Import Readwise

Do not proceed until the CLI is installed and authenticated.

Step 1: Ask the user what to import

Suggest importing by topic first — it's the most useful starting point:

• By topic — search for documents and highlights related to a subject (recommended)
• Import specific documents (by URL, title, or search)
• Filter by date range or source type (books, articles, podcasts, tweets)
• Mine their highlights for material relevant to the wiki

Step 2: Search and browse

Use the Readwise CLI to explore the user's library interactively:

# Search for documents by topic
readwise reader-search-documents --query "<topic>" --limit 20 --json

# List recent documents
readwise reader-list-documents --limit 20 --json

# Search highlights by topic
readwise readwise-search-highlights --vector-search-term "<topic>" --limit 30 --json

Show results to the user and let them pick what to import. This is the interactive phase — it's fine to have search results in context here since they're just metadata (titles, authors, snippets).

Note: The CLI --json flag outputs raw JSON arrays, not objects with a results key. Pipe through jq carefully — e.g. jq '.[].title', not jq '.results[].title'.

CLI flag gotcha: reader-get-document-details uses --document-id (NOT --id). See fetch-readwise-document skill for the full flag reference.

Step 2.5: Update home immediately

Once you know what sources were found, update wiki/home.md right away — before fetching or ingesting anything. Write a brief overview of what's coming: the topics found, how many sources, what the wiki will cover. This gives the user something to read and shows progress while the import runs.

Step 3: Fetch into raw/

Import in small batches. Fetch and fully ingest 3-5 sources first so the user can see the wiki taking shape before importing more. A wiki with a few well-connected pages is more useful than a queue of unprocessed raws. After the first batch is ingested and the user can browse it, ask if they want to continue with more.

Once the user has picked what to import, delegate to the appropriate skill:

• For documents: invoke fetch-readwise-document with the selected doc IDs. It handles metadata, streaming the body to disk, and verification.
• For highlights: invoke fetch-readwise-highlights with the agreed-upon search queries. It handles vector search, deduplication, grouping, and writing highlight files.

Both skills chain into ingest automatically to create wiki pages from the raws.

Important: All files in raw/ must be markdown (.md), never JSON. Temp JSON files from CLI queries go in /tmp/, not raw/. If you need to store structured data from Readwise, convert it to a readable markdown document before saving to raw/.

Step 4: Parallel ingest with subagents

This is the most important performance step. After fetching raw files, do NOT ingest them one at a time. Use the Agent tool to parallelize:

1. Read the current wiki state yourself first — read wiki/index.md and wiki/home.md to understand what exists.
2. Dispatch one subagent per source (or per 2-3 related sources) using the Agent tool. Launch them all in a single message so they run concurrently. Each subagent brief should include:
   • The path to the raw file(s) to ingest
   • The current wiki index (so it knows what pages exist)
   • The CLAUDE.md schema reference
   • Instructions to: read the raw, create source-summary page, create/update concept pages, cross-link aggressively, update index.md and log.md
3. After all subagents complete, do a single pass yourself to:
   • Deduplicate any index.md entries (multiple agents may have added overlapping entries)
   • Update wiki/home.md with the full picture
   • Check for cross-linking gaps between the new pages

Example dispatch pattern:

Agent({
  prompt: "Ingest raw/source-a.md into this wiki. Schema is in CLAUDE.md. Current index: [paste index]. Create source-summary at wiki/sources/, propagate claims to concept pages, cross-link from 2-3 existing pages, update index.md and log.md.",
  description: "Ingest source-a"
})
Agent({
  prompt: "Ingest raw/source-b.md into this wiki. Schema is in CLAUDE.md. Current index: [paste index]. Create source-summary at wiki/sources/, propagate claims to concept pages, cross-link from 2-3 existing pages, update index.md and log.md.",
  description: "Ingest source-b"
})
// ... all in the same message for parallel execution

Why this matters: Serial ingestion of 5 sources takes 5x as long. Parallel subagents cut wall-clock time dramatically. The dedup pass at the end is cheap.

Step 5: Update wiki infrastructure

After all subagents complete:

1. Deduplicate wiki/index.md entries (multiple subagents may have added the same pages).
2. Update wiki/home.md to reflect everything that was imported.
3. Verify wiki/log.md has timestamped entries.
4. Scan for cross-linking gaps — pages created by different subagents may not link to each other yet.

Report what was imported and what pages were created.