A Karpathy-inspired wiki + knowledge graph + hook setup that stops Claude Code from re-exploring your codebase every conversation.
Teaches your Claude Code instance to look things up the way an experienced engineer would:
- Query a knowledge graph first (fast index, low token cost)
- Read a hand-curated wiki second (context the graph can't infer)
- Only touch source code when the first two didn't answer
Enforced by hooks so Claude can't skip the order even if it wants to.
Clone this repo, open it in Claude Code, and say:
Read SETUP.md and set this up in my environment.
Claude will ask you a handful of questions (project name, what to call your wiki, where it lives, etc.) and wire everything up: the wiki directory, hook config, CLAUDE.md, graphify install, and a verification run.
Expect 5 to 10 minutes for the first-time setup, most of which is graphify building its initial graph.
- A wiki (you name it, default suggestion is your project name + "-brain") with
entities/,concepts/,maps/pages Karpathy-style - graphify (open source knowledge graph over your code)
- Hooks that remind Claude of the work order on every prompt, tool call, and git push, plus the Boil the Ocean standard so partial fixes don't ship
- A
CLAUDE.mdin your repo that documents the workflow for future sessions (and future teammates) - Guardrails against directory drift and credential leaks (optional, can skip)
- Stop hook - warns at session end if you have uncommitted or unpushed work (catches the "I forgot to commit" failure mode)
/adrslash command - capture architectural decisions into your wiki with one command- Scheduled jobs - daily TLS cert sweep, nightly Go test-coverage report, daily Claude-spend rollup, weekly security + SEO drift checks. macOS launchd + Linux cron both supported. See
docs/scheduled-jobs.md.
See WHY.md for the token-waste story. Short version: a 50kloc codebase re-explored from scratch every conversation costs dozens of tool calls and hundreds of thousands of tokens before Claude even understands where anything is. This fixes that.
Karpathy's LLM wiki approach is the spiritual inspiration. graphify by @graphifyy does the heavy lifting for the graph layer. Everything here is the glue.
MIT.