version: "1.0.0" name: fetch-readwise-document description: Fetch one or more Readwise Reader documents into raw/ without loading bodies into context. Streams content to disk via jq pipe, then chains into ingest. allowed-tools: Bash(*) Read Write Edit Glob Grep

Fetch Readwise Document

Goal: grab a Reader document (or several) and drop it into raw/ without ever loading the full body into context. Only small metadata (title, author, url, date, category, doc_id) is allowed in context. The body gets streamed from CLI to file via a pipe.

Preconditions

• readwise CLI installed and authenticated.
• jq installed (brew install jq if missing).
• A raw/ directory exists.

Core pattern (single doc)

Given a Readwise URL (https://read.readwise.io/read/<id>), a bare doc_id, or a search query:

Step 1 — Resolve doc_id

If the user gave a URL: the id is the last path segment. No CLI call needed.

If the user gave a specific title, use --title-search (note: --query is also required even when filtering by title):

readwise reader-search-documents --query "<title words>" --title-search "<title words>" --limit 5 --json \
  | jq -r '.[] | "\(.document_id)\t\(.title)\t\(.author)\t\(.category)"'

For a topical search (no exact title), use --query alone. Add --author-search when you know the author. Show candidates to the user if ambiguous.

Step 2 — Fetch metadata (mandatory)

reader-get-document-details does not return image_url, source_url, published_date, word_count, or site_name. Pull those from reader-list-documents:

readwise reader-list-documents --id <DOC_ID> \
  --response-fields title,author,url,source_url,category,published_date,saved_at,site_name,word_count,image_url \
  --json | jq '.results[0]'

• image_url — cover/header image. Embed as ![](url) in the raw header.
• source_url — the original URL (not the read.readwise.io shell). Use as the canonical **Source:**.

If image_url is null, skip the markdown image embed. Don't fail the fetch.

Pick a filename slug: <author-last-or-source>_<short-title-slug>.md, lowercase, hyphen-separated, no punctuation, max 60 chars.

Step 3 — Stream the body to disk

This is the critical command. Never run reader-get-document-details without piping into jq and redirecting to a file.

{
  printf '# %s\n\n![](%s)\n\n**Source:** %s\n**Readwise URL:** https://read.readwise.io/read/%s\n**Readwise ID:** %s\n**Date:** %s\n**Author:** %s\n**Category:** %s\n**Cover image:** %s\n\n---\n\n' \
    "<TITLE>" "<IMAGE_URL>" "<SOURCE_URL>" "<DOC_ID>" "<DOC_ID>" "<DATE>" "<AUTHOR>" "<CATEGORY>" "<IMAGE_URL>"
  readwise reader-get-document-details --document-id <DOC_ID> --json \
    | jq -r '.content'
} > raw/<slug>.md

Drop the ![](%s)\n\n line if image_url is null.

Step 4 — Verify without reading the body

wc -l raw/<slug>.md && head -n 10 raw/<slug>.md

head -n 10 only shows the header you wrote. If line count is 0, something went wrong.

Step 5 — Report, then chain into ingest

Tell the user: filename, word count (from metadata), and that the body is on disk. Do not summarize the content — you haven't read it. Then invoke the ingest skill on the raw file.

Multi-doc pattern

Resolve all doc_ids first (Step 1), fetch metadata for all (Step 2), then loop:

for id in <ID1> <ID2> <ID3>; do
  slug=$(...)  # derived per-id from metadata
  {
    printf '...header...'
    readwise reader-get-document-details --document-id "$id" --json | jq -r '.content'
  } > "raw/$slug.md"
done
wc -l raw/*.md

Hold off on ingest until all fetches are done, then ingest the batch using parallel subagents (see import-readwise skill Step 4).

JSON shapes (don't re-probe these)

• reader-search-documents --json → top-level array. Each item: document_id, title, author, category, url, matches[].
• reader-list-documents --json → `{count, nextPageCursor, results: [...]}`. Access with jq -r '.results[0] | ...'.
• reader-get-document-details --json → flat object with keys id, title, author, category, tags, notes, content. The body is at .content. No image_url, source_url, published_date.

Tweet caveat

When a user saves a tweet that is a reply, Reader stores the parent thread as the document. The source_url (from list-documents) points at the actually-saved tweet. The image_url is the parent author's avatar. Surface this to the user when fetching tweet replies.

CLI flag reference (don't guess these)

The Readwise CLI uses --document-id, NOT --id. Here are the exact flags:

# Get document details — use --document-id (NOT --id)
readwise reader-get-document-details --document-id <DOC_ID> --json
 
# List documents — use --id to filter by doc ID
readwise reader-list-documents --id <DOC_ID> --json
 
# Search documents
readwise reader-search-documents --query "<text>" --json
readwise reader-search-documents --query "<text>" --title-search "<title>" --json
 
# Search highlights
readwise readwise-search-highlights --vector-search-term "<text>" --limit 30 --json

Common mistakes to avoid:

• reader-get-document-details --id → WRONG, use --document-id
• reader-list-documents --document-id → WRONG, use --id

Rules

• Never run reader-get-document-details without | jq -r '.content' > <file>. No exceptions.
• Never Read or cat a raw/ file you just wrote unless the user explicitly asks.
• Never probe JSON shapes with jq 'keys'.
• Prefer the flag reference above over --help — but --help is fine as a fallback.
• Prefer `--title-search` over --query when the user names a specific title.
• If metadata is missing fields, use null or unknown — do not fetch the body to find them.
• Confirm with the user before overwriting an existing raw/ file.