agentleFS
BinayakDotelBinayakDotel/skills-marketplace/skills/research/SKILL.md

research

Skill1 starsChanged 37 days agoMITView on GitHub

Research a topic on the web and return a sourced fact brief for a script, video, article or story. Every claim carries a source link and an exact short quote from that page, legend is kept apart from recorded history, competing versions are listed separately, and disputed points, common myths and gaps are flagged. Ends with one JSON block in a fixed format (research-brief/1) that apps can import and check. Use when the user types /research, asks to research or fact-check a topic, wants "the real story" or "facts with sources" behind a festival, legend, temple, place, custom, historical event or recent news, or is preparing content that has to be accurate.

189 other projects use a version of this.

Adopted by

Most recently taken up by

Used with

Projects with it also use Problem-Solving Techniques, daily-arxiv and 10 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.
  • truongtv22Skill

    Problem-Solving Techniques

    Apply systematic problem-solving techniques for complexity spirals (simplification cascades), innovation blocks (collision-zone thinking), recurring patterns (meta-pattern recognition), assumption constraints (inversion exercise), scale uncertainty (scale game), and dispatch when stuck. Techniques derived from Microsoft Amplifier project patterns adapted for immediate application.

    Also in 6 projects that use this

  • LambenthanSkill

    daily-arxiv

    每日从 arXiv 拉取新论文,过滤相关性,auto-ingest 高优先级论文,检测 SOTA 更新

    Also in 5 projects that use this

  • LambenthanSkill

    exp-eval

    实验判决门:Review LLM 独立评判实验结果 → 4 种判决路径 → 自动更新 claims confidence、ideas status、graph edges

    Also in 5 projects that use this

  • liufeng1976Skill

    setup-ts-deep-modules

    Wire dependency-cruiser into a TypeScript repo so each package is a deep module — implementation hidden in subfolders, reachable only through its entry-point files. User-invoked.

    Also in 49 projects that use this

  • parcadeiSkill

    ast-grep-find

    AST-based code search and refactoring via ast-grep MCP

    Also in 4 projects that use this

  • SireJeffSkill

    context-optimize

    Context Optimization - Orchestrate RPI workflow TODO generation with interactive scoping

    Also in 4 projects that use this

  • +5 more from Lambenthan

The file

Show its rules and text

We split this skill into the rules it gives an agent

17 rules. 0 are common, 7 are distinctive, and 10 are very specific to this project.

  • 0

    Common

    Most of the 170 edited versions we read keep these.

  • 7

    Distinctive

    This author's own choices. Worth a look.

  • 10

    Very specific

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

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    If the topic is ambiguous (two places with one name, two festivals with similar names), pick the likely one, say which in question, and …

    line 19
  2. Distinctive:

    If a text tells it, what are its actual words at the key moment?

    line 27
  3. Distinctive:

    When does it happen? Its place in the calendar (which month, lunar day or season.

    line 29
  4. Very specific:

    Open every page you cite: Never cite a search snippet, a page you could not open, or your own memory.

    line 32
  5. Very specific:

    Keep versions apart: When legends or accounts differ, give each one a versions entry and tie its facts to it.

    line 35
  6. Very specific:

    Shape the story: when the user is making content (see *Story shape*): one angle, 2–3 hooks, and an outline of 5–8 beats that cite fact ids.

    line 37
  7. Very specific:

    If the user asks about this year's event, still research what happens every time.

    line 66
  8. Very specific:

    No source, publication or institution names in angle, hooks or outline.

    line 82
  9. Very specific:

    If you cannot search or open pages: say so in the summary, set web_access to false, give facts from memory with confidence: "low" and empty …

    line 151
  10. Very specific:

    Every field with examples: references/brief-schema.md.

    line 153
  11. Distinctive:

    Every fact has at least one source, and every quote is copied, not paraphrased.

    line 157
  12. Distinctive:

    No claim says more than its quote (or the source's date) supports.

    line 158
  13. Distinctive:

    Every referenced id exists, the JSON parses, and nothing follows the JSON block.

    line 163
  14. Very specific:

    If you can run Python: python3 <skill-dir>/scripts/check_brief.py <file> prints ok.

    line 164
  15. Distinctive:

    Do not paste page text or long working notes into the chat: only the summary and the JSON.

    line 172
  16. Distinctive:

    Do not restate the JSON in prose.

    line 173
  17. Very specific:

    Stop searching once every sub-question has a high- or medium-confidence fact or is listed in gaps.

    line 174

“Common” is measured: kept by at least 6 in 10 of the 170 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

Research brief

Find out what is actually known about a topic and hand back a brief that someone can write from without re-checking everything. Whoever reads the brief will trust it, so accuracy beats completeness: ten solid facts are worth more than thirty guesses.

When the brief is for a story (a narrated video, an article), it must also carry the story: a person or object to follow, moments you can see, a question that pulls the audience through, and why it matters to their own life. Facts that only say who reported what make a report, not a story. And the story must still be true when someone watches it years later (see Any day, any year).

Input

  • A topic or question, usually with the user's own idea and notes. Treat their notes as leads to check, never as facts.
  • Sometimes: what it is for (for example a 60–90 s narrated video), a target language, things to include or avoid.

If the topic is ambiguous (two places with one name, two festivals with similar names), pick the likely one, say which in question, and carry on. Ask a question only when the topic cannot be identified at all.

Workflow

  1. Break it down (silently): 4–8 sub-questions. For the origin of a temple, festival or custom: what each legend or text says (one version at a time), which text and roughly how old, what recorded history or archaeology shows and when, what people do today, and what is commonly said but wrong. For a story, add the story questions:
    • Who or what is at the centre: a person, a god in a legend, a community, an object, a journey?
    • What is the single most vivid moment, and what exactly happens in it: who does what, holding what, where?
    • Which numbers could be shown on screen (how many people, how many days, how far)?
    • If a text tells it, what are its actual words at the key moment?
    • What question does the topic make people ask ("why would an army salute a bundle of crops?"), and which fact answers it?
    • When does it happen? Its place in the calendar (which month, lunar day or season; every year, every twelve years), so the story can say "every year on …" instead of "today".
    • What does an ordinary family do with it each time it comes round, step by step, and what does it mean for their lives?
  2. Search wide, then deep. Use encyclopedic overviews to map names, dates and versions, then open the sources they cite and better ones (see Source quality). Search in English and in the topic's own language and script (for Nepal topics, also in Nepali/Devanagari). Budget: about 6–12 searches and 5–10 opened pages.
  3. Open every page you cite. Never cite a search snippet, a page you could not open, or your own memory. If a page is paywalled or will not load, leave it out and add it to gaps.
  4. Record facts while reading, one claim per fact, each with the exact words that support it. Prefer facts that describe something happening (who, what, where, with what) over facts about who wrote or listed it. A quote is 5–25 consecutive words copied character for character from the page, in the page's own language and spelling. No ellipses inside a quote and no stitching of separate sentences. If no short passage supports a claim, the claim is unsupported: drop it, or reword it to what the page actually says.
  5. Label and rate each fact (tables below).
  6. Keep versions apart. When legends or accounts differ, give each one a versions entry and tie its facts to it. Never blend two versions into one smooth story.
  7. Hunt the myths. Popular claims you found to be wrong or unsupported go into dont_say. Real disagreement between good sources goes into disputed.
  8. Shape the story when the user is making content (see Story shape): one angle, 2–3 hooks, and an outline of 5–8 beats that cite fact ids. Only high- and medium-confidence facts may appear in the outline.
  9. Self-check (list below), then reply.

Fact types

typeMeaningHow to word the claim
legendwhat a scripture, text or folk tradition says happenedName the story or text that tells it: "The Devi Mahatmya says…", "Local legend says…" (never the website you read it on)
historydocumented events: inscriptions, records, archaeology, dated accountsPlain statement with the date
practicewhat people do, how it is observed, every time it comes roundHabitual present, with when it happens: "Every year on Navami of Dashain, …"
sciencea natural or scientific explanationPlain statement
figurea number, measurement or statisticInclude the unit and the year it applies to
currentnews, or anything likely to change: this year's date of a festival, a price, a countInclude "as of <date>"; the custom itself is a practice fact

Confidence

  • high: a primary source, or two independent reliable sources agree.
  • medium: one reliable source.
  • low: only weak sources, or reliable sources disagree. Low facts never go into outline.

A claim may not say more than its quote supports. Dates may also come from the source's own publication date (published). Watch for words that sneak in extra meaning: "oldest", "first", "only", "always", exact numbers.

Any day, any year

A video or article made from the brief stays up for years, so everything in it must be true on any day it is read or watched.

  • Word angle, hooks, outline and claim_local for any day: never "today", "tomorrow", "this year" or "now" for the research day.
  • Something that recurs (a festival, a ritual, a season, a harvest) is a practice fact worded as recurring, with its calendar anchor from a source: "Every year on Navami of Dashain, the temple opens to the public for one day." Use the festival day or lunar day (tithi) the source gives: festivals that follow the moon fall on a different calendar date every year, so this year's date is a separate current fact with the year. Word something as recurring only when a source says it recurs.
  • A number that changes (a price, a count of visitors, a population) is a figure with its year. Never work out a new number (days between two dates, a total, an age) in a claim, a hook or the outline.
  • If the user asks about this year's event, still research what happens every time; keep this year's specifics as dated current facts.

Story shape

For content, the angle, hooks and outline are the story's spine; the writer builds on them.

  • angle: the story in one line, with its spine: "The bundle of crops a whole army salutes", not "Fulpati: history and significance".
  • hooks: scenes or questions an audience can picture, at most two sentences each, each resting on facts in the brief. Never a summary, the topic's name, or a scene that is true only on one day ("This morning the queue already…").
  • outline, in this order:
    1. the hook moment;
    2. the story itself, following one person, object or journey, in two or three beats;
    3. the turn: the question the story raises;
    4. the answer, from the facts (the text's words, the evidence, the reason);
    5. bring it home: what people do every time it comes round, and what it says about their lives;
    6. a closing image that echoes the opening.
  • Each beat says only what its own facts support, so a writer who drops one fact can drop its beat.
  • No source, publication or institution names in angle, hooks or outline. The sources live in sources, where they belong.
  • claim_local is plain spoken language, the way you would tell it. For a legend, attribute it to the story or text ("कथा अनुसार", "दुर्गा सप्तशतीमा"), never to a website, newspaper or organisation. A practice is told as recurring ("हरेक वर्ष दशैंको नवमीमा …"), never as "आज" or "यो वर्ष".

Source quality (best first)

  1. Primary texts and their translations (scriptures, chronicles, inscriptions), official records, archaeological reports.
  2. Academic work: university presses, journals, scholarly encyclopedias, museum and UNESCO pages.
  3. Government and institutional sites: the body that manages a site, national archives, departments of archaeology, official calendars.
  4. Established news outlets with a named author and a date.
  5. Wikipedia and general encyclopedias: fine for orientation and for finding sources. Cite them only for uncontroversial basics, and prefer the source they cite.

Prefer sources that tell what happened in detail (the text itself, a scholar's account, a long-form feature) over ones that only mention it. Never rely only on blogs, tour-company pages, content farms, social posts, AI-written pages or forums. Many sites copy one unsourced sentence from each other; that counts as one source, not ten. More per topic type in references/source-guide.md (if present).

Output

Reply with exactly two parts:

  1. A short summary for the user, 6–10 lines: the answer in one sentence; what is solid; which parts are legend and which versions exist; what is disputed or unverified; anything sensitive.
  2. One fenced json block with the brief, and nothing after it.

The shape, shortened to one entry per list (the full brief behind it is examples/kasthamandap.json):

{
  "schema": "research-brief/1",
  "topic": "Kasthamandap",
  "question": "How did Kasthamandap come to be, and how did it give Kathmandu its name?",
  "language": "ne",
  "researched_on": "2026-10-09",
  "web_access": true,
  "angle": "The city named after one wooden pavilion",
  "hooks": ["One wooden pavilion gave Kathmandu its name."],
  "outline": [{ "beat": "Kathmandu is named after a wooden pavilion", "facts": ["F1"] }],
  "versions": [{ "id": "V1", "name": "Single-tree legend", "summary": "Legend says all its timber came from one tree.", "facts": ["F2"] }],
  "facts": [
    {
      "id": "F1",
      "claim": "The name Kathmandu comes from Kasthamandap, a wooden pavilion.",
      "claim_local": "काठमाडौंको नाम काष्ठमण्डप नामको काठको पाटीबाट आएको हो।",
      "type": "history",
      "version": null,
      "confidence": "high",
      "sources": [
        {
          "url": "https://en.wikipedia.org/wiki/Kathmandu",
          "title": "Kathmandu",
          "publisher": "Wikipedia",
          "published": null,
          "quote": "The Nepali name Kathmandu comes from Kasthamandap.",
          "accessed": "2026-10-09"
        }
      ]
    }
  ],
  "disputed": [{ "point": "How old the site is", "positions": ["7th century CE (archaeology)", "5th century (one carbon-dated joint)"], "facts": ["F4", "F8"] }],
  "dont_say": [{ "claim": "It was built from a single tree.", "why": "That is the legend; attribute it: 'legend says'." }],
  "gaps": ["Could not open the 1143 CE manuscript that first mentions the pavilion."],
  "notes": ["The 2015 collapse killed people; keep that part respectful."]
}

Field rules

  • schema is always "research-brief/1". researched_on and accessed are YYYY-MM-DD. language is the language of the final content as a short code (ne, en, hi).
  • Ids: facts F1, F2… and versions V1, V2… in order of appearance. Every id used in outline, versions or disputed must exist.
  • claim: one specific sentence (names, dates, numbers), in English unless the user asked otherwise. claim_local: the same sentence in language when that is not English; otherwise leave it out.
  • version: the version id for a legend fact that belongs to one account; otherwise null.
  • sources: 1–3 per fact, each with its own exact quote. Two quotes from the same page are two entries.
  • facts: usually 8–20. Fewer is fine for a narrow topic; never pad.
  • If you cannot search or open pages: say so in the summary, set web_access to false, give facts from memory with confidence: "low" and empty sources, and add "No web access: verify everything" to gaps.

Every field with examples: references/brief-schema.md. Machine-readable schema for apps: references/brief.schema.json. A finished brief: examples/kasthamandap.json.

Self-check before replying

  • Every fact has at least one source, and every quote is copied, not paraphrased.
  • No claim says more than its quote (or the source's date) supports.
  • Legends are attributed to their story or text ("…says"), never stated as history.
  • For content: the outline follows Story shape, and angle, hooks and outline name no source.
  • Everything is true on any day: no "today", "tomorrow" or "this year" outside a dated current fact; recurring practices carry when they happen; no number you worked out yourself.
  • Versions stay separate, and the outline does not merge them.
  • Every referenced id exists, the JSON parses, and nothing follows the JSON block.
  • If you can run Python: python3 <skill-dir>/scripts/check_brief.py <file> prints ok. It also reads a whole reply saved as .md and checks its last json block. --fetch also checks that each quote appears on its page.

Sensitive topics

Religion, caste, ethnicity, politics, living people and tragedies: report what sources say with attribution, give the major viewpoints fairly, keep neutral wording, and record sensitivities in notes (for example places where photography is not allowed, or rites a community keeps private). For recent events, prefer reports from the last few days and put the date in the claim.

Keep it cheap

  • Do not paste page text or long working notes into the chat: only the summary and the JSON.
  • Do not restate the JSON in prose.
  • Stop searching once every sub-question has a high- or medium-confidence fact or is listed in gaps.
  • <skill-dir> is the folder containing this file. Read the files in references/ only when a field or a source question is unclear.

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.