Lachlan Kermode

Can't sleep at night for want of knowledge gardens (Zettelkasten-style) in Typst? Stay awake no longer because rookery is here: rookery.ohrg.org

Rookery is also the basis for a more structured and local-first way of doing context engineering, the vision for which I wrote about on my blog: www.ohrg.org/ideas/birdke...

Can't sleep at night for want of knowledge gardens (Zettelkasten-style) in Typst? Stay awake no longer because rookery is here: rookery.ohrg.org

I built rookery to migrate my knowledge management system (read: scores of plain-text files in a folder) from Org-mode to Typst. It has now become the basis for my todo tracking, my book/paper/meeting notes, my brainstorming, my slideshowing, and---most importantly---my writing.

More sage thinking about LLMs and the symbolic function of writing from Cantrill: "to use an LLM to write is to void the social contract between writer and reader: we readers shouldn’t be expected to labor to understand a sentence that the writer themselves didn’t work to create."

I also think this might be the foundation for more reasonable thinking on why writing CODE with LLMs might be more acceptable than writing prose with them.

I've been stewing on what to say about the discomforting industrial metaphorics of Steve Yegge's beads and gastown for a while, and here it is: www.ohrg.org/birdkeeping

I've just released github.com/freecomputin..., an extension set for your coding agent to work effectively with jj and beads.

The basic idea is that LLM agents do much better when you parcellate work for them in advance. This kind of planning also opens up more economical downstream agential workflows such as using a local model to implement work that a more capable model from one of the frontier labs has planned.

I've just released github.com/freecomputin..., an extension set for your coding agent to work effectively with jj and beads.

Let's bring back Weizenbaum's excellent phrase and accusation "the artificial intelligentsia"

Does anyone know of an intro to CS that includes Weizenbaum's _Computer Power and Human Reason_? What with this AI business, it seems more essential than ever to get this in front of computer people as early as possible

Coding workflow: 1. Have Anthropic's Opus family design a feature or fix as a set of beads issues: github.com/Dicklesworth... 2. Using either Anthropic's Sonnet family or GLM 4.7 (an open weights model), turn on caveman mode: github.com/juliusbrusse... 3. "br/jj pair": github.com/breezykermo/...

4. Follow the implementation with abacus - github.com/ChrisEdwards... - and tuicr - github.com/agavra/tuicr

Finally beginning to feel that I'm hitting a happy medium between 1) letting Claude Code run away with arbitrary design decisions in favor of functionality and 2) writing code through sharply scoped directives (or by hand) and thus developing real awareness of the codebase as I build it out

Beads - github.com/gastownhall/... - and abacus - github.com/chrisedwards... - are pretty essential to my workflow here, as they let me more effectively pair program with the agent, editing a plan for future work (beads issues) as I move through implementation

Inspired by some folks at @eeg.cl.cam.ac.uk, I've started keeping research weeknotes. Rationale for the undertaking and link to the weeknotes are below

There's unfortunately no RSS feed for these yet! But I hope to add this to Rheo (the tool I use for the notes) soon

Forget LLMs: PowerPoint had already ruined our ability to think straight and seriously