How to annotate a text properly
Highlighting and annotation get bundled together, but they do different jobs and only one of them survives contact with time.
A highlight records that a sentence mattered. Three weeks later you can see the mark and have no idea what you were thinking. An annotation records why — and the why is the part that was doing the work.
The cost is that annotation is slower, so it needs a system compact enough that you will still be using it in week nine.
A symbol set small enough to remember
Elaborate colour-coded schemes get abandoned by the third chapter. Five symbols is roughly the limit of what people maintain without thinking:
! — a claim
The author is asserting something arguable. These are the sentences essay questions are built from, and they are easy to walk past because claims are often phrased flatly.
? — I do not follow
Mark and keep going. Do not stop to resolve it — later sections often answer it, and stopping is how people fail to finish chapters.
→ — connects to
Another chapter, a lecture, a contradicting source. Write the target in four words. These connections are what separate a good answer from a complete one.
✓ — evidence
A statistic, study or example that proves a nearby claim. Pairs with "!" — an exclamation mark with no tick anywhere near it is a claim the author did not support.
⌀ — I disagree
The most valuable and least used mark. Your objection at first reading is worth recording even if you later decide you were wrong.
What to actually write in the margin
Margin notes should be four to eight words. Long notes do not get written, and when they do get written they do not get re-read.
The useful ones fall into a few shapes: a translation of a dense sentence into plain language, a connection to something else, an objection, or a flag about scope — “only true at STP”, “sample was 30 undergrads”, “contradicts Lewis p.88”, “exam Q 2024”.
What is not useful is restating the sentence. “Important” and “key point” tell you nothing you did not already know from the fact that you marked it.
Annotating when you cannot write on it
Library books, borrowed texts and locked PDFs rule out marginalia, and the workarounds are mostly worse than people expect.
Sticky notes work but detach and reorder themselves. A separate notebook works but divorces the note from the sentence, so you end up with a page of observations you cannot locate in the text.
The least bad method is a running document with page and line references: “p.44 l.12 — contradicts Lewis”. It is slower, but it is searchable, it survives returning the book, and it accumulates across sources in a way marginalia never does. If you are working from a PDF you can copy text from, pasting the passage above your note makes it genuinely self-contained.
Digital annotation: searchable, but easy to over-collect
Digital tools make annotations searchable and backed up, which is a real advantage over a shelf of marked-up books you will never open again.
The failure mode is different, though. Clipping is so cheap that people collect hundreds of highlights per book and never process any of them — the digital equivalent of a page that is 60% yellow. A collection you never revisit is not a knowledge base, it is a graveyard with good search.
If you work digitally, impose the same scarcity you would on paper and do the processing step: convert annotations into questions, and review them on a schedule. The tooling does not remove the need for retrieval practice, it just makes it easier to pretend you are doing it.
Where automatic highlighting fits
Annotation and automatic highlighting are complements rather than alternatives, because they fail in opposite places.
A statistical highlighter is good at the mechanical part: scanning a long text and marking the sentences that are structurally central, numerically dense or definitional. It has no opinions, no memory of your lecture, and no idea what your essay question is.
You are good at exactly those things and bad at maintaining attention across forty pages. So run the mechanical pass first to get a sensible shortlist, then spend your attention annotating that shortlist — the objections, connections and scope flags that no algorithm can produce. That division also follows the one-in-ten rule: machine narrows, you annotate.