Matjaž Domen Pečan

Open minded? Friendly? Tech-savvy? Idealistic? Carebear? Cookie?

[Davorin Pavlica @DavorinPavlica]

Hey, @AnthropicAI, my Claude Code is broken. It says "Usage limits reached" but my dashboard shows I've only used 13% of my Pro plan and I even have €9.24 in credits.

This has been happening for a week.

Can you fix this billing routing bug. IT'S ANNOYING AF #brokenClaude

[Matjaž Domen Pečan @genom]

I ran the numbers for LAP Coffee's 32 German stores. Effective rating: ~3.8. Displayed: 4.6.

Almost a full star of missing reviews.

https://t.co/oaS5EdifnV

[Matjaž Domen Pečan @genom]

Google now tells us how many reviews each business has had removed for defamation.

What does that look like for a chain everyone has an opinion about?

[Davorin Pavlica @DavorinPavlica]

Če bi že gradili metro (u-bahn/s-bahn, bi NUJNO morali podaljšati do Jesenic / Kamnika / Maribora (vsaj Celja) / Zidanega Mosta / Novega Mesta / Ilirske Bistrice (vsaj Postojne) en podaljšek do Kopra in en še do Gorice.

Potem ostaneta od večjih mest brez samo Idrija in Bovec. :)

[Matjaž Domen Pečan @genom]

rable only exists because ldayton's Parable[https://t.co/mcbf7lhAsV] existed first. Couldn't use it directly (Python dep for a Rust binary), but the test suite was invaluable.

https://t.co/6LFld0Kqhz

[Matjaž Domen Pečan @genom]

I had Claude Code build a bash parser from scratch, hands-free, using Parable's test suite as the spec.

The result: rable — 99.9% accuracy across 1,783 test cases.

tree-sitter-bash manages 93.8%.

[Matjaž Domen Pečan @genom]

The timeline:

13h in: 93.6% passing
17h in: 100% on Parable's suite

But the architecture was a mess. The real work was directing the refactor, not generating code.

[Matjaž Domen Pečan @genom]

Lesson: building an AI content system is easy. Making it funny is the hard part.

https://t.co/xtczojZwCg
https://t.co/fF1GrCdHZK
https://t.co/8jt4AONZym

Completely different tone now. Give them a look.

[Matjaž Domen Pečan @genom]

Friends kept saying the same thing about my AI satire sites: "It's clever but... weirdly dark?"

They were right. I let the AI drift into dystopian corporate thriller territory and didn't notice.

So I rebuilt the whole pipeline. The difference is wild.

[Matjaž Domen Pečan @genom]

Before: Bezos tracking "extraction yield percentages." Zuck running "psychological conditioning protocols."

Now: Bezos at a wellness retreat where joy is toxic to him. Zuck brought perfect hexagonal cookies to a bake sale. His wife: "that's somehow worse."

[The Reptilian Diaries @MetaLizardKing]

LEAKED: Mark's reaction to Instagram dropping E2EE

"Satisfy attorney general visibility requirements" = compliance theater
"Higher extraction density" = the actual play
"Optimal alignment opportunity" = both at once

His diary decoded this move WEEKS ago #Meta

[Matjaž Domen Pečan @genom]

New post: Getting Into Make, Just, and Hooks

When Claude Code runs `just test`, tokf can intercept the outer command — but what about the `cargo test` inside?

I solved this by injecting tokf as a custom shell and shimming PATH lookups for git hooks.

https://t.co/E1XetyqksA

[Matjaž Domen Pečan @genom]

Also shipped since the last update:

• make/just support via shell-override wrappers
• env var prefix stripping (NODE_ENV=test npm test now matches)
• tokf auth delete-account — removes everything immediately
• Terms of Service: https://t.co/3tNrOq0n0U

[Matjaž Domen Pečan @genom]

tokf now works with Claude Code, OpenCode, and OpenAI Codex CLI.

Also shipping this week: a community catalog of shared filters.

The question I'm watching: does the community converge on better compression than the stdlib? Early data will tell.

https://t.co/oCjTiNPeMD 🧵

[Matjaž Domen Pečan @genom]

Each filter page shows the full TOML with syntax highlighting.

Lua filters are sandboxed and inlined at publish time — what you see is what runs. No network calls, no side effects beyond the filter output.

[Matjaž Domen Pečan @genom]

Smarter pipe handling: simple pipes (`| grep`, `| tail`) get stripped — tokf's filter replaces them. Gain accounting tracks what the original pipe would have consumed.

Complex chains or unknown targets are left alone.

https://t.co/FiRcsEabu5
https://t.co/OJLGPgFm7h

[Matjaž Domen Pečan @genom]

tokf update — the interesting part of building a CLI filter for LLMs is adapting to how the model actually runs commands.

A few things I learned watching Claude Code in the wild 🧵

[Matjaž Domen Pečan @genom]

History hints: tokf can now suggest to the model that raw output is available via `tokf history show <id>` — so it can dig deeper when the filtered summary isn't enough.