text-distill
Skill0 starsChanged 7 days agoMITView on GitHub
Rewrites any text into a factual, logically structured document with no filler. Handles video and audio transcripts, monologues, lectures, essays, interviews, dialogues, Zoom or Teams meeting recordings, and chat exports, and turns them into a step-by-step guide, lecture notes, meeting minutes, or a summary of agreements. Keeps every fact, number, contact, link, and decision and who said what; removes greetings, repetition, and speech noise; anonymizes people on request; reports what to verify. Use when the user asks to "remove the fluff", turn a transcript into instructions or notes, clean up a rambling text, anonymize it, write minutes from a call recording, or extract what was agreed in a chat, even if they just paste a long text or point to a .txt, .srt, or .md file without naming the skill. Writes a new Markdown file and never changes the source. Do not use for a one-line summary, translation, literary editing that keeps the author's voice, or transcribing audio or video itself.
In this repository4 other files it gives its agents.
- publish-skillSkill
- codebase-auditSkill
- codex-imagegenSkill
- modular-architecture-refactorSkill
The file
Show its rules and text
We split this skill into the rules it gives an agent
15 rules. About 0 are common, 14 are distinctive, and 1 is very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈14
Distinctive · estimate
This author's own choices. Worth a look.
1
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 36
If the document ends up holding secrets (passwords, access tokens, API keys, full card details), keep them too, but warn the user in the …
- Distinctive: line 37
Keep names of organizations, products, services, and interface elements: they are facts.
- Distinctive: line 42
Do not silently guess at doubtful ones: pick the most likely reading and list it in the report as "to verify".
- Distinctive: line 46
If numbers or statements in the source contradict each other, do not silently pick one.
- Distinctive: line 55
Identify the genre and the form: Read the matching section of references/genres.md.
- Distinctive: line 66
No introductions like "This document describes…".
- Distinctive: line 75
every listed fact is in the document.
- Distinctive: line 76
every statement in the document has support in the source, and guesses are marked or removed.
- Distinctive: line 77
no fact is repeated across sections.
- Distinctive: line 79
every statement from a conversation says whose it is.
- Distinctive: line 81
no leftovers of speech remain.
- Very specific: line 84
Write a .md file next to the source or where the user said.
- Distinctive: line 87
Never modify or rename the source files.
- Distinctive: line 88
If the text was pasted into the chat, return the result in the chat.
- Distinctive: line 90
Report to the user, briefly: Do not repeat the document.
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Text Distill
Rewrite the source so that a reader can understand it and act on it without opening the original. Every fact from the source is there, nothing is invented, nothing is superfluous.
A good result is not a short retelling but a complete document of smaller size: it holds everything that changes the reader's understanding or actions, and nothing that does not.
Principles
1. Facts are untouchable. Keep every number, amount, deadline, condition, exception, sequence of actions, contact, address, and link, name of a product, service, section, or button, every warning and pitfall. Add nothing of your own: knowledge you have about the topic must not appear in the document as if the source said it.
A guess is not a fact. If you inferred something the source does not say directly (where a countdown starts, an intention like "I'll order it today" turned into "ordered", what an estimate includes), do not state it in the document. Either leave it out or mark the spot with a short "(to verify)" and put the details in the report. The reader of the document does not see the report, so a guess written as a fact misleads them.
2. Filler is everything that does not change understanding:
- greetings, self-introductions, goodbyes, thanks;
- "as you can see", "so, yeah", "hope that's clear", "no big deal";
- the same point said twice, thinking aloud, digressions;
- data from a demonstration: someone's balance, the specific date on the presenter's screen, how many people are in their team. The general rule stays; the example goes;
- incidental details that affect nothing: where the keys were picked up, who joined the call late, who apologized for the noise. The test: if the line were deleted, would anything change for the reader? If not, delete it.
Each fact once. Repeating a point in another section is filler too. Do not restate a fact in a summary, a table, and a recommendations list. A "Key points" block belongs only in a long document (more than two screens). A "(to verify)" mark goes in one place, where the fact is written.
3. People: keep who said what.
- In dialogues, interviews, meetings, and chats, keep the participants' names as the source gives them. Who reported, proposed, agreed to, or owns something is part of the fact, and the reader needs it to know whose information it is. Use one form of each name throughout.
- A chat export may label its owner only as "self" or "me". Take the name from the conversation if it appears there (someone addresses them by name); otherwise keep the export's label and say so in the report.
- A lone narrator (a tutorial, a lecture) is not a participant whose authorship matters: their self-introduction is filler, and a guide speaks impersonally.
- Keep concrete data exactly as the source gives it: phone numbers, addresses, emails, usernames, links, account, contract, and order numbers, amounts. In work chats and calls these are often the very facts the reader needs; a summary that says "a phone number was shared" without the number is useless.
- If the document ends up holding secrets (passwords, access tokens, API keys, full card details), keep them too, but warn the user in the report, so they decide before sharing the document.
- Keep names of organizations, products, services, and interface elements: they are facts.
- Anonymize only when the user asks ("anonymize", "no names", "for publication"). Then replace names with roles ("curator", "designer"; "Participant 1" when no role is clear), remove the contacts, addresses, and IDs that identify people, and give the role-to-name mapping only in the report, never in the document.
4. Speech-recognition errors. Automatic transcripts confuse similar-sounding words, mangle names, and lose punctuation. Whisper-style models also add phrases nobody said, especially at the end of a recording and in pauses: credits such as "Subtitles by …" or "Thanks for watching", "To be continued…", and their equivalents in the source's language.
- Fix obvious errors and list them in the report.
- Do not silently guess at doubtful ones: pick the most likely reading and list it in the report as "to verify".
- Spell the same name the same way everywhere.
5. Contradictions and questionable advice.
- If numbers or statements in the source contradict each other, do not silently pick one. Either keep both with a note, or drop them and say so in the report.
- If the source recommends something risky or dishonest (entering knowingly false data in an official form, bypassing a check, an unsafe action), do not turn it into the document's recommendation. Keep the neutral fact ("a tax ID is required") and raise it in the report: it is the user's call.
6. Structure serves the reader, not the order of speech. Speech jumps between topics; the document follows a logical order: conditions → actions → result, or context → decisions → tasks. Choose the form by genre.
Workflow
- Read the whole source and look around it. Read long texts in parts and keep a running list of facts so that nothing is lost. Check whether a file with the same name and a
.srtor.vttextension sits next to it: transcripts often come with subtitles, which give timestamps for the places to verify. If there are several files, write one document per file by default. If they form one topic, number the results in logical order and cross-reference them by number. - Identify the genre and the form. Read the matching section of references/genres.md. If the user named the form ("write minutes"), follow it.
- List the facts: actions, conditions, numbers, terms, decisions, tasks, open questions, warnings.
- Build the structure.
- The title names the subject of the document, not the source: "Buying products in store X", not "Video transcript".
- Sections follow a logical order.
- Actions go in a numbered list, one action per item.
- Uniform data (form fields, options, plans, deadlines, owners) goes in a table.
- Important warnings go in a blockquote (
>) so that they stand out. - Bold marks interface elements and key values.
- No introductions like "This document describes…".
- Write.
- Guides: impersonal imperative ("Click", "Enter").
- Notes and summaries: neutral statements.
- Short sentences, the same term for the same thing throughout.
- The source's language, unless the user asks for another.
- Check before saving:
- every listed fact is in the document;
- every statement in the document has support in the source, and guesses are marked or removed;
- no fact is repeated across sections;
- the numbers match the source;
- every statement from a conversation says whose it is;
- contacts, addresses, and links match the source character for character;
- no leftovers of speech remain.
- Save.
- Write a
.mdfile next to the source or where the user said. - Name it by content:
<Topic> — <aspect>.md, in the document's language, without characters that file names forbid. - Number a series:
01 …,02 …. - Never modify or rename the source files.
- If the text was pasted into the chat, return the result in the chat; if it is longer than a couple of screens, also save it as a file in the working folder.
- Write a
- Report to the user, briefly. Do not repeat the document; it is in the file.
- which files were created and what each one covers;
- what was removed (by category, not as a list of phrases);
- which recognition errors were fixed;
- what to verify against the source: contradictions, guesses, questionable advice. If a subtitle file or timestamps exist, give the time of each place;
- secrets the document contains (passwords, tokens, keys), if any;
- how participants without a name in the source were labeled, and the role-to-name mapping if the user asked for anonymization.
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
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.

