grey
Six fields, six colour rules. State encoded in colour, not just decoration.

The original colour design gives context four zones: green below 50%, yellow from 50 to 75, orange from 75 to 90 and red above 90. The original prose and figure differed at the exact 90% boundary; choose an explicit comparison in your own script.

The model label uses magenta for Opus, cyan for Sonnet and grey for Haiku. A workspace has its own colour. Branch state is white when clean, yellow when dirty and red for a conflict. Worker count turns yellow at four or more. The clock stays dim grey.

The 50% point has a purpose in my workflow: write a session handoff at half the context window. Seeing that threshold gives me a cue to save continuity before I need it.

Build the version you will actually use

Start with your own reply rules in ~/.claude/CLAUDE.md. Mine took an afternoon to draft and three weeks of use to settle. Notice which replies are hard to read, then change the rule that would help.

If you also use Codex, maintain the corresponding instructions in ~/.codex/AGENTS.md.

For the interface, my layout put the script at ~/.claude/scripts/statusline.py. Configure it in ~/.claude/settings.json with statusLine.type set to command and statusLine.command pointing to the script. Read the session JSON, handle absent values, print the fields you need and exit.

The original article planned a sanitised script, workspace example, instruction template and installer. Those downloads are not supplied here. The examples show the arrangement to build, rather than a complete ready-to-install package.

Work on both surfaces

The language tells me what needs my attention. The statusline tells me what state the work is in. Together they reduce the amount I have to hold in my head just to use the tool.

Choose the signals you keep looking for, make them visible, and revise them as you work.

[ comments ]

guests welcome. members show first in the list

no comments yet. start the thread_

1 / 7