A Saturday in Claude Code

Saturday morning, before exercise, I asked Claude:#

is there a way for external process to inject text into TUI (say claude code’s input window)?#

This started the creation of pair, a small tool that makes Claude Code’s input box better. It actually works for any TUI program. Writing this down for people curious about what “agentic coding” is. Full transcript of this coding session is available here, extracted by the same process that created pair. The transcript took 49 rounds of exchanges. See? It’s not one shot or a few shots. Agent is your knowledge base and your execution engine. But you define its purpose through those exchanges.#

And the initial draft of this blog post was done as the last prompt to the agent — recursion all the way down!#

How it started #

Pure curiosity. Claude Code’s input box had been irritating me for a while now:#

• Not a real editor. Yes, you can do Ctrl-g to bring up nvim, but returning to Claude often messes up previous printed lines. You can’t search in that buffer, no mouse support either. Plus, you can’t see the history anymore while in the editor, a huge drawback.
• Accident-prone. Easy to accidentally hit Return and send the prompt prematurely, particularly when you copy-paste from somewhere.
• Copy from Claude gives hard wraps. Select a paragraph from Claude’s output and your clipboard has \n baked in at every terminal column. Paste it anywhere that respects soft-wrap and it looks mangled. I often do this, with a leading >, so I can indicate what I’m referring to.

I want to make my workflow better.#

By the afternoon there was a new repo named pair, a pensive committed to my notes, four issues filed, a lessons.md entry, an atlas/architecture.md. By dinner, there was homebrew set up for it.#

This is the story of how that happened, and an example of how far personalized software can go.#

The wall is shorter than it looks #

The first half hour was just back-and-forth on terminal mechanics with Claude, mostly the agent giving me ideas of what was possible.#

tmux send-keys works. zellij has zellij action write-chars. TIOCSTI is dying (Linux moved it behind a flag, OpenBSD removed it). OSC52 is mostly write-only. The pieces are there if you compose them right.#

About 60 minutes in, the design had assembled itself in my head and the agent’s context:#

• Run zellij with a horizontal split.
• Top pane: claude (or codex, or gemini — same plumbing works for any TUI agent).
• Bottom pane: nvim editing a persistent draft file.
• A keybind in nvim that takes the buffer, types it into the agent pane via zellij action write-chars, then clears the draft.
• A keybind in the agent pane that grabs whatever the user has selected and dumps it into nvim, quoted with a > prefix.

Nothing exotic.#

From chat to artifact #

I stopped and asked Claude to “make a pensive out of this.” A pensive is what I call a thinking-out-loud note that lives under docs/vision/ in what I would call brain repo. It’s just a staging ground of persisted memory and discussion I had with the agent, continuously being organized. Pensive itself is not a spec, not a plan, just the moment when the thought happened, in my voice, with the open questions still open. In my “agentic workbench” called ariadne, I have a meta skill called datatype, and pensive is merely one kind of semi-typed markdown file.#

I didn’t write the pensive myself. Claude distills the conversation into prose, picks a slug, places the file at the right location. I copied it over to the pair repo for your reference.#

Issue, then a new repo #

Once the pensive was in, I filed an issue against brain/workshop/issues/. Halfway through writing it I changed my mind: this isn’t a brain thing — it’s its own thing. So, new repo.#

The bootstrap was one command:#

cd ../pair && ../ariadne/construct/setup --vendor

Ariadne is my base layer for “AI-assisted single-developer workflow” — AGENTS.md, Makefile.workflow, the workshop/ folder structure, the atlas/ map, a set of xx-* skills for typed documents and issue management. --vendor copies all of it into the new repo so it’s self-contained, no symlink dependency on my home checkout. From an empty directory to fully scaffolded in about a second.#

I moved the issue from brain to pair/workshop/issues/000001-pair-nvim-driven-tui-agent-setup.md. Renumbered, updated the pensive cross-reference path, ran make issue-sync to push the renumbered issue to GitHub. That is the entire “starting a new project” overhead. I am pretty amazed every time.#

Building it #

The build itself was Claude doing the typing and me doing the testing and steering. pair took shape very fast. Well, there wasn’t that much code to write; rather, you need a plethora of knowledge in several rather arcane areas, and AI is perfect for that.#

Normal back-and-forth: test what’s implemented, notice ergonomics that can be improved, rinse and repeat.#

The side issues #

This is the part I want to point at.#

While building issue #000001, several side concerns surfaced. Each became its own issue. Some examples:#

• #000002 — each pair session may have a customized name.
• #000003 — make a blog post (this one) about how this was made.
• #000004 — Claude filed this one on its own, while I was complaining that the auto-attach behavior was surprising in a long-lived-session world.

As you can see, AI is great at tracking things, if you give it some structure for how things should be tracked. All of that is in the ariadne layer. You might wonder why I put all these processes inside a single repo. This design originated from my strong desire to remove the chrome, the incidental complexities. Everything is a file (with processes around it), carrying its full history, in a uniform way. Plain and simple.#

The shape of this kind of work #

A few hours, a few breaks. Artifacts:#

• One pensive.
• A new repo pair with seven files of code and config, a README, an architecture atlas, four issues, a lessons.md entry, ~30 commits to main.
• A working tool that’s homebrew installable, by the time of this writing.

Most of those artifacts I would not have produced on my own. I have no idea about those terminals and how they behave. But I have a clear view of what I want, and this vision gave the AI its purpose. After all, it’s the purpose that defines us. ;)#

This is the magic of “agentic coding”!#