Reading CLAUDE.md, AGENTS.md, and other agent-generated markdown

The volume of markdown an AI coding agent writes is high and growing. A single Claude Code session can rewrite CLAUDE.md, generate a fresh plan.md, drop a multi-page design doc into docs/, and emit a status summary — in one afternoon. By next week, half of those files will be different again.

MD View is the read-side window for that workflow. Native, fast, local, with file-watching that updates the view the moment a file hits disk. This page walks through the patterns that come up repeatedly.

Pattern 1: the always-open CLAUDE.md window

Pin a window in a corner of your monitor with CLAUDE.md open. As Claude Code edits the file during a session, the rendered view updates without you doing anything. You glance over, see what changed, decide whether the agent’s interpretation matches your intent, and steer.

mdview .\CLAUDE.md

Combined with a small window size and the dark theme, the watcher becomes ambient — present, not in the way.

Pattern 2: AGENTS.md as a contract surface

AGENTS.md (or agents.md in some setups) tends to be the canonical instruction file an agent reads on every invocation. When you tweak it, you want immediate feedback that the new content parses cleanly — headings render, code blocks tag right, tables line up. Open it with MD View, edit in your editor, watch the rendered view confirm the structure.

Pattern 3: workspace mode for an agent’s working directory

When an agent is generating multiple .md files in a project — a plan.md, an architecture.md, a series of RFCs — open the whole folder as a workspace:

mdview .

The left rail shows every .md in the tree. Click between them. Sync scroll if comparing two related docs. Workspace mode is a Pro feature; the in-app trial unlocks it for 14 days.

Pattern 4: catch hallucinated paths with verification

Run the verification layer (Ctrl+Shift+V) over an agent-edited doc. Internal links to files that don’t exist flash red. Stale anchors after a heading rename surface in the sidebar. The check is deterministic, instant, offline — exactly the kind of cheap pass that catches the most embarrassing class of agent error.

Pattern 5: the diff overlay for “what did Claude change?”

Press Ctrl+Shift+D on a file the agent just rewrote. The git-diff overlay shows changed paragraphs in context — green for added, orange for modified, red strikethrough for deleted. Read in flow. Catch a sentence the agent broke, accept the rest.

This is faster than a side-by-side diff for prose: you read the new version with the changes highlighted, instead of reading patches.

Pattern 6: pipe to mdview from your shell

When the agent’s output isn’t on disk yet — it’s in the terminal, or in a stream — pipe it:

claude generate plan | mdview
gh pr view 42 --json body --jq .body | mdview

Stdin documents open as ephemeral — no recents entry, no watcher. Close the window when you’re done. See stdin piping in the CLI reference.

Pattern 7: read CLAUDE.md changes across machines

When you sync CLAUDE.md via git between machines, MD View on the receiving end picks up the new version through the file watcher the moment git pull writes it. No reload, no reopen — just the new content showing in the same window.

Why MD View specifically

You can read .md in any text editor, in VS Code’s preview pane, in Typora, in your browser. They all work. MD View’s wedge for this particular workflow is:

Fast cold start. Effectively instant on modern hardware — so pressing Ctrl+O to glance at a file doesn’t break flow.
Reader-only. No editor preview pane to fight with, no source- toggle button to accidentally hit, no plugin alerts. The window exists to render markdown.
Real watcher. No “press F5 to reload” — the view updates the moment the file does.
Native. About 15 MB on disk, no Electron bundle, no second instance of Chrome running just to read a 4 kB markdown file.
Verification layer + diff overlay. Built specifically for the failure modes of agent-edited prose.