Progressive Summarization: How to Distill Notes With AI
Progressive summarization is a way to distill notes in layers: save the source, bold the key passages, highlight the best of those, then write a short summary in your own words. AI can now do the first layers in seconds. The last one is still yours, and it is the one that makes notes worth keeping.
Founder of The Viable Edge, with 20+ years of experience helping businesses build competitive advantage through strategic transformation. Expert in AI-era business strategy and systematic implementation.
Published · Updated
Progressive summarization is a note-taking technique from Tiago Forte: you distill a note in layers each time you revisit it. Layer one is the saved passage, layer two bolds the key lines, layer three highlights the best of those, and layer four is a short summary in your own words. With AI, let the model draft layers two and three, and write layer four yourself.
Progressive summarization is a technique for distilling notes a little more each time you use them. Created by Tiago Forte as the distill step of his CODE method, it adds a layer of compression on each visit, so a note gets shorter and sharper only if it proves useful. Notes you never revisit cost nothing extra.
It is the D in CODE (capture, organize, distill, express), the workflow half of the second brain method. The organizing half is the PARA method. Progressive summarization answers the question PARA leaves open: once a note is filed, how do you make it quick to use later?
The idea that matters is "progressive". You do not summarize everything at capture time. You summarize a note when you return to it, which means your effort goes to the notes that earn it.
What are the four layers of progressive summarization?
Layer one is the original passage you saved. Layer two bolds the sentences that carry the point. Layer three highlights the best of the bold lines. Layer four is an executive summary of a few bullets in your own words at the top of the note. Each layer is roughly a tenth of the one before.
Layer
What you do
When
1. Capture
Save the passage, with its source
When you read it
2. Bold
Bold the lines that carry the point
First time you revisit
3. Highlight
Highlight the best of the bold lines
When you use it again
4. Summary
Write 3 to 5 bullets in your own words
Only for notes you keep returning to
A fifth layer, remixing the idea into something of your own, belongs to the express step. Most notes never get past layer two, and that is the design working.
How does AI change progressive summarization?
AI does layers two and three almost instantly: give it a long note and it will pull out the key sentences and the single best line. That removes the step most people skipped. What AI cannot do is layer four, because a summary in your own words is how you check that you understood the idea.
The weak point of the method was always time. Bolding and highlighting every note you revisit takes attention most people never had, so notes stayed long and unread. A model can now draft those layers on demand, every time, for free.
The risk moves to trust. A model's highlight is a guess about what matters to you. Read it against the original, and keep the source line on every note so you can. Our second brain retrieval test is a quick way to check whether AI answers from your notes are landing on the right passages.
Free download
Start with the Second Brain Starter Kit
The capture inbox, note template, source log, weekly checklist, connect-your-AI guide and retrieval self-test, as plain markdown files that work in Obsidian or any editor, with Claude or any AI. Enter your email and the download starts.
No spam, ever
Unsubscribe anytime
Which parts of distilling notes should you not hand to AI?
Keep the layer-four summary and the decision about what a note is for. Writing three bullets in your own words is where the idea becomes yours, and it is what makes the note useful to an AI later, because the model then reads your judgment instead of a stranger's paragraph. Hand off the bolding, never the summary.
There is a practical reason as well. When your AI tools answer from your notes, the layer-four summary is usually what they quote. A summary you wrote reflects your position. A summary the model wrote reflects the model's, and over a few hundred notes the knowledge base drifts toward generic.
A workable split:
AI drafts: key sentences, a one-line highlight, related notes to link.
You write: the three-bullet summary and one line on why the note matters to your work.
You check: anything the AI drafted, against the source, before relying on it.
How do you start using progressive summarization?
Start with notes you already use. Pick five you opened this month, add a three-bullet summary at the top of each in your own words, and keep the original text below. From then on, add a layer only when you return to a note. Ask your AI tool to draft the bold lines first if the note is long.
You do not need a special app. Plain markdown handles every layer: **bold** for layer two, ==highlight== in Obsidian or a second bold pass for layer three, and a short list under the title for layer four. Our free second brain template already has a note template with a space for the summary and the source.
What are common progressive summarization mistakes?
Summarizing at capture time instead of on revisit, which spends effort on notes you may never use. Bolding so much that nothing stands out. Letting AI write the final summary, which makes every note sound the same. And dropping the source, which makes any summary impossible to check later.
If a note has more bold than plain text, it has not been distilled; start layer two again with a stricter eye. And if your notes are meant to feed a team, the same discipline at company scale is in how to build a knowledge base for AI.
For consultants and service providers
Build second brains for clients
Build and Deliver Professional Second Brains is a self-paced course on doing this for other people: scoping the engagement, building a knowledge base from a client's real material, testing retrieval, and handing it over, with a working knowledge base kit to build with. Plain files, no coding, no vector database.
