claudeai-sync — pull what's new from claude.ai into the vault

Account owner.work@example.com (Claude Max). claude.ai chats do NOT live on disk — they are pulled live through the logged-in session (Claude-in-Chrome). Semi-automatic: PULL (step 1) is done by me while the session is open; everything after that is deterministic and idempotent.

Scripts: $IMPORTS_ROOT/claude-ai/ · originals → $OBSIDIAN_ROOT/_originals/claude-ai-export/ · live layer → 01-Conversations/Claude-AI/.

⚠️ The main landmine

The API exposes artifacts ONLY with rendering_mode=messages. raw/default silently return 0 artifacts. Always use messages + verify the artifact counter.

⚠️ Anti-recents (hunting for one specific chat by hand)

When looking for ONE specific chat by hand (not a full PULL) — NEVER conclude "the chat isn't there" from a glance at the recents list: (1) use the built-in "Search chats…" with keywords; (2) check Projects/archive/pinned; (3) confirm the active ACCOUNT by the email in the profile menu. More reliable still — the full API listing chat_conversations?limit=1000 (the PULL step) instead of reading /recents with your eyes. Canon: memory [[web-ui-search-not-recents]] (incident 2026-07-23).

Steps

1. PULL (in-session, ~30 sec)

• Claude-in-Chrome → navigate https://claude.ai/recents.
• javascript_tool (same-origin authed fetch), org = <YOUR_ORG_UUID> (roles chat+claude_max; a second org on the account is API-only, returns 403, ignore it):
  • the list: /api/organizations/{ORG}/chat_conversations?limit=1000&offset=0
  • each chat: /api/organizations/{ORG}/chat_conversations/{uuid}?tree=True&rendering_mode=messages → cache into window.__convs
  • projects: /api/organizations/{ORG}/projects + /{uuid} (descr + prompt_template) + /{uuid}/docs (content inline)
  • Blob-download it as ONE file claude-ai-export-YYYY-MM-DD.json (schema claude-ai-export/v1: {conversations, projects, filesManifest}).
    • ⚠️ conversations MUST be an OBJECT {uuid: conv}, NOT a list — the converter claudeai_export_to_vault.py:120 calls convs.items() (a list → AttributeError: 'list' object has no attribute 'items'). filesManifest is a list of {conv, msg, kind, meta}. Under rendering_mode=messages artifacts arrive as <antArtifact …> INSIDE text items (NOT as tool_use blocks) — that is normal, the converter parses them.
• Move it out of Downloads into raw\ + archive into _originals\claude-ai-export\ (sha256). The full JS lives in the session journal of 2026-06-12 / memory [[claude-ai-export-to-vault]].
• (A delta pull is optional — the converter is idempotent by uuid; a full pull only adds what's new.)

2. SYNC (deterministic)

python $IMPORTS_ROOT/claude-ai/claudeai_sync.py

Converts into a temporary staging area → copies into the live layer only the NEW notes (existing ones, including linked ones, are left untouched) → refreshes the MOC + dashboard → writes the paths of new artifacts into _new_artifacts.txt. Prints new_conversations / new_artifacts / new_projects.

3. CONCEPT-LINK the new artifacts (MANDATORY — concept-creation-rules.md §1)

If _new_artifacts.txt is not empty:

• Workflow claudeai-artifact-curation (sonnet, one agent per artifact; ⚠️ above ~40 you hit a server-side rate limit, finish with resume) → concept proposals.
• python apply_curation.py --result <workflow .output> — validates slugs against the real 06-Concepts/09-Bridges, writes links idempotently (a <!-- curation --> block + related_concepts/summary/value_score/tags).
• Recurring themes with no concept → create new concepts (threshold §1 ≥3), show them to the operator, then re-run apply (the new ones get picked up). One-offs — don't breed concepts for those.

4. RAG + commit

python $IMPORTS_ROOT/brain_embed_update.py        # or wait for the nightly run at 04:00
python $IMPORTS_ROOT/vault_backup.py

(claudeai_sync.py --reindex --commit does step 4 in one call; reindex rc=3 = the lock is busy, the nightly task will catch up.)

5. PING

A short summary to the operator (Telegram Saved Messages works fine): "claude.ai: +N chats, +M artifacts, linked."

Notes

• Idempotent: a repeat run with no new data = "nothing new", and existing links are never lost.
• Provenance: artifacts get origin: claude-ai, authored_by: claude (NOT #anton-original); project instructions get origin: anton.
• Binaries (images/docs, 524 of them in the manifest) — the operator asked NOT to pull those (2026-06-12).
• Auto mode: the nightly Windows task Claude-AI Sync Daily runs the back half over any fresh export; the PULL stays in-session. Full headless ("it downloads itself") = moving to a dedicated Chrome profile + Playwright + a deterministic RAG linker — deferred (the operator chose the safe semi-automatic path).

Like this skill? It is one of 100 in second-brain-starter-kit: the second brain we built for ourselves and run every day at Palo Alto AI Research Lab. Install the whole set with npx skills add tonydzi/second-brain-starter-kit. Everything is open source and free, so take what you need.

Flagships worth a look on their own: secondop-panel (a second opinion from a panel of external models), claude-memory-tidy (stop your agent's memory from rotting), telegram-mcp-kit (your own Telegram over MCP in about 15 minutes).

Author: Anton Dziatkovskii, Palo Alto AI Research Lab. Telegram @tonydzi - WhatsApp +1 341 222 9178 - X @Tony_Stef_

Engineers: want to test-drive this setup? Message me. I hand out free starter seeds to engineers who test and report back, and custom skill requests are welcome.