Backend · Artificial Intelligence

Obsidian Is the IDE, the LLM Is the Programmer, and the Wiki Is the Codebase

By 犀利豆 · Jul 16, 2026

Read original on juejin.cn ↗ Google Translate ↗ Alt translation

The bottleneck in AI-assisted development is no longer code generation but context continuity. This workflow shifts the maintenance burden of a project wiki from humans to the LLM, making long-cycle, multi-service refactors feasible where IDE-only approaches collapse under session resets and lost rationale.

Summary

Traditional IDEs treat AI as a disposable Q&A sidecar: context resets with every session, and decisions vanish into chat history. A workflow built on Obsidian and the Claudian plugin flips this by making a Markdown vault the persistent home for all project context—architecture, DB schemas, decision records, and AI dialogue.

The approach draws directly from Andrej Karpathy's gist that positions the LLM as a programmer maintaining a wiki codebase. A `CLAUDE.md` schema file disciplines the model into a maintainer rather than a chatbot, while regular Lint operations scan for dead links and orphan notes to keep the knowledge graph alive.

A two-week, 15-decision refactoring case study shows the leverage: a master plan note anchored every session, commit hashes were written back as a changelog, and context carried forward with a single "read this note and continue" prompt. The larger the project, the more this infrastructure becomes a necessity, not a nicety.

Takeaways

— Every project gets its own Obsidian vault with a root `CLAUDE.md` that defines the tech stack, conventions, and gotchas so the LLM never starts cold.

— AI output is treated as persistent documentation, not disposable chat; key decisions, commit hashes, and rationale are written back into linked notes.

— A typical requirement flow moves from a requirement note through AI analysis and code changes, with every decision recorded for the next session.

— Regular Lint scans expose dead wikilinks and orphan notes, producing a concrete action list of missing concepts and structural holes in the knowledge graph.

— Dead links that are referenced frequently but never written represent concepts the developer repeatedly felt should exist—Lint turns that intuition into a prioritized backlog.

— Orphan note clusters signal a missing Map of Content (MOC) index, a structural problem invisible when viewing any single note in isolation.

— The IDE remains the place for writing, debugging, and profiling code; Obsidian handles the "brain" work of design, context, and decision tracking.

Conclusions

A vault without periodic Lint degrades into a graveyard of dead links and isolated islands within two to three years, regardless of how diligently notes are created.

The shift from "AI as Q&A tool" to "AI as disciplined wiki maintainer" hinges on a single configuration file—`CLAUDE.md`—which acts as a schema that constrains model behavior.

Karpathy's framework succeeds now because LLMs have zero marginal cost for bookkeeping tasks like updating cross-references across dozens of files, a chore that broke human-maintained wikis.

The division of labor is clean and non-overlapping: the IDE owns execution (code, tests, profiling), while Obsidian owns cognition (design, context, decisions). Neither tool is trying to replace the other.

Concepts & terms

Vault

An Obsidian vault is a folder on disk containing plain Markdown files. It is the unit of organization for a project's knowledge base, holding notes, attachments, and configuration.

CLAUDE.md

A schema file placed in the vault root that defines project context, conventions, and constraints. It is read automatically by the Claudian plugin at session start and disciplines the LLM into acting as a structured wiki maintainer rather than a generic chatbot.

Wikilink

A double-bracket link syntax (`[[note-name]]`) used in Obsidian to create bidirectional connections between notes. The graph of these links forms the knowledge network that the LLM maintains and that Lint operations audit.

Map of Content (MOC)

An index note that links to a collection of related notes, serving as a navigational hub. Orphan clusters often indicate a missing MOC that should tie a series together.

Lint (for knowledge bases)

An automated health check that scans a vault for dead wikilinks, orphan notes, contradictions, and outdated statements. It treats the wiki as a codebase subject to the same quality discipline as software.

Architecture Decision Record (ADR)

A single note capturing the context, alternatives considered, chosen option, and rationale for an architectural decision. Linked ADRs form a traceable decision history across a project's lifetime.

From the discussion

Featured comments

用户51707184966

Really interesting idea! With all the AI coding I've been doing lately, I've definitely noticed that interacting with AI is still a systematic conversation made up of individual text fragments. Right now my approach is to manually copy and archive the key outputs. I want to connect it to Obsidian but don't know where to start.

犀利豆 · 1 likes

Use this plugin, Claudian

用户51707184966 → 犀利豆

Thanks! I'll give it a try

See top comments, translated →

Source: juejin.cn ↗ Google Translate ↗ Backup ↗