Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Use when someone hands you a source — an article, meeting notes, a transcript, a spec, a saved thread — and wants to be able to skim it fast later and still drill into the full detail. Distills it in nested layers WITHOUT rewriting it: keeps the original text intact, bolds the load-bearing sentences inside it, highlights a smaller subset of that bold, and adds a short top-line gist in your own words above the source. Do NOT use to produce a standalone summary that replaces the source, to cut text to a length budget, to copyedit, or to extract structured fields.
.claude/skills/progressive-summarization/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-01 | ✗→✓ | ▲ Improved | 210% | 0% |
| case-02 | ✗→✓ | ▲ Improved | 299% | 0% |
| case-03 | ✗→✓ | ▲ Improved | 260% | 0% |
| case-05 | ✗→✓ | ▲ Improved | 214% | 0% |
| case-06 | ✗→✓ | ▲ Improved | 363% | 0% |
Asked to "summarize this so I can review it later," a base model writes a fresh, shorter paraphrase and throws the source away. That is the wrong move for something you want to revisit. A paraphrase is lossy and final: when you come back in three months and the two-line summary isn't enough, the detail is gone, and you can't tell which words in it were the author's and which the model invented.
Progressive summarization does the opposite. You never rewrite the source — you keep it whole and add emphasis on top of it, in nested layers, so a future reader can zoom from a one-glance gist down to the full original at whatever depth they need. Nothing is discarded; the layers just let you skim.
Output them top to bottom so the reader moves shallow → deep:
before deciding whether to read further. This is the only layer you compose; every other layer is pure selection.
delete, reorder, paraphrase, or "clean up" a single sentence. This is what makes the technique non-lossy.
that carry the argument, the decision, the number, the ask. A reader who reads only the bold should still get the point.
essential, the handful of fragments you'd keep if you could keep almost nothing.
reordered, or reworded any of the original, you broke the technique. The whole value is that the full text is still there.
highlight a fragment you didn't bold, and you never bold text that isn't in the source. Each layer is a tighter selection of the layer beneath, not a new pass over the material.
minority that carries the weight; highlight a minority of that. Over-marking destroys the point of the layers — a page that's all yellow is a page with no signal.
to emphasize in the author's text, not restating it.
Use standard Markdown so the layers survive copy-paste:
**load-bearing sentence**==essential fragment== (Obsidian/Notion style; <mark>essential fragment</mark> also works)Given a source paragraph, the output looks like:
> Gist: The team is delaying the launch two weeks to fix the checkout bug; revenue impact is > expected to be small. > > We met Thursday to review launch readiness. The checkout flow still fails for about 3% of > users on mobile, which the team traced to a payment-timeout race. After discussion, ==we > decided to push the launch two weeks== rather than ship with the known defect. Marketing has > been notified. Finance estimates the delay costs roughly $40K in deferred revenue, which > leadership judged acceptable against the reputational risk of a broken checkout.
Notice: the full paragraph is still there, the bold alone reads as a summary, the highlight alone is the single decision, and the gist sits on top — four depths, one preserved source.
Other measured skills in the registry, with their headline benchmark lift.