Skip to content

Releases: asakin/llm-context-base

v1.1.0 - The Schema Gets Real

12 Apr 07:32
@asakin asakin
v1.1.0
63396c3
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Choose a tag to compare

View all tags
v1.1.0 - The Schema Gets Real Latest
Latest

v1.1.0 — The Schema Gets Real

Everything in llm-context-base now runs on YAML frontmatter. The metadata standard that was designed to make every file queryable without reading it — it's now a proper schema, not a loose convention.

This release also ships the features that emerged from real daily use: an onboarding system that teaches you the tool as you use it, a lint system that knows when "archived" pages shouldn't be flagged as stale, and a tools manifest so you can declare what your wiki talks to without installing anything upfront.

Headline: YAML frontmatter migration

All metadata blocks are now YAML frontmatter (--- fences at the top of every file). This is the one change that affects existing users — if you started on v1.0.0, your files use the old bullet-list format. Both work, but new files will use YAML, and the lint system expects it going forward.

New features

  • Onboarding tip system — contextual 💡 hints during training and cooldown phases. The AI picks tips based on what you haven't tried yet. Shown at session start and inline when relevant.
  • Tools manifest — declare the tools your wiki integrates with (Obsidian, MCP servers, export targets) in _config/tools.md. The AI offers to install them on first session.
  • Project Index — 1-Projects/README.md now auto-maintains a table of all projects with status, summary, and last-updated date.
  • Knowledge ops log — append-only log at 2-Knowledge/log.md tracks every ingest, query, lint, and edit. Provenance for your wiki.
  • Confidence field + superseded status — metadata standard now supports confidence: (how sure you are) and status: superseded with a required failure_reason.
  • Lint staleness exemption — pages with terminal statuses (archived, superseded, cancelled) are no longer flagged as stale. Finally.
  • Conversation Tone section — _config/config.md now has a dedicated section for controlling how the AI talks to you vs. how it writes files.
  • Folder-per-project model — projects can now be a folder with an _overview.md and supporting files, not just a single page.

Docs

  • LLM Wiki pattern comparison — how llm-context-base compares to Karpathy's original gist and other implementations
  • Chat LLM support — how to use Claude Chat or ChatGPT with your wiki via GitHub sync
  • Obsidian setup rewrite — Minimal theme, appearance config, graph settings, Terminal plugin
  • Journal sync and privacy — options for keeping 3-Journal/ private while syncing everything else
  • Claude Code setup guide — starter settings.json and recommended permissions

Fixes

  • docs/ and _meta/ files no longer require the metadata standard (they're framework files, not wiki content)
  • Framework mode detection uses a sentinel file instead of fragile git-remote checks
  • Training footer formatting cleaned up, questions numbered, meta-questions about the system banned
  • Obsidian defaults to reading mode with hidden inline titles

Full Changelog: v1.0.0...v1.1.0

Assets 2
Loading

v1.0.0 - Foundation

08 Apr 19:34
@asakin asakin
v1.0.0
779b933
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Choose a tag to compare

View all tags
v1.0.0 - Foundation

llm-context-base v1.0.0 - Foundation

The first stable release. No new features, just the existing system made solid and documented for public use.

What this is

A git template for building your own LLM-powered personal knowledge base, built on Andrej Karpathy's LLM Wiki pattern. The difference from a bare wiki: it learns how you work.

What's included

  • Two-section README - personal section maintained by your AI, framework docs tracked cleanly upstream
  • Metadata standard - every file gets a queryable header so your AI scans summaries, not full docs
  • Training period - 30-day adaptive phase where the AI asks questions, suggests structure, and logs what it learns. Goes quiet after.
  • Inbox-first capture - everything lands in _inbox/ first. You never hesitate to capture because you don't have to decide where it goes.
  • Instruction modules - lint, knowledge query, write, and definition-of-done behaviors, loaded just-in-time
  • Templates - decision records, knowledge articles, project briefs

Supported tools

Claude Code, Cursor, Copilot, Windsurf, and anything that reads a CLAUDE.md or system prompt from a folder.

Get started

Click Use this template, clone it, fill in _config/config.md, and start talking to your AI.

Apache 2.0 - Built by @asakin

Loading