pyQuejica: because "Synthesizing..." is too dignified

#ai #claude #cli

Years ago, when the first robot vacuums hit the market — Roombas, Congas, and their kind — I had an idea I called Proyecto Quejica (Project Whiner): give them a voice. Not a helpful voice. A complaining voice.

Imagine your vacuum cleaning the floor while muttering "How many times do I have to tell you to clean up after yourself..." or "This place is an absolute disaster...". And when it bumped into the leg of a chair — as those early models constantly did — it would hurl an insult into the air.

No utility whatsoever. That was the point. Some things deserve to exist purely for the pleasure of making them.

I forgot about that idea for years. Then I started using Claude Code.

The problem with "Synthesizing..."

While Claude Code is reasoning or executing tools, it shows you a rotating label. A status indicator. Something to let you know the model is working.

The labels look like this:

Synthesizing... Reflecting... Analyzing... Planning...

They are, frankly, too polished. Too corporate. They describe the activity with the dignity of a consultant's slide deck. And they are hardcoded in the bundle.

When I first noticed this I thought: what if they didn't have to be?

Finding the labels in Claude Code

Claude Code is distributed as an npm package. The main bundle is a single minified JS file — cli.js — sitting somewhere in your node_modules. The exact path depends on how you installed it (global npm, nvm, Homebrew), but which claude and a couple of symlink resolutions will get you there.

Inside that minified bundle, there's a pattern like this:

return[...VAR,...q.verbs]

That VAR is the array you're looking for. Find its initialization and you've found the thinking labels.

The tricky part: newer versions of Claude Code ship as a compiled ELF binary (claude.exe), not plain JS. The array is still there — as a UTF-8 encoded byte sequence in the binary — but you have to patch it in-place, preserving the exact byte length. Shorter replacement? Pad with spaces. The JS parser inside the runtime ignores them. Longer replacement? You're out of luck — you can't expand a binary without corrupting it. Keep your verbs short.

The patch script

patch_claude_verbs.py handles both cases.

It locates the claude binary, resolves symlinks, detects whether it's dealing with a JS bundle or an ELF binary, finds the verb array, replaces it, and on macOS applies an ad-hoc codesign so the patched binary runs without Gatekeeper complaints.

It's idempotent — run it ten times and it only ever patches once, preserving the original backup. It also handles the one real maintenance cost of this whole thing: every npm update overwrites the bundle and removes the patch. Just run the script again.

python3 patch_claude_verbs.py          # apply the patch
python3 patch_claude_verbs.py --status # show current verbs
python3 patch_claude_verbs.py --restore # restore original

And the result:

· maldiciendo…

(Cursing. While reading your codebase. Honest.)

Gemini CLI: a different animal

Gemini CLI uses a different architecture. Instead of one big bundle, the code is split into chunks. There's no central array of thinking labels — instead, specific tool-action strings are scattered across those chunk files:

"Searching the web for:", "Executing command", "Thinking"...

patch_gemini_verbs.py scans all the chunk files, finds those strings, and replaces them with a translation dictionary you control:

TRANSLATIONS = {
    'Thinking': 'Firibicundiando',
    'Searching the web for:': 'Snooping around the internet about:',
    'Executing command': 'Doing the dirty work:',
    ...
}

The result in the terminal:

Firibicundiando... (esc to cancel, 2s)

Firibicundiando is a made-up Spanish word. Morphologically impeccable, semantically empty, expressively perfect. It is, genuinely, a better status indicator than "Thinking".

The verb corpus

The current list for Claude Code has 180 verbs. A selection:

Physical effort: Bufando (huffing), Resoplando (puffing), Agonizando (agonizing), Arrastrándose (dragging itself)
Mental block: Cavilando (brooding), Elucubrando (speculating wildly), Desvariando (rambling), Fraguando (scheming)
Despair: Claudicando (giving up) — meta, given the model's name — Naufragando (sinking), Desmoronándose (crumbling)
Technical: Refactorizando, Rollbackeando, Commiteando, Defragmentando
Corporate irony: Stakeholdeando, Onboardeando, Alineando, Escalando
Existential: Nihilizando, Epifaniando, Existiendo, Absurdizando
Religious: Encomendándose (commending itself to God), Flagelándose (flagellating), Mortificándose

The selection criterion: what would an experienced, slightly cynical software developer actually mutter while doing this task? Some highlights:

"Mintiendo" (lying) — if it appears during a git commit, that's pure statistical art.
"Procrastinando" (procrastinating) — breaks the fourth wall. The model admits it's not doing what it should.
"Rezando" (praying) — implies even the AI isn't sure this is going to work out.
"Claudicando" — the model named Claude, giving up. This one needed to be in the list.

Why bother

There's a version of this post that argues for some deeper point about anthropomorphism, or human-computer interaction, or the semiotics of status indicators in conversational AI.

I'm not going to make that argument.

The honest answer is: Synthesizing... felt too smug. I had a free afternoon and a text editor. The Roomba that mutters insults when it hits the furniture still deserves to exist. And so does a language model that admits it's Rumiando while it reads your spaghetti code.

Some things don't need to be useful to be worth doing.