Achhina's Digital Garden

2026-09-06 Declaratively Syncing CLAUDE.md to claude.ai Instructions

3 min read

2026-09-06 Declaratively Syncing CLAUDE.md to claude.ai Instructions

What I set out to do

Answer a small question: is there a way to sync the top-level ~/.claude/CLAUDE.md to claude.ai the way skills sync server-side? It turned into shipping a declarative sync.

What I actually did

  • Confirmed there is no CLAUDE.md sync. The only claude.ai to Claude Code channels in the binary are syncClaudeAiSkills/syncClaudeAiPlugins and managed (enterprise) settings; nothing memory- or instructions-shaped.
  • Found the account-wide “Instructions for Claude” field is conversation_preferences on claude.ai/api/account_profile, by reading the web bundles rather than the docs.
  • Chased whether the Claude Code OAuth token could write it. It cannot: api.anthropic.com/api/account_profile returns oauth_token_not_accepted, and /api/oauth/profile plus /api/oauth/account/settings (which the token does read) do not carry the field. I was wrong twice along the way, first “browser-only” then “a cookie will not work from curl”, and testing corrected both.
  • Reconstructed the live Claude Desktop session from disk: decrypt its Chromium cookie jar (Safe Storage key from the keychain, PBKDF2 then AES-128-CBC), and replay the real cf_clearance cookie to clear Cloudflare. No fingerprint spoofing; it is my own session.
  • Captured the exact write shape by intercepting the web UI’s own fetch: PUT {"conversation_preferences": "<text>"}, a single-field partial update. My earlier guessed four-field payload (which would have nulled full_name) was wrong, and capturing is why nothing got clobbered.
  • Shipped the sync: nix/scripts/claude-ai-instructions-sync.py plus a Darwin-gated activation module. Source of truth is the same AGENTS.md that deploys to CLAUDE.md; only the Communication Style and Writing Style sections are pushed (the rest names Claude Code tools that do not exist in a chat). Best-effort, hash-gated, keychain read via the stable /usr/bin/security. just check and hm build green, committed as 3869049.

What was striking

  • The keychain-grant-binds-to-binary constraint (the same one behind 2026-07-25 Obsidian MCP TCC Break from a Homebrew node Upgrade) flipped the mechanism. The obvious launchd agent (the gmailctl pattern) cannot show the approval prompt, and its grant would not survive nix store-path churn, so the sync runs as an interactive hm switch activation step instead. Reading via /usr/bin/security, a path that never moves, is what makes a one-time “Always Allow” stick.
  • Replaying my own cf_clearance got plain curl past Cloudflare cleanly. The edge was not the wall I assumed.
  • The whole thing is a live-session tap, not a durable credential: it depends on Claude Desktop staying logged in. Fine for a switch-time push, and deliberately not a background job.

Next

  • Run hm switch interactively once to grant the keychain and push the first sync.
  • Decide later whether the Knowledge Capture section should sync (the vault MCP connector exists on claude.ai, but it references the /note skill that does not).

Graph View