Connect everything
The more Claude can access on demand, the better the work.
The job is to build paths to discovery: sharp AGENTS.md, well-described Skills, a .md wiki of everything that matters, and clones of relevant remote sources. Connectedness is what turns a pile of files into a system.
Sharp AGENTS.md as the entry point
AGENTS.md is the first thing Claude reads on every session. Treat it as the index, not the encyclopedia. It should:
- Point to the top 5–10 things Claude might need: the spec location, the research folder, the Skills directory, the deploy runbook.
- Describe each link in one line. The description is how Claude decides whether to follow it on a given turn.
- Stay short. A bloated
AGENTS.mdis itself a context-budget problem (see Managing context).
The entry point’s job is to make the next hop obvious.
Skill frontmatter as a discovery hook
Skills live outside AGENTS.md but are discovered through their frontmatter descriptions (see Skills). A Skill with a vague description hides; one with a specific description fires when the matching task appears. Treat the description as a query-target pair: what phrases in a user prompt should make this Skill load?
The .md wiki
Beyond AGENTS.md and Skills, most projects accumulate a folder of .md files that capture the project’s knowledge over time: research artifacts, architecture decisions, competitor notes, customer interviews, meeting summaries. This is the project wiki.
Organize it with a flat-ish structure and clear filenames:
- Prefer a shallow folder hierarchy (two or three levels deep, max).
- Name files by what they contain rather than when they were written.
vector-db-comparison.mdbeats2026-03-research.md. - Keep an index file (e.g.
docs/index.mdordocs/README.md) that lists and briefly describes each wiki entry, soAGENTS.mdcan point at one place.
This is essentially the pattern Andrej Karpathy has called an LLM-backed personal wiki: LLMs summarize, cross-reference, and maintain the knowledge base, while you feed raw material in. The pattern has spawned a small ecosystem of Obsidian plugins, local-first variants, and visual explorers.
The knowledge graph
Markdown as the medium covered the mechanics of cross-linking (markdown links vs. wikilinks). The payoff at the project level: when .md files reliably cross-link, the folder becomes a graph, and Claude traverses that graph the way a new engineer would: clicking through, building understanding across multiple files.
Obsidian has deep community experience turning .md folders into browsable knowledge graphs; its graph view visualizes what the LLM has connected. Worth looking at once even if you don’t adopt Obsidian full-time, because it makes the shape of your project legible.
Cloning remote sources into local .md
Anything Claude needs to reference repeatedly should live locally, not behind an API call. Common cases:
- Linear tickets: periodically export relevant tickets to
docs/tickets/. Claude can search the history without rate-limited API calls. - Slack threads: copy decision-relevant threads into
docs/decisions/. Raw chat transcripts are fine; Claude reads them as they are. - Google Docs / Notion pages: convert to markdown at the start of any session that needs them. One-shot conversion beats on-demand fetching.
- Vendor documentation: for deep dives, clone the whole docs site. Claude can
grepit instantly (see Deep research).
The rule: remote when it’s dynamic, local when it’s reference material.
Paths from broad to specific
A well-connected project has obvious paths from the general to the specific:
AGENTS.md
└─ "See specs/current/ for architecture"
└─ specs/current/index.md
└─ "See specs/current/auth.md for auth design"
└─ specs/current/auth.md
└─ "Implementation in src/auth/"
Claude follows these paths the same way a new engineer would. Invest in making the paths obvious and the project becomes progressively easier to work in, for Claude and for humans.
Resources
- Obsidian: markdown-native knowledge tool with wikilinks and a graph view
- Zettelkasten method: the older note-linking methodology the LLM-wiki pattern draws from
- Andrej Karpathy has written and talked about the LLM-backed personal wiki idea on his X account; search there for the most current framing