Building a Personal Relevance Filter

Every platform you use has an opinion about what you should see. Twitter thinks you want engagement. YouTube thinks you want watch time. Google thinks you want clicks. LinkedIn thinks you want — actually, I have never been able to figure out what LinkedIn thinks, but whatever it is, it is not what I want. These are relevance filters, and they are designed to serve the platform’s definition of relevance, not yours.

A personal relevance filter is different. It is a system — part human judgment, part automation, part AI — that filters incoming information according to criteria you have defined explicitly. Not what an algorithm guesses you want. Not what generates the most engagement. What you have deliberately decided is worth your limited attention, based on your actual goals, responsibilities, and interests.

This is harder than it sounds. Defining what is relevant to you requires knowing what your priorities are, which requires the kind of self-knowledge that most of us have in much lower quantities than we think. But the exercise of defining it is valuable even when the definition is imperfect, because an imperfect explicit filter is dramatically better than an implicit one you never chose.

Step 1: Define Your Information Priorities

Before you can filter, you need criteria. Before you can build criteria, you need to answer some uncomfortable questions about what actually matters in your professional and intellectual life.

I use a framework with three tiers:

Tier 1: Mission-Critical

These are topics where being uninformed creates immediate, concrete risk. For a software engineer, this might be security vulnerabilities in your stack. For a portfolio manager, market developments in your sectors. For a policy analyst, legislative changes in your domain. For a product manager, changes in your competitive landscape.

The defining characteristic of Tier 1: if you miss something here, something bad happens. Not “you feel out of the loop.” Something actually goes wrong.

Most people have 2-4 Tier 1 topics. If you have more than 5, you are either in an unusually broad role or you are confusing “would be nice to know” with “must know.”

Tier 2: Professional Development

These are topics that make you better at your job over time, but where missing any individual piece of information has no immediate consequence. Research in your field. Best practices in your discipline. Trends that will affect your work in the next 1-3 years. Adjacent areas that inform your primary work.

The defining characteristic of Tier 2: this information has a long shelf life and cumulative value, but low urgency on any given day.

Most people have 4-8 Tier 2 topics. These are where most of your “defer” reading should come from.

Tier 3: Intellectual Curiosity

These are topics you follow because they interest you, broaden your perspective, or contribute to your ability to think well about the world. They are not directly related to your current work but contribute to the well-rounded understanding that makes you a more creative and effective thinker.

The defining characteristic of Tier 3: entirely optional, but life is impoverished without it. Do not eliminate this tier in the name of productivity. Curiosity is a resource, and starving it makes you worse at everything else.

Most people have 3-6 Tier 3 topics, though the number is less important than the honesty. If your Tier 3 list includes things you feel you “should” be interested in rather than things you actually are, drop them.

The Priority Map

Here is a concrete example. This is roughly my own priority map, simplified:

Tier 1 — Mission-Critical:

• AI/ML development and capabilities (my field)
• Security and privacy developments in AI systems
• Regulatory changes affecting AI deployment

Tier 2 — Professional Development:

• Software architecture and systems design
• Technical leadership and engineering management
• Research methodology and epistemics
• Cognitive science and decision-making research

Tier 3 — Intellectual Curiosity:

• History of technology and media
• Urban planning and transportation
• Climate and energy systems
• Literature and long-form journalism

Your map will look different. The structure should be the same: a small number of high-priority topics, a moderate number of development topics, and a handful of curiosity topics. Write it down. Actually write it down, in a place you will see it regularly. This map is the foundation of everything else in this chapter.

Step 2: Create a Scoring Rubric

With your priority map defined, you can create a simple scoring rubric for incoming information. This does not need to be quantitative — nobody is actually going to assign numerical scores to each email. But having a rubric gives you a fast, repeatable way to sort information.

Here is a rubric that works:

Relevance Score:

• Directly relevant: Addresses a Tier 1 topic and contains actionable information
• Moderately relevant: Addresses a Tier 1-2 topic with useful but non-urgent content
• Tangentially relevant: Touches on a Tier 2-3 topic or is adjacent to your priorities
• Not relevant: Does not connect to any of your defined priorities

Quality Score:

• High quality: From a known, credible source; based on evidence or primary reporting; adds new information or perspective
• Medium quality: From a generally reliable source; synthesizes existing information well; adds some value
• Low quality: From an unknown or unreliable source; largely derivative; mostly opinion without evidence

Timeliness Score:

• Time-sensitive: Value decreases significantly within 24-48 hours
• Current: Best consumed this week, but no urgency
• Evergreen: Will be equally valuable whenever you get to it

The combination of these three scores tells you how to handle an item:

This table is simplified, but the principle holds: you are cross-referencing what it is about, how good it is, and how time-sensitive it is. The combination determines your action, not any single dimension.

Step 3: Use LLMs to Pre-Screen and Categorize

Now for the automation. You can use LLMs to apply your scoring rubric to incoming information, dramatically reducing the time you spend in the Preview stage of your triage pipeline (Chapter 9).

Creating a Personal Relevance Assistant

The key tool here is the system prompt — a set of standing instructions that configure the LLM’s behavior for a specific purpose. Here is how to create one for information filtering:

System Prompt:

You are my personal information filter. Your job is to assess incoming
content against my priority map and help me decide how to process it.

MY PRIORITY MAP:
Tier 1 (Mission-Critical):
- [Your Tier 1 topics]

Tier 2 (Professional Development):
- [Your Tier 2 topics]

Tier 3 (Intellectual Curiosity):
- [Your Tier 3 topics]

SCORING RUBRIC:
For each piece of content I share with you, provide:

1. RELEVANCE: [Directly relevant / Moderately relevant / Tangentially
   relevant / Not relevant] — and which specific priority it connects to
2. QUALITY INDICATORS: Note the source, author if known, and any
   indicators of content quality
3. TIMELINESS: [Time-sensitive / Current / Evergreen]
4. RECOMMENDED ACTION: [Process now / Defer / Skim / Ignore]
5. ONE-LINE SUMMARY: What is this about, in one sentence?

IMPORTANT GUIDELINES:
- Err on the side of filtering OUT rather than IN. My time is limited
  and I would rather miss something tangentially relevant than be
  overwhelmed with marginal content.
- If content is relevant to multiple priorities, note all of them.
- If you are uncertain about relevance, say so rather than guessing.
- If content seems designed to provoke engagement rather than inform
  (clickbait, rage-bait, controversy for controversy's sake), flag it
  and recommend Ignore regardless of topic relevance.

You can use this system prompt in several ways:

Batch processing: Paste a list of article headlines and URLs. The LLM will score and sort them for you. This is particularly effective for processing RSS feed items or newsletter roundups.

Email triage: Forward emails (or paste their contents) to the LLM for screening. This works best for newsletters, alerts, and automated reports. (Obviously, be mindful of confidentiality when pasting content into LLMs.)

Real-time filtering: Keep a conversation open with this system prompt and drop items in as you encounter them during the day. This works well if you are someone who encounters information throughout the day through multiple channels.

Example in Action

Let me show you what this looks like in practice. Suppose you are a product manager at a health-tech company, and your priority map looks like:

Tier 1: Healthcare regulation changes, competitive product launches, data privacy law

Tier 2: Product management methodology, health-tech market trends, user research methods

Tier 3: Behavioral economics, digital health innovation, medical ethics

You paste in a batch of ten headlines from your morning RSS reader:

1. "FDA Proposes New Framework for AI-Based Medical Devices"
2. "10 Productivity Hacks for Product Managers in 2026"
3. "Competitor X Launches Patient Portal with AI Triage Feature"
4. "The History of Stethoscopes: A Surprisingly Interesting Story"
5. "HIPAA Enforcement Actions Up 40% in Q1"
6. "Why Behavioral Nudges Fail in Healthcare Settings"
7. "React 22 Released with Major Performance Improvements"
8. "New Study: Patients Prefer AI Symptom Checkers to WebMD"
9. "How to Run Better Sprint Retrospectives"
10. "European Parliament Debates AI Medical Device Directive"

The LLM with your relevance filter system prompt might return:

1. FDA AI Framework — DIRECTLY RELEVANT (Tier 1: healthcare regulation)
   Time-sensitive. PROCESS NOW. Major regulatory development affecting
   your product category.

2. PM Productivity Hacks — NOT RELEVANT. Generic listicle.
   IGNORE.

3. Competitor X Portal — DIRECTLY RELEVANT (Tier 1: competitive
   launches). Time-sensitive. PROCESS NOW. Direct competitor feature
   launch.

4. History of Stethoscopes — NOT RELEVANT. Interesting but no
   connection to priorities. IGNORE.

5. HIPAA Enforcement — DIRECTLY RELEVANT (Tier 1: data privacy law).
   Current. PROCESS TODAY. Enforcement trends affect your compliance
   posture.

6. Behavioral Nudges in Healthcare — MODERATELY RELEVANT (Tier 3:
   behavioral economics + Tier 2: health-tech). Evergreen. DEFER for
   reading block.

7. React 22 — TANGENTIALLY RELEVANT (your product uses React, but this
   is a Tier 2 topic at best). Evergreen. SKIM if time permits.

8. AI Symptom Checkers Study — MODERATELY RELEVANT (Tier 2:
   health-tech market trends). Current. DEFER for reading block.

9. Sprint Retrospectives — TANGENTIALLY RELEVANT (Tier 2: PM
   methodology). Evergreen. IGNORE unless you're currently having
   retro problems.

10. EU AI Medical Directive — DIRECTLY RELEVANT (Tier 1: healthcare
    regulation). Current. PROCESS TODAY. Regulatory environment for
    potential EU expansion.

In about 30 seconds, you have gone from ten undifferentiated items to a clear action plan: two items need immediate attention, two should be processed today, two go to the reading queue, and four can be ignored. Without the filter, you would have spent ten minutes scanning all of them and probably still made suboptimal choices about which to prioritize.

RSS Feeds + LLM Filtering

RSS is the unsung hero of personal information management. In an era of algorithmic feeds, RSS gives you chronological, unfiltered access to sources you have explicitly chosen. No engagement optimization. No algorithmic curation. Just the content, in order.

The limitation of RSS is volume. If you subscribe to enough feeds to cover your priority map thoroughly, you end up with more items per day than you can manually triage. This is where the LLM filter becomes powerful.

Setting Up the Pipeline

Step 1: Curate your feeds by tier.

Organize your RSS subscriptions according to your priority map:

• Tier 1 feeds: Official sources for your mission-critical topics. Government agencies, regulatory bodies, official company blogs, primary industry publications. Keep this list short and authoritative.
• Tier 2 feeds: Quality publications in your development areas. Trade journals, research digests, curated newsletters, thoughtful blogs by practitioners in your field.
• Tier 3 feeds: Whatever interests you. Magazines, blogs, newsletters, podcasts. This is where you allow yourself breadth.

Step 2: Process feeds at different frequencies.

• Tier 1 feeds: Daily. These are your first triage priority.
• Tier 2 feeds: 2-3 times per week, or during dedicated reading time.
• Tier 3 feeds: Weekly, during your Friday review.

Step 3: Use LLM filtering for high-volume feeds.

For feeds that produce more than 5-10 items per day, batch the headlines through your relevance filter. Most RSS readers let you export or view items in a format you can paste into an LLM conversation. The filter will identify the 2-3 items worth your attention and let you skip the rest.

Automation Options

If you are technically inclined, you can automate parts of this pipeline:

• RSS reader API + LLM API: Write a script that pulls new items from your RSS reader, sends them through the LLM filter, and tags or stars the relevant ones. Most RSS readers (Feedbin, Miniflux, FreshRSS) have APIs, and LLM APIs are straightforward to use.

• Zapier/Make/n8n workflows: Connect your RSS reader to an LLM step that categorizes items, then routes relevant ones to your read-later app or task manager.

• Local scripts: A Python script that runs every morning, pulls your RSS items, sends them through an LLM API with your system prompt, and produces a daily briefing email. This is maybe 50-100 lines of code and is surprisingly satisfying to build.

I hesitate to give specific tool recommendations because the landscape changes quickly, but the architecture is stable: content source → LLM filter → prioritized output. The specific tools matter less than the pattern.

Email Triage with AI Assistance

Email is the most universal information channel and the one where most people waste the most time. Applying relevance filtering to email is high-impact but requires some care around privacy and confidentiality.

What You Can Automate

Newsletters and digests: These are the easiest to filter. They are not confidential (they were sent to your entire subscriber list), and they are often high-volume. Batch them through your relevance filter during morning triage.

Automated reports and alerts: Dashboards, monitoring alerts, system notifications. Most of these are noise most of the time. Use the LLM filter to identify when one actually requires attention.

Industry news roundups: Weekly digests from industry publications. Run them through the filter and extract only the items that hit your priority map.

What You Should Not Automate

Confidential communications: Do not paste internal emails, client communications, or anything covered by NDA into an external LLM. If you use an enterprise LLM deployment with appropriate data handling, different rules may apply.

Relationship-sensitive correspondence: Emails from your manager, direct reports, or key stakeholders deserve personal attention, not LLM screening. The content matters, but so does the subtext — tone, urgency, what is left unsaid — that an LLM cannot reliably assess.

Anything requiring judgment about people: Performance-related communications, team dynamics, sensitive HR matters. Keep humans in the loop for human issues.

A Practical Email Workflow

Here is a workflow that balances automation with appropriate caution:

Morning email triage (15 minutes):

1. Scan sender and subject line for all new emails
2. Immediately process anything from key people (manager, direct reports, active project stakeholders)
3. Batch all newsletters, digests, and automated reports → paste into LLM with relevance filter → process only the flagged items
4. For everything else, apply your triage rubric manually
5. Anything that does not need a response today → archive with a “this week” label

This hybrid approach uses the LLM for what it is good at (screening high-volume, low-sensitivity content) and preserves human judgment for what it is not (nuanced, relationship-sensitive communication).

The “I Don’t Know What I Don’t Know” Problem

The most dangerous limitation of any relevance filter — human or AI — is that it can only filter based on criteria you have defined. If your priority map does not include a topic, information about that topic will be filtered out even if it is profoundly important.

This is the unknown-unknowns problem, and it is the reason that a relevance filter must not be airtight. You need deliberate gaps in your filter — channels where unfiltered, unprioritized information can reach you.

Building in Deliberate Diversity

Here are concrete strategies for maintaining exposure to information outside your defined priorities:

The Wildcard Feed: Subscribe to 2-3 broad, high-quality sources that cover a wide range of topics. Not everything they publish will be relevant to your priorities — that is the point. Once a week, scan these feeds without your relevance filter. Look for things that surprise you, challenge your assumptions, or open up topics you had not considered.

Good wildcard sources tend to be:

• Generalist publications with strong editorial standards
• Cross-disciplinary journals or magazines
• Curated newsletters from people with very different backgrounds from yours
• The “recommended reading” sections of publications you already trust

The Random Expert Strategy: Once a month, find someone with deep expertise in a field unrelated to yours and read something they have written for a general audience. A marine biologist writing about ocean ecosystems. A historian writing about economic panics. A philosopher writing about personal identity. The goal is not to become an expert in their field; it is to expose yourself to different ways of thinking about problems.

The Dissent Channel: Deliberately subscribe to one or two sources that frequently disagree with your existing views. Not fringe sources — credible, well-argued sources that reach different conclusions than you typically do. If you are a technology optimist, read a thoughtful technology critic. If you lean toward government intervention, read a credible case for market solutions. Your relevance filter should never filter out good-faith disagreement.

Conversation as discovery: Some of the most important information reaches us through conversations with other humans, not through media consumption. Make it a habit to ask people — colleagues, friends, acquaintances — “What are you reading/thinking about lately?” Other people’s attention is an information discovery channel that no algorithm can replicate.

The LLM as Unknown-Unknown Detector

You can also use your LLM assistant to help identify blind spots:

Monthly blind spot check prompt:

Here is my current information priority map:
[Your priority map]

And here is a summary of the topics I've been consuming information
about this month:
[List of topics from your recent reading]

Based on this, what important topics or developments might I be missing?
Consider:
- Emerging trends that could affect my Tier 1 priorities
- Cross-disciplinary connections I might not see
- Risks that are growing but not yet on most people's radar
- Areas where my Tier 1 topics intersect with fields I'm not tracking

This is not a perfect solution — the LLM has its own blind spots — but it is a useful supplement to your own reflection.

Tuning Your Filter Over Time

A relevance filter is not a set-it-and-forget-it system. Your priorities change. Your role changes. The world changes. A filter that was perfectly calibrated six months ago might be significantly miscalibrated today.

Monthly Filter Review

Once a month, spend 20 minutes reviewing and adjusting your filter:

Review your priority map:

• Have any Tier 1 topics resolved or become less critical?
• Have any Tier 2 or 3 topics become urgent enough to move to Tier 1?
• Are there new topics that should be added at any tier?
• Are there topics that should be removed because you have lost interest or they are no longer relevant?

Review your sources:

• Which sources consistently provided high-value content this month?
• Which sources consistently provided content that your filter scored as Ignore?
• Are there gaps in your source coverage for your Tier 1 topics?
• Unsubscribe from underperforming sources. Subscribe to new ones that fill gaps.

Review your filter accuracy:

• Did you miss anything important because your filter excluded it?
• Did your filter let through too much low-value content?
• Are there patterns in the false positives (filter says relevant, you disagree) or false negatives (filter says irrelevant, but it actually mattered)?
• Adjust your system prompt and scoring rubric based on these patterns.

Quarterly Priority Map Overhaul

Once a quarter, do a more thorough review:

• Revisit your Tier 1 list from scratch. What are you actually responsible for right now?
• Look at what information you actually used in the last quarter. Which topics generated information that influenced your decisions or work?
• Identify any topics where you invested significant reading time but got little return. Consider downgrading or removing them.
• Talk to your manager, colleagues, or clients about what they think you should be paying attention to. External perspective is valuable for catching blind spots.

The Difference Between a Relevance Filter and a Comfort Filter

This is the section I almost did not write because it is the most uncomfortable one. But it is also possibly the most important.

A relevance filter selects information based on your defined priorities, the quality of the source, and the timeliness of the content. A comfort filter selects information based on whether it confirms your existing beliefs, makes you feel good, or avoids topics that cause you anxiety or discomfort.

These are different things, and they can look the same from the outside. Consider:

• “I’m filtering out political news because it’s not relevant to my work.” That might be a legitimate relevance judgment. It might also be avoidance of uncomfortable reality.

• “I’m unsubscribing from this economist because their analysis is usually low-quality.” That might be a legitimate quality judgment. It might also be because their analysis challenges your preferred economic framework.

• “I’m ignoring this emerging risk because it’s speculative.” That might be a legitimate timeliness judgment. It might also be because the risk is scary and you would prefer not to think about it.

I am not saying that every act of filtering is secret comfort-seeking. Most of the time, when you filter out political news because it is not relevant to your work, that is exactly what is happening. But the overlap between “not relevant” and “not comfortable” is large enough to warrant periodic self-examination.

How to Check Yourself

The inversion test: For any topic you are filtering out, ask yourself: “If the content coming through on this topic consistently confirmed my existing beliefs, would I still filter it out?” If the answer is no — if you would happily consume it when it agreed with you — then you are not filtering for relevance. You are filtering for comfort.

The discomfort metric: Periodically review what your filter is catching. If nothing in your information stream is making you uncomfortable, challenging your assumptions, or introducing unwelcome complexity, your filter may be too tight. Good information sometimes hurts.

The outsider review: Describe your filter criteria to someone you trust. Ask them if they see any topics where your “relevance” judgment might actually be “comfort” judgment. Other people can see your blind spots more easily than you can.

This is not about being a masochist who seeks out distressing content. It is about maintaining the intellectual honesty that makes a relevance filter useful rather than harmful. The whole point of building your own filter — instead of relying on an algorithm’s — is that you can make it smarter and more honest. Do not waste that opportunity by recreating a comfort bubble and calling it a relevance system.

Putting It All Together: A Complete Implementation

Let me walk through a complete implementation from scratch, step by step. This assumes you are starting with nothing and want a working system within a week.

Day 1: Define Your Priorities (30 minutes)

Sit down with a blank page. Write your priority map with all three tiers. Be specific — not “technology” but “cloud infrastructure security for AWS and GCP.” Not “business” but “SaaS pricing strategy and competitive positioning.”

Day 2: Inventory Your Sources (30 minutes)

List every information source you currently consume: email newsletters, RSS feeds, social media accounts, podcasts, websites you visit regularly. For each one, note which tier of your priority map it serves. If it does not serve any tier, mark it for removal.

Day 3: Set Up Your System Prompt (20 minutes)

Using the template from this chapter, create your personal relevance filter system prompt. Include your full priority map and scoring rubric. Save it somewhere accessible.

Day 4: Organize Your Feeds (30 minutes)

If you do not already use an RSS reader, set one up. (Miniflux, Feedbin, NetNewsWire, or Inoreader are all solid choices.) Import your current sources, organized by tier. Unsubscribe from any newsletters or feeds that did not connect to your priority map.

Day 5: Run Your First Filtered Triage (30 minutes)

Use your new system: RSS feeds sorted by tier, headlines processed through your LLM relevance filter, actions taken based on the scoring rubric. Note what works and what feels wrong. Adjust the system prompt if the filter is too aggressive or too permissive.

Day 6-7: Refine and Build Habits

Run the system for two more days. By now, you should have a feel for the rhythm: morning triage with LLM-filtered feeds, midday quick check, end-of-day review. Note your weekly review time in your calendar.

Ongoing: Monthly Reviews

Follow the monthly filter review process described above. The system will get more effective over time as you tune it to your actual patterns.

The Payoff

When this system is working well, it feels like having a competent assistant who reads everything and briefs you on only what matters. Your morning triage becomes fast and focused. Your reading time is spent on content that is actually relevant and high-quality. Your exposure to new ideas is maintained through deliberate channels rather than left to algorithmic chance.

More importantly, you stop feeling guilty about all the things you are not reading. The filter is making those decisions for you, based on criteria you defined. When you ignore something, it is not because you failed to get to it — it is because you made an explicit choice that it was not worth your time. That is a fundamentally different experience from the low-grade anxiety of an overflowing inbox that you know you will never clear.

The filter will not be perfect. You will occasionally miss something that mattered, or spend time on something that turned out to be a waste. But perfect coverage was never achievable anyway. What is achievable is a system that makes your information consumption deliberate rather than reactive, efficient rather than exhausting, and honest rather than comfortable. That is the goal, and it is within reach.