agentleFS
berkayburakkberkayburakk/tubecrew/skills/research-facts/SKILL.md

research-facts

Skill2 starsChanged 30 days agoMITView on GitHub

Build the sourced fact pack for a video - every number with its primary source, link, type and date in one facts.json that the script and visuals read from, plus a do-not-use list. Use at the start of research, when a new filing or event lands, or when someone asks "where does this number come from" or "check my numbers".

The file

We split this skill into the rules it gives an agent

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

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈5

    Distinctive · estimate

    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:

    Prefer the primary source.

    line 32
  2. Very specific:

    Keep three dates apart: when you accessed it, what period it describes (as_of), and when the event happened.

    line 33
  3. Distinctive:

    Do not mix similar-sounding things: revenue, annualized revenue, profit, valuation and money raised are five different numbers.

    line 34
  4. Distinctive:

    After the script is written:

    line 50
  5. Distinctive:

    Every number in the script must match a fact.

    line 56
  6. Distinctive:

    No advice phrasing (buy, sell, you should invest).

    line 72
  7. Very specific:

    When the creator's stack (~/.tubecrew/stack.json) has verifier: codex, also run the same audit through Codex, a different model family, …

    line 79

≈ “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.

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

Research facts

The script never contains a number that is not in facts.json. That one rule removes most of the corrections a factual channel ever has to pin.

$SKILL below is this skill's folder.

1. Collect

For every number you might use, add an entry (start from $SKILL/templates/facts.example.json):

"revenue_2025": {
  "value": 12.5, "unit": "billion $", "type": "primary",
  "source": "Acme Corp annual report 2025, page 41",
  "url": "https://example.com/acme-annual-report-2025",
  "accessed": "2026-10-01", "as_of": "fiscal year 2025",
  "quote": "Total revenue was $12.5 billion."
}
  • type is one of primary (the company, regulator or official dataset said it), news (a report about it), estimate (an analyst or model), market (exchange or market data: a price, an index level, a yield), own-calculation (yours; write the formula in calc using other keys).
  • A market fact needs as_of with the date and the session or time zone ("2026-03-04 close, ET"). Take a closing price from the exchange's own historical series, not an aggregator; they can differ on a single day.
  • Prefer the primary source. A news article is a pointer to the primary source, not the end of the search.
  • Keep three dates apart: when you accessed it, what period it describes (as_of), and when the event happened.
  • Do not mix similar-sounding things: revenue, annualized revenue, profit, valuation and money raised are five different numbers; announced, signed and completed are three different states.
  • Anything you found but must not say (unverifiable, contradicted, too old) goes into do_not_use with the reason, so nobody "rediscovers" it later.

2. Validate

python $SKILL/scripts/facts_check.py facts.json

It fails on missing links, bad dates, unknown types, market entries without a dated session or time zone, and own-calculation entries whose formula does not recompute to the stated value. Fix until it exits 0.

3. Gate the script

After the script is written:

python $SKILL/scripts/facts_check.py facts.json --script script.md

Every number in the script must match a fact. A number may be rounded to the precision it is written with (12.52 billion as "12.5 billion") as long as it stays within 5% of the fact; anything looser needs a hedge word ("about", "more than", "up to") in the direction that keeps it true. A fact of 0.45 is "about 0.5 percent", never a bare "0.5 percent" (11% off). Dates and times are reported, not checked. A do_not_use claim in the script fails the gate. Exit 1 means fix the script, not the fact file.

How numbers are said

  • Without "about" or "around", a number is exact; do not shorten 176.52 to 176 silently.
  • "Up to" for ceilings, "reportedly" for news, "estimated" for estimates, "percentage points" when two percentages are subtracted.
  • Superlatives ("the biggest", "the first", "the only") need their own source.
  • Hook and thumbnail numbers need two independent sources that agree.
  • Nothing that was not reported is drawn as fact (someone's cost basis, an implied share, a made-up chart).
  • Rounded parts add up to the total shown with them; otherwise show the residual as its own "≈" term. A label ("non-cash") belongs only to the item it describes.
  • No advice phrasing (buy, sell, you should invest). Explain; do not recommend.

Hand-off

The fact-checker agent audits the final script against primary sources before any voiceover is made. Give it the script path and facts.json; the pack is a lead for it, not proof.

When the creator's stack (~/.tubecrew/stack.json) has verifier: codex, also run the same audit through Codex, a different model family, with the read-only command in the setup skill's references/tools.md. A claim either checker marks WRONG gets fixed; a disagreement goes to the creator.

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.