\n<\/p>\n
Languages: English \u00b7 \u0420\u0443\u0441\u0441\u043a\u0438\u0439 \u2014 Agent instructions:
\nEnglish \u00b7 \u0420\u0443\u0441\u0441\u043a\u0438\u0439<\/p>\n
llm-wiki (original gist by Andrej Karpathy)
\nis the idea of keeping a codebase’s documentation as an LLM-readable wiki:
\nevery meaningful directory carries an index.md, pages
\nare small and single-purpose, links are bidirectional (parent \u21c4 child), and one
\npage owns each detail. An agent navigates the tree top-down through index.md
\nfiles instead of blindly grepping.
\nllm-wiki-tags is exactly that, plus tags \u2014 nothing about llm-wiki
\nchanges; tags are added on top.<\/p>\n
A tag is a short kebab-case slug written as the token @tag:. The same
\ntag is placed both in code and in documentation, which creates an explicit
\ncode \u21c4 documentation link that does not depend on the directory tree. With
\ntags you can:<\/p>\n
find code and docs by tag \u2014 one search returns every file (code or docs)
\nthat carries a concept, even when they are scattered across the tree;
\nsee which tags a piece of code has \u2014 read the tags at the top of a file or
\ndirectory to learn which cross-cutting concepts it participates in;
\nsee which tags a document has \u2014 the same, for a doc page;
\nkeep a single tag registry (docs\/tags.md) describing what each tag means.<\/p>\n
Tags complement index.md navigation (which follows the directory tree) with a
\nsecond, orthogonal axis: a concept that spans several folders is reachable in one
\nstep.<\/p>\n
Slugs are lowercase kebab-case ((a-z0-9-)+); the same @tag: token is
\nplaced in both code and docs.<\/p>\n
Documentation (.md): in YAML front matter at the top of the file, a
\ntags field holding the space-separated tokens; for a directory, in that
\ndirectory’s index.md. A tag that applies to just one section may instead sit
\nas a @tag: line in the body next to it.
\n—
\ntags: “@tag:payments @tag:retry”
\n—<\/p>\n
Code: a comment in the language’s syntax containing the token, placed above
\nthe element. A tag can mark a package, a file, a class (or
\nequivalent) or a method\/function. Per-language rules live in
\nlanguages\/ \u2014 Java,
\nPython, Go.
\nMultiple tags: repeat the token, space-separated: @tag:ui @tag:mechanism.<\/p>\n
# every code + doc location that carries a tag
\ngrep -rn “@tag:” .<\/p>\n
# which tags a given file has
\ngrep -oE “@tag:(a-z0-9-)+” path\/to\/file<\/p>\n
# every tag used in the repo
\ngrep -rhoE “@tag:(a-z0-9-)+” . | sort -u
\nEvery tag is registered once in docs\/tags.md with a one-line description.<\/p>\n
llm-wiki-tags keeps (and makes explicit) llm-wiki’s documentation rules:<\/p>\n
Every meaningful directory has an index.md giving a one-line
\ndescription of each file and each sub-directory in that folder.
\nIndex files describe stable concepts, not changelogs. Prefer many small pages
\nover one large document; put local detail next to the code it describes.
\nBidirectional navigation: parent indexes link to child pages; child pages
\nlink back to the parent index and to related pages.
\nOne page owns a detail; other pages link to it \u2014 do not duplicate.
\nRead before you act: before a task, follow index.md files from the
\nnearest directory down to the code you will touch.
\nUpdate as you go: during or after the task, update the affected index.md
\nfiles (and tags) in the same change.<\/p>\n
Point your coding agent at the instruction page and ask it to install:
\n\u043f\u043e\u0441\u0442\u0430\u0432\u044c https:\/\/github.com\/mpashka\/llm-wiki-tags\/blob\/main\/INSTRUCTIONS.md
\n# or, in English:
\ninstall https:\/\/github.com\/mpashka\/llm-wiki-tags\/blob\/main\/INSTRUCTIONS.md<\/p>\n
The agent reads INSTRUCTIONS.md (or
\nINSTRUCTIONS.ru.md) and sets up the index.md tree, the
\ntag mechanism and the docs\/tags.md registry in the current repository, and
\nrecords the convention in the repo’s agent guide so future agents keep following
\nit.<\/p>\n