gwern/gwern.net

Haskell source files are usually clean because I have an Emacs save-hook which strips trailing whitespace on Haskell/R/Bash/Python files. This doesn't help on the CSS/JS/PHP because usually someone else (like Said Achmiz or a LLM) is editing them. This leads to noisy diffs (and slightly wasteful token use). We should add a simple git hook to strip trailing whitespace from all text files in the repo before each commit is saved. (As far as I know, no text files in it depend on trailing whitespace.) Claude-4.7-opus suggests a build/git-hooks/pre-commit VC-controlled Bash script enabled by git config core.hooksPath build/git-hooks, that operates on the git index directly (rather than the working tree). We can further ensure that text files are valid UTF-8 Unix text files without legacy BOMs, which I believe should also be true of all repo text files, and is worth ensuring. Claude's prototype: #!/usr/bin/env bash # build/git-hooks/pre-commit — sanitize staged text files. # - validate UTF

We enforce -Wall -Werror on Haskell, and it would be good to have clean safe Bash; all Bash scripts or utilities should be shellcheck-clean. Some of the shellcheck warnings cannot be fixed because they are warning about intended functionality or are just dubious. In those cases, presumably there is some way to whitelist per-line (some sort of comment?), or to add a shellcheck config local to this repo disabling a particular category of warnings.

https://gwern.net/lorem-christmas In Christmas mode, everything should cheerfully alternate green-red, like the list markers do. Currently, the SVG sun/star/moon triplet, and other SVG UI icons like the collapse circle, do not.

Gwern.net has a number of sources of "toil", which are a continual suck of time & energy in terms of rote maintenance. Tagging annotations (to be fixed by #31 & #32) Cleaning/checking archived web pages (agentic vision-enabled LLMs?) -- Excerpting/abstracting/summarizing (LLMs again, possibly Hypernym?) General sync.sh lint fixing (to be fixed by agentic LLMs drawing on the completed Manual of Style) Fixing/blacklisting as many 404s as possible Right now, there is some support for semi-automated fixing of 404s, by me piping in a hand-selected list of errors into the edit-distance-based guessing utility nginxredirectguesser.hs, which looks at existing nginx redirect rules, does some exact and edit-distance distance searches (weighted by popularity), and tries to create a new rule. I have to hand-review them because it makes many mistakes, and it is quite tedious. I also cannot use it to generate more broad rules, like the obvious patterns generated by scanning tools/malware/bugs. The

Pleometric used a bunch of Claude-4 calls to try to analyze the current repo and create a 'documentation website' (using the CMS Docusaurus v3.9.2) explaining the structure of the repo, backend, frontend, design patterns, etc: https://gwern.pleometric.net/ To create the docs, I first built a rudimentary map with notes I wrote while reading the codebase. I then fed this map to multiple Claude Opus 4.5 agents, with each cohort documenting a different part of the codebase. I manually reviewed the output and did a secondary fact-checking step with GPT 5.2 and Claude again. It is pretty good overall but still has substantial errors (often relating to key docs or source code not being in the agent's context window?). I would like to incorporate it since a 'technical implementation whitepaper' overview is missing, and not provided by existing docs like the per-module source code comments (however extensive) or "Design" or "Design Graveyard". But it's unclear how to do so as I do not want to