agentleFS
OHOKAYYEAHOHOKAYYEAH/claude-skills/skills/distill/SKILL.md

distill

Skill1 starsChanged 35 days agoMITView on GitHub

Distill any text: a pasted transcript, meeting captions, an article URL, or a file. Returns a TL;DR plus the substance at a set compression ratio in one of four shapes (themes, argument, timeline, meeting), each point anchored to the source. Fire on "summarize this transcript / article / meeting", "distill this", "what did they decide". NOT a YouTube link (/tldw, if installed), NOT Claude's own last reply (/tldr), NOT a near-verbatim cleanup (/readable).

76 other projects use a version of this.

Adopted by

Most recently taken up by

Used with

Projects with it also use layout, shape and 8 more.

Other files people use alongside this one.

Files found in the same public projects as a version of this one, more often than they turn up anywhere else. Each project counts once, large collections count for less, and this file's own repository is left out.
  • owshq-mecSkill

    layout

    Improve layout, spacing, and visual rhythm. Fixes monotonous grids, inconsistent spacing, and weak visual hierarchy. Use when the user mentions layout feeling off, spacing issues, visual hierarchy, crowded UI, alignment problems, or wanting better composition.

    Also in 8 projects that use this

  • HAR5HA-7663Skill

    shape

    Plan the UX and UI for a feature before writing code. Runs a structured discovery interview, then produces a design brief that guides implementation. Use during the planning phase to establish design direction, constraints, and strategy before any code is written.

    Also in 11 projects that use this

  • ReScienceLabSkill

    archive

    Archive session learnings, debugging solutions, and deployment logs to .archive/yyyy-mm-dd/ as indexed markdown with searchable tags. Use when completing a significant task, resolving a tricky bug, deploying, or when the user says \"archive this\". Maintains .archive/MEMORY.md index for cross-session knowledge reuse.

    Also in 3 projects that use this

  • klauswgSkill

    ckm:design-system

    Token architecture, component specifications, and slide generation. Three-layer tokens (primitive→semantic→component), CSS variables, spacing/typography scales, component specs, strategic slide creation. Use for design tokens, systematic design, brand-compliant presentations.

    Also in 3 projects that use this

  • iovisorSkill

    evolve

    Evolve uBPF through a gated spec-first workflow that propagates approved changes through requirements, design, validation, implementation, and RFC alignment artifacts.

    Also in 3 projects that use this

  • termdockSkill

    recall

    Retrieves cross-session memories about past decisions, lessons, and patterns. Use when user asks about prior work, past decisions, or mentions a feature. Triggers include check memory, what did we decide, how did we solve.

    Also in 5 projects that use this

The file

We split this skill into the rules it gives an agent

7 rules. 0 are common, 5 are distinctive, and 2 are very specific to this project.

  • 0

    Common

    Most of the 72 edited versions we read keep these.

  • 5

    Distinctive

    This author's own choices. Worth a look.

  • 2

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Pasted: never retype a paste. Claude Code keeps it verbatim in the session log, and this copies it out:

    line 34
  2. Very specific:

    File: use it where it is; don't copy it.

    line 40
  3. Distinctive:

    Read all of it, in parts if one Read won't hold it.

    line 51
  4. Distinctive:

    Faithful to the source: No outside knowledge anywhere except the Reader's note (or the argument shape's "How strong is the case").

    line 59
  5. Distinctive:

    Compress, don't transcribe: Keep an anecdote only when it *is* the point.

    line 60
  6. Very specific:

    Write $DIR/summary.SHAPE-RATIO.md (e.g.

    line 67
  7. Distinctive:

    Keep only the Note items that apply, and drop the line when none do.

    line 81

“Common” is measured: kept by at least 6 in 10 of the 72 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Distill a long text so the reader gets its substance without reading all of it: real compression, not an annotated copy, with every point anchored to where it came from. The top block (TL;DR, or Thesis in the argument shape) is the one-glance version; everything below it is the depth. The source stays in context afterwards, so follow-ups are answered from what it actually says.

1. Parse the call

  • Source, first match wins: a file path in the arguments; a URL in the arguments; the text the user pasted in this conversation (the newest paste, unless they point at another); otherwise ask for it and stop.
    • A YouTube link belongs to the /tldw skill, if you have it, which fetches the transcript; say so and stop. Without it, ask for the transcript text.
    • Claude's own previous reply belongs to /tldr.
  • Shape: themes, argument, timeline, or meeting in the words overrides the default for this call. With no shape named, a speaker-labelled transcript with three or more voices takes the configured meeting_shape. Say which shape you picked in one line, so a re-shape is one word away.
  • Ratio: 5:1 or ratio 5 overrides the default. "Shorter" roughly doubles it; "more detail", "deep", or "deeper" roughly halves it.
  • Changing a default ("make meeting my default", "use 15:1 from now on"): run python3 "${CLAUDE_SKILL_DIR}/scripts/config.py" set shape meeting (or set ratio 15, or set meeting_shape themes), confirm what it prints, and carry on if there's also a source to do.
  • Anything else is focus ("just the part about pricing", "what did the CTO say about the roadmap"). Lead with it.

Already distilled this source in this session (a re-shape, a new ratio, a follow-up)? Skip to step 4; reading it again costs thousands of tokens for nothing.

2. Get the source on disk

Pick the output folder: ${DISTILL_DIR:-$HOME/distill}/SLUG/, where SLUG is a short kebab-case name for the source plus its date when one is known (pricing-talk-2025-03-14). A source distilled before reuses its folder, so check ls first and reuse a matching one.

  • Pasted: never retype a paste. Claude Code keeps it verbatim in the session log, and this copies it out:
    python3 "${CLAUDE_SKILL_DIR}/scripts/grab_paste.py" \
      --out "$DIR" --match "A PHRASE FROM THE PASTE"
    It prints {"path", "words", ...}. On no_paste (short pastes arrive untagged), write the text to $DIR/source.txt yourself; on no_session, do the same and mention it.
  • File: use it where it is; don't copy it. A PDF gets converted first, so it can be counted and read in parts: pdftotext "$F" "$DIR/source.txt", then work from source.txt.
  • URL: fetch it with WebFetch, asking for the article's full text verbatim with headings, and save that as $DIR/source.txt. WebFetch runs the page through a model, so name the route in the header's Note line. A paywall, login wall, or empty result: ask the user to paste the text.

Then count the words a reader reads. Caption timing lines and speaker labels don't count:

python3 "${CLAUDE_SKILL_DIR}/scripts/count.py" SOURCE_FILE

3. Read the source

Read all of it, in parts if one Read won't hold it. A summary of the first half is not a summary. Caption files interleave people who talk over each other, so track who said what by the labels, not by position.

4. Write it

Read the shape's layout first: ${CLAUDE_SKILL_DIR}/references/SHAPE.md (themes.md, argument.md, timeline.md, or meeting.md). Then apply these rules in every shape:

  • Length follows the ratio. The target is the source's counted words ÷ ratio (with a focus, the words of the part in focus). Density moves it by about a quarter either way: a repetitive pep talk compresses harder, a dense report runs longer. Never pad, and never drop a distinct claim to fit. When cutting, cut support before claims, and cut names, numbers, citations, and attributions last. There's no upper cap. A very short source may fall below what the shape needs; then write the top block and the few points that exist, and skip empty sections.
  • Anchor every point with the source's own markers, never computed ones. Captions → the clock time the point starts ((14:04)); a timestamped lecture → (12:34); an article or report → its section heading ((§ Pricing)) or page ((p. 12)); a source with no markers gets no anchors.
  • Faithful to the source. No outside knowledge anywhere except the Reader's note (or the argument shape's "How strong is the case"). If the source is too garbled or empty, say so in one line and stop.
  • Compress, don't transcribe. Keep an anecdote only when it is the point. Cut setup, repetition, jokes, crowd work, tangents, sponsor reads, and plugs.
  • Speakers. Name someone only when a label or the text makes it clear; otherwise "the host", "the author", "someone in the room". A label shared by several people (a conference-room mic, a dial-in line) gets named in the Reader's note, not silently split into guesses.
  • Mangled names and jargon. Captions and speech-to-text garble proper nouns. When one looks wrong, flag it once ("likely Kalshi, captions say Kosci") instead of silently guessing.
  • Reader's note (last section, all shapes but argument): the one place outside knowledge is allowed, labelled as yours: claims made without evidence, facts that look wrong, what the source leaves out, speaker-label caveats. Skip it when there's nothing worth flagging.

5. Save and check

Write $DIR/summary.SHAPE-RATIO.md (e.g. summary.meeting-10.md). A focused run saves as summary.SHAPE-RATIO.focus-SLUG.md, so it never replaces the whole-source summary:

# TITLE

- **Source:** FILE PATH or URL, or "pasted in session" (saved copy: $DIR/source.txt)
- **Made:** DATE · SHAPE at RATIO:1 · SOURCE_WORDS → SUMMARY_WORDS words · by YOUR_MODEL_ID
- **Note:** caption transcript · fetched via WebFetch · shared-mic speaker labels

---

SUMMARY

Keep only the Note items that apply, and drop the line when none do. Then count the body mechanically, because models under-count their own output by about a quarter:

python3 "${CLAUDE_SKILL_DIR}/scripts/count.py" SUMMARY_FILE --kind summary

More than about a quarter off target: revise and count again before showing it. Put the final count in the Made line.

6. Present

Show it in the chat; that's the point. End with one line: the saved path, and the three follow-ups: questions ("what did the host say about liability?"), a re-shape ("as meeting"), or a new depth ("5:1"). When the source is a transcript, add that /readable makes a cleaned near-verbatim edition from the same saved copy.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.