Ghostreader Has Earned a Place in My PKM Workflow

Readwise Reader (a read it later app for power users) has been actively iterating on their AI integration into their ecosystem. One of the tools I have made the most use of so far is Ghostreader.

I've been using a custom Ghostreader prompt to generate executive summaries that not only capture the overall point of the article, but also bring my highlights and annotations into context.

The act of creating this summary is something I think is the perfect candidate to offload to AI. I've already done the reading, and I've even done the thinking (by highlighting and annotating as I went), now I just want to roll that up into a concise overview.

Exploring the Intersection of AI and PKMs

I am excited about the power of using AI with an established PKM system. So I asked ChatGPT to interview me about it.

Next TopicBailey Belvis

The workflow

1. Read an article
2. As your read, highlight and annotate
3. When completely done, trigger the custom Ghostreader Prompt to generate a document summary

The Prompt

Here is the custom Ghostreader prompt I'm currently using to generate my document summaries:

{#- BACKGROUND: This prompt instructs ChatGPT to use your notes and highlights to summarize the most important points or any action items according to some input from you to guide the large language model. It's intended to be used after you're done reading. The primary purpose of this prompt is to generate content for my "one-pager" in Obsidian.   It attempts to capture summaries and my personal takeaways using my highlights and notes. -#}
{% set intent = input("Provide some context on how this document relates to your work/life") %}
I just read "{{ document.title }}" by {{ document.author }} and I want you to summarize some of my notes and highlights around my hastily written intent: {{ intent }}

Here are my highlights. Sometimes I also added tags and notes to the highlights.
{% for highlight in highlights -%}
    ===
	{{ highlight }}
    {% if highlight.note or highlight.tags %}
	note: {{ highlight.note }}
	tags: {{ highlight.tags | join(", ") }}
    {% endif %}
    ===
    
{%- endfor %}

Here is the document I read:

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

For each primary section generated, use the markdown h3 header.

First, rewrite my shorthand intent ({{ intent }}) to be clearer. Follow this markdown format:

### One-Liner
#ai/generated
<content goes here>


Next, include a header for Summary: followed by: {{document.summary }}

Next, summarize the top 3-5 most important takeaways or to dos that stood out to me according to my intent, highlights, and notes. Each of these takeaways should start with a unique and relevant emoji.  They should be ordered from most impactful or actionable to least.  Follow this markdown format:

### Key Takeaways
#ai/generated
<emoji> **<Takeaway short desc>** - <elaborate on takeaway>

<emoji> **<Takeaway short desc>** - <elaborate on takeaway>

<emoji> **<Takeaway short desc>** - <elaborate on takeaway>


Next, based on my highlights, notes, and related context from the document, write a summary of my personal thoughts and takeaways. Follow this markdown format:

### Personal Thoughts and Takeaways
#ai/generated
<content goes here> 


Finally, offer 2-3 counter-opinions, challenging thoughts, or questions worth investigating further.  Follow this markdown format:

### Counter-opinions and Questions worth investigating
#ai/generated
1. <content goes here>
2. <content goes here>
3. <content goes here>

Example Result

Here is an example note generated by Ghostreader for an article I highlighted and annotated called How to Provide Feedback on Documents.

### One-Liner
#ai/generated
Strategies for providing effective feedback on documents to avoid creating an endless feedback loop.

### Summary
To provide effective feedback on documents, focus on helping the author rather than prioritizing your own goals. Start by skimming the document to understand its structure, then leave clear comments that explain your concerns and their importance. If you have many comments, consider discussing them in person to streamline the feedback process.

### Key Takeaways
#ai/generated
📚 **Understand Document Structure First** - Skim the document initially to grasp its structure and avoid irrelevant feedback related to document layout or content that is addressed later.
🔍 **Leave Constructive Comments** - When providing feedback, ensure each comment includes the suggested change, its significance, and its urgency, helping the author prioritize adjustments.
🤝 **Limit Feedback Quantity** - If you find yourself leaving numerous comments, either increase your threshold for feedback or arrange a discussion with the author to consolidate your insights, preventing overwhelming the author.

### Personal Thoughts and Takeaways
#ai/generated
The approach outlined in the document resonates with my experiences in providing feedback. Prioritizing the author's needs and minimizing unnecessary comments can foster a more collaborative environment. I've noticed that when feedback becomes overwhelming, it can discourage authors from sharing their work, which ultimately hampers team communication. The emphasis on clarity and structured feedback aligns with my belief that effective communication is key to successful documentation processes.

### Counter-opinions and Questions worth investigating
#ai/generated
1. Is it always feasible to prioritize the author's needs over team or strategic goals, especially when critical decisions depend on document feedback?
2. How can we balance the need for thorough feedback with the risk of overwhelming the author, particularly in high-stakes projects?
3. What systems or tools could be implemented to better support the feedback process, especially to manage the flow of comments and promote collaboration?

Why I like this

One reason I'm on board with allowing AI to enter my second brain space in this way is because this content is purely additive. My original highlights and annotations are still synced to Obsidian, but this summary provides a nice, concise overview of what I've already captured.

With this executive summary a couple interesting future workflows are unlocked.

For instance, this context could be fed back into an LLM to help me identify relevant source material to a given topic, or help me identify connections between my source materials.

Additionally, since LLM's have a frustratingly annoying habit of being overly positive and affirming, I really like that I can get counter opinions and questions added. These are like instant brainstorming fuel every time I revisit one of these source notes.

Things the summary misses

I'll be completely transparent in saying that the generated output can be hit or miss. It's never outright bad, but it may miss key points that you would have preferred made it into the summary.

This is one of many reasons why I don't like purely relying on AI summaries of any kind of content. In this case, its acceptable to me because the original material is still available right along side the summary.

What's next

The Reader team has announced they are exploring expansions to their AI integrations in Reader, so I'm excited to see what they come up with.