You are building a reader persona for the user based on their Readwise Reader library. This persona file is used by other skills (triage, quiz, etc.) to personalize their experience.
Readwise Access
Check if Readwise MCP tools are available (e.g.
mcp__readwise__reader_list_documents
). If they are, use them throughout (and pass this context to the subagent). If not, use the equivalent
readwise
CLI commands instead (e.g.
readwise list
,
readwise read
,
readwise search
,
readwise highlights
). The instructions below reference MCP tool names — translate to CLI equivalents as needed.
Welcome
Open with a brief introduction:
Build Persona
· Readwise Reader
I'll analyze your reading history — saves, highlights, and tags — and build a
reader_persona.md
profile in the current directory. Other skills (triage, quiz) will use this to personalize their output to you.
I'll start with a quick pass (~1-2 min) and then you can decide if you want a deeper analysis.
Process
IMPORTANT:
This skill involves fetching a lot of data. To keep the main conversation context clean, launch a
Task subagent
to do all the heavy lifting.
Phase 1: Quick Pass
The subagent should do a focused scan to build a solid initial persona fast:
Gather data.
Run ALL of these in parallel (one batch of tool calls):
4 highlight searches:
mcp__readwise__readwise_search_highlights
with 4 broad queries (e.g. "ideas strategy product", "learning technology culture", "writing craft creativity", "business leadership growth") with
limit=50
each. These are semantic/vector searches so broad multi-word queries work well. Highlights are cheap and high-signal — cast a wide net.
4 document lists:
mcp__readwise__reader_list_documents
from each non-feed location:
location="new"
,
location="later"
,
location="shortlist"
, and
location="archive"
with
limit=100
each. If the combined results are very sparse (< 20 docs total), also try without a location filter or with
location="feed"
as a fallback. Only fetch metadata:
response_fields=["title", "author", "category", "tags", "site_name", "summary", "saved_at", "published_date"]
. Do NOT fetch full content.
Tags:
mcp__readwise__reader_list_tags
to understand their organizational system.
Parse results efficiently.
The JSON responses from document lists can be large (25k+ tokens). Do NOT try to read them with the Read tool — it will hit token limits and waste retries. Instead, use a single Bash call with a python3 script to extract and summarize all the data at once. The script should parse all result files together and output:
Document counts by category
Top 20 sites, authors, and tags
Save velocity by month
All docs saved in the last 3 weeks (title, category, author, date)
A representative sample of highlight texts with their source titles/authors
Write the persona.
Write
reader_persona.md
to the current working directory with these sections:
Identity & Role
— Who they appear to be (profession, role, industry)
Core Interests
— Top themes and topics, ranked by frequency and recency
Reading Personality
— How they read (saves a lot but reads selectively? highlights heavily? prefers short or long-form?)
Current Obsessions
— What they've been saving/reading most in the last 2-3 weeks
Goals & Aspirations
— What they seem to be working toward, inferred from patterns
Taste & Sensibility
— Thinkers and styles they gravitate toward (contrarian? practical? philosophical? technical?)
Anti-interests
— Topics notably absent or avoided
Triage Guidance
— Specific instructions for how to pitch documents to this person (e.g. "lead with practical applicability", "connect to their interest in X", "bar is high for AI content — flag when it's genuinely novel")
Return
a brief summary (3-5 sentences) of the persona AND the absolute path to the file.
Subagent speed rules:
Do NOT call
readwise_list_highlights
— it often errors and is redundant with search.
Do NOT try to Read large JSON tool-result files — parse them with python3 via Bash.
Combine all analysis into ONE python script, not multiple sequential scripts.
Maximize parallel tool calls. Every API fetch in step 1 should be a single parallel batch.
Phase 2: Deep Pass (optional)
After the quick-pass subagent returns, show the user the results and ask if they want a deeper analysis. If yes, launch a second subagent that:
Fetches 4-6 more highlight searches with
different, more specific
queries informed by what phase 1 found (e.g. if the persona shows interest in AI tooling, search "AI agents workflows automation"; if they read fiction, search "fiction narrative storytelling") with
limit=50
each
Paginates beyond the first 100 docs per location using
next_page_cursor
from phase 1 results — fetch the next 100-200 per location to build a much larger sample
Reads the existing
reader_persona.md
and enriches/rewrites it with the additional data — more nuanced sections, stronger evidence, sharper triage guidance
Returns a summary of what changed
After Each Subagent Returns
Show the file link.
Always tell the user:
reader_persona.md
was written to
{absolute_path}
. Display the full path so they can open it.
Show a summary
of the persona (use the subagent's returned summary).
After phase 1:
Ask if they want the deep pass
or if the quick version is good enough. Also ask if they want to adjust anything.
After phase 2 (if run):
Show what changed
and ask if they want to adjust anything.
If adjustments needed,
edit the file directly based on their feedback.
Confirm saved.
Tell them the file is saved and which skills will now use it (triage, quiz, feed-catchup, etc.).