readwise-ai · v1.2

Fetches articles from your Readwise Reader library, generates an AI-powered daily briefing shaped by your personal taste profile, and saves it back to Readwise as an article. Optionally generates a short watch summary (for an Apple Watch companion app).

How it works

1. Fetches articles from Readwise Reader (feed, new, or both)
2. Filters and prioritises them by your tag preferences
3. Sends them to an OpenAI model with your taste profile as context
4. Saves the generated briefing back to Readwise as a new article

The daily briefing is a longer editorial summary. The watch summary is a short (~200 word) version of the last few hours, designed for quick glancing.

Requirements

• Python 3.11+
• A paid Readwise subscription with Reader access (Ghostreader custom prompts require a paid plan)
• An OpenAI API key (see below)

OpenAI API key and costs

This tool uses the OpenAI API directly with your own key — it does not use the Readwise-provided AI model. This means you pay OpenAI separately for every run.

Getting an API key

1. Create an account at platform.openai.com
2. Go to API keys and create a new secret key
3. Add billing details and set a monthly spend limit (recommended — easy to forget it's running)
4. Paste the key into your .env file as OPENAI_API_KEY

Token usage estimates

The tool is prompt-heavy. Both your taste profile and the full article list are sent on every call. Rough per-run estimates with default settings (max_articles = 100, watch every 3 hours):

Run type	Typical input	Output	Total
Daily briefing (×1/day)	12,000–30,000 tokens	~2,000 tokens	~14,000–32,000
Watch summary (×8/day)	5,000–8,000 tokens	~400 tokens	~5,500–8,500
Daily total			~60,000–100,000 tokens
Monthly total			~1.8M–3M tokens

Input tokens dominate — the article summaries and your taste profile are the main driver. Reducing max_articles in config.toml is the most effective way to cut costs.

Model choice: quality vs. cost

The model is set in config.toml under [openai] model. This is the single biggest lever on both output quality and cost.

Model tier	Quality	Relative cost	Notes
GPT-5 mini variants	Good	Low	Fine for the watch summary; briefing quality noticeably lower
GPT-5 standard	Very good	Medium	Solid default for most users
GPT-5 premium variants	Excellent	High	Meaningfully better editorial quality and cross-topic synthesis; cost scales sharply

Check OpenAI's pricing page for current model names and rates — the lineup evolves quickly. Set a monthly spend cap in your OpenAI account settings if you run this on a schedule.

Tip: You can use different models for different runs. The watch summary is short and tolerates a cheaper model well. Reserve a premium model for the daily briefing if cost is a concern.

Installation

git clone https://github.com/samikki/readwise-ai.git
cd readwise-ai

python -m venv venv
source venv/bin/activate
pip install -r requirements.txt

Configure secrets

cp .env.example .env

Edit .env and fill in your tokens:

READWISE_TOKEN=your_readwise_token_here
OPENAI_API_KEY=your_openai_key_here

Your Readwise token is at readwise.io/access_token.

Configure settings

cp config.example.toml config.toml

Edit config.toml to set your name, preferred language, OpenAI model, tags to prioritise/ignore, and how many days of articles to fetch.

Set up your taste profile

cp taste_profile.example.md taste_profile.md
cp local_profile.example.md local_profile.md

Edit these files to describe your reading interests and priorities. The taste profile is injected into the prompt — the more specific you are, the better the briefing.

Set up Ghostreader in Readwise

This tool feeds article summaries and tags from your Readwise library directly into the LLM prompt. If those fields are empty, the briefing LLM has nothing but a title to work from — quality suffers significantly.

The recommended way to populate them automatically is Readwise Reader's Ghostreader feature, which runs a custom AI prompt on every article you save. To configure it: open read.readwise.io, go to Settings → Ghostreader, and set up custom prompts for Document summary and Auto-tag.

You need two prompts:

---

1. Document summary prompt — runs on every new article and populates the summary field.

Ghostreader prompts use Jinja2-style template variables. A starting point:

You are a concise summariser. Your goal is to give me the key information from this article in as few words as possible.

Rules:
- First sentence: the single most important takeaway, or a direct answer if the title is a question or teaser.
- Follow with 2–3 sentences covering other essential points.
- No preamble. Never start with "This article" or "The author". No filler.
- If the article is behind a paywall or the content is missing, start your summary with [P].
- Write in present tense. Aim for 100–150 words maximum.

Title: {{ document.title }}
Author: {{ document.author }}

{% if (document.content | count_tokens) > 6000 %}
{{ document.content | central_sentences | join('\n\n') }}
{% else %}
{{ document.content }}
{% endif %}

Adapt the output language to match your preference (e.g., add "Always write in Finnish." if that's your reading language).

---

2. Auto-tag prompt — runs on every new article and assigns it a topic tag.

Design a flat taxonomy that matches your interests. Keep categories at the same level of specificity — mixing broad and narrow (e.g., "Technology" alongside "AI") causes the model to default to the broader one. A starting point:

Categorise this article into exactly one of the following topics:

AI: Artificial intelligence, machine learning, LLMs, and AI tools and research.
Technology: Software, hardware, cybersecurity, programming, and tech industry news not primarily about AI.
Science: Physics, biology, mathematics, astronomy, medicine, and scientific research.
Business: Companies, economics, investing, markets, and professional strategy.
Politics & Society: Government, policy, geopolitics, and social issues.
Health: Physical and mental health, nutrition, fitness, and medicine.
Arts & Culture: Literature, music, film, art, and cultural commentary.
Productivity: Personal organisation, note-taking, time management, and self-improvement.
Entertainment: Humour, games, TV, sports, and light content.
Other: Anything that does not fit the above categories.

Return only the category name. No explanation, no punctuation.

Title: {{ document.title }}
Author: {{ document.author }}

{% if (document.content | count_tokens) > 2000 %}
{{ document.content | central_sentences | join('\n\n') }}
{% else %}
{{ document.content }}
{% endif %}

Customise the taxonomy to your own interests — the example above is generic. Once you have your tags defined, update IGNORE_TAGS in config.toml to filter out any categories you never want to appear in your briefing.

Usage

Daily briefing

# Generate and post to Readwise
python build.py

# Last 3 days from "new" inbox only
python build.py --days 3 --sources new

# Dry run — print output, don't post
python build.py --dry-run

# Also save to a local file
python build.py --output summary.html

Watch summary (short, last few hours)

python build.py --watch

Options

Flag	Default	Description
--days N	1	How many days back to fetch articles
--sources	feed later	Readwise Reader locations: new, later, shortlist, archive, feed
--max-articles N	100	Max articles to include in the prompt
--dry-run	off	Print output without posting to Readwise
--output FILE	—	Also save HTML to a local file
--watch	off	Generate short watch summary instead

Automating with cron

crontab -e

Add entries like:

# Daily briefing at 8 AM
0 8 * * * /path/to/readwise-ai/run_build.sh >> /path/to/readwise-ai/logs/build-$(date +\%Y-\%m-\%d).log 2>&1

# Watch summary every 3 hours
0 */3 * * * /path/to/readwise-ai/run_watch.sh >> /path/to/readwise-ai/logs/watch-$(date +\%Y-\%m-\%d).log 2>&1

Make sure the scripts are executable:

chmod +x run_build.sh run_watch.sh

File layout

File	In git	Purpose
.env	No	Your API tokens — never committed
.env.example	Yes	Template for new users
config.toml	No	Your settings — never committed
config.example.toml	Yes	Template with neutral defaults
taste_profile.md	No	Your personal taste profile — never committed
taste_profile.example.md	Yes	Template and instructions
local_profile.md	No	Optional local/contextual priorities — never committed
local_profile.example.md	Yes	Template and instructions
templates/prompt_template.md	Yes	Prompt template for the daily briefing
templates/watch_prompt_template.md	Yes	Prompt template for the watch summary
build.py	Yes	Main entry point
run_build.sh	Yes	Shell wrapper for cron
run_watch.sh	Yes	Shell wrapper for watch summary cron

Watch app

There is a companion Apple Watch app in readwise-ai-watchapp that fetches and displays the latest watch summary.