You hear a sharp point on a podcast or read an article that changes how you think. Then it goes into Apple Notes with a dozen other half-thoughts, or nowhere. The problem isn't capturing more. It's deciding what's worth keeping for you, and finding it again later. Here's a workflow that does both, and how to turn it into a skill.
The workflow
| Step | What happens | Time |
|---|---|---|
| Lens | You write down what you care about, once | 10 min, once |
| Capture | The full text goes in, not a vague memory of it | 1 min |
| Filter | AI keeps only what's new and useful to you | 1 min |
| Store | One card per source, in a place you already search | Automatic |
| Resurface | A weekly review turns cards into ideas and actions | 10 min a week |
Set your lens once
The lens turns a generic summary into your filter. Keep it short, and update it when your priorities change.
MY LENS
Goals right now: [2-3 things you're working toward]
Questions I'm chasing: [open questions you'd love an answer to]
Projects this could feed: [current work, content, decisions]
Already know well, skip it: [topics you don't need explained]
Always flag: [a claim that challenges my view, a tool worth testing, a number worth checking]
Get the full text in
AI filters what you give it. A title or a link it can't open leads to a confident summary of nothing.
| Source | How to get the text |
|---|---|
| Article | Paste the text, or share the link if your assistant can open it |
| YouTube | Under the video: More → Show transcript, then copy it (only for videos with captions) |
| Podcast | Spotify shows transcripts in the mobile app for many episodes. For a full copy you can paste, use a transcript tool (see below) |
| PDF or report | Upload the file |
| Your own voice note | Dictate your reaction right after. It's often the most valuable part |
What I use for podcasts. I'm still looking for an MCP server that pulls a full Spotify transcript directly. For now I use SpotScribe and paste the transcript in myself. The free tier transcribes up to 60 minutes per episode and has a built-in AI chat on top, and so far I'm loving it.
On a separate note, Audioscrape is a cool MCP server I came across. Its core job is a little different: it lets your assistant search a large catalog of already-transcribed podcasts and pull speaker-labeled transcripts with timestamps. It's become a staple in my stack.
Filter it into a keep card
The core prompt. It forces a verdict and caps the output, so you get a card, not a second article.
Using my lens below, filter this source.
1. Verdict: KEEP, SKIM or SKIP, with one line on why.
2. If KEEP: at most 3 insights that are new to me. For each, give:
- the insight in one sentence
- why it matters for my goals or projects
- a quote or timestamp from the source as evidence
3. One thing that challenges what I currently think, if there is one.
4. One action: something to try, test, use or look up.
5. Tags from my lens.
Leave out anything I already know or that doesn't connect to my lens.
[lens]
[source text]
SKIP is a good result. It means you can stop thinking about that source.
Store it where you'll find it
Save one card per source in the tool you already search, like Obsidian, Notion or a markdown folder. Use the same header every time:
---
source: [title + link]
type: article | video | podcast | pdf
date: [today]
tags: [from the lens]
status: new
---
If you keep topic notes, add a one-line link from each insight to the matching note. That's what turns a pile into connected notes.
Make it stick
Saving isn't remembering. Once a week, give your new cards back to the AI:
Here are this week's keep cards. Group them by theme, show me which ones connect to each other or to my current projects, and pick the 3 most useful. Quiz me with one question on each. Then list any card that led to nothing, so I can archive it.
The quiz matters most. Having to recall an idea does more for memory than rereading it, and archiving keeps the collection worth searching.
Turn it into a skill
In Claude Code or Codex, save it as a skill so one command filters and files. Both read the same SKILL.md format.
---
name: keep
description: Filter an article, transcript or PDF through my lens and save a keep card to my notes.
---
1. Read my lens from notes/lens.md.
2. Get the full source text. If you can't access it, ask me to paste it. Never summarize from the title.
3. Apply the keep-card filter: verdict, max 3 new insights with quotes, one challenge, one action, tags.
4. If KEEP, save to notes/inbox/YYYY-MM-DD_<slug>.md with the standard header. If SKIP, save nothing and tell me why.
5. Link each insight to a matching topic note if one exists.
In ChatGPT or the Claude apps, put the lens and the filter prompt in a Project's instructions instead. Then paste each source into a new chat in that Project.
Where it fails
- Invented quotes. Summaries can misquote a source. Requiring a quote or timestamp makes that easy to spot-check.
- Lost nuance. A filter drops caveats. Go back to the source before you repeat a claim, publish it or act on it.
- An echo chamber. A tight lens only keeps what you already agree with. The "challenges what I think" line is there on purpose.
- Hoarding again. If everything comes back KEEP, your lens is too loose. Tighten it.
Start with the next three
- Write your lens (10 minutes)
- Run the filter on the next 3 things you'd normally save or forget
- Store the KEEP cards with the header above
- Put a 10-minute weekly review on your calendar
- After two weeks, turn it into a skill or Project if you're still using it
Go deeper: How I Use AI to Index My Brain · The 10-Minute Weekly Context Audit · Stop Building Your AI System Inside One Tool · Follow-Up Prompts That Fix Weak AI Answers