agentleFS
wxxb789wxxb789/hoa/skills/official-document-verification/SKILL.md

official-document-verification

Skill3 starsChanged 53 days agoMITView on GitHub

Verify a load-bearing claim against the exact print of the document that carries it: pin the print before reading, scan by number and by content, and get an artifact a plain client cannot fetch. Use when a conclusion rests on a bill, a statute, an audit report, a regulator's page, or any artifact whose wording and version decide the answer. Not for ordinary web reading where the page itself is the source of record.

The file

We split this skill into the rules it gives an agent

8 rules. About 0 are common, 7 are distinctive, and 1 is very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈7

    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
  1. Distinctive:

    Use when a load-bearing claim rests on a bill, a statute, an audit report, a regulator's page, or any artifact whose exact version and …

    line 20
  2. Distinctive:

    Keep status facts and text facts on separate records: the publisher's action history answers "has it moved", the fetched text answers "what …

    line 34
  3. Very specific:

    Search section numbers and the citation of the older law, then open every hit and read its heading.

    line 42
  4. Distinctive:

    Follow with a content scan for the named things — frozen capability lists, program elements, defined terms.

    line 45
  5. Distinctive:

    When the numbers live in a chart or an embedded image, read the rendered figure: screenshot the element and read it as an image, then …

    line 57
  6. Distinctive:

    Keep a per-figure status in the evidence note: figure readable / table unreadable = *partially blocked*.

    line 61
  7. Distinctive:

    When every path on a domain returns the same block page, that is an unresolved item.

    line 63
  8. Distinctive:

    When a page has a reviewer or another writer, a draft edit is not a proposal — it is an unauthorized landing.

    line 77

≈ “Common” is an estimate: at least 1 in 5 of the 350,495 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

Official document verification

Use when a load-bearing claim rests on a bill, a statute, an audit report, a regulator's page, or any artifact whose exact version and exact wording decide the conclusion — and when that artifact is behind bot protection, a jurisdictional gate, or a page a plain HTTP client cannot reach.

Companion skills: grounded-citations (citation ledger mechanics), evidence-backed-analytical-writing (claim hygiene and report structure), ops-claim-verification (cron results and "fixed" claims). This skill covers what those do not: getting the right version, proving what it does and does not say, and proposing a change instead of making one.

1. Pin the version before reading a single line

  • Bills and statutes exist as several distinct prints: introduced → reported (RS) → engrossed (EH) → enrolled → public law. "The bill" is not a version. Every absence claim is scoped to the print actually fetched, so name it (chamber, print, size, line count) in the note that carries the claim.
  • Keep status facts and text facts on separate records: the publisher's action history answers "has it moved", the fetched text answers "what does it say". A quiet action history does not constrain a future text, and text content does not establish procedural status.
  • Report each finding per version. A provision present in the House print and absent from the Senate print is a single-chamber proposal — write that, not "proposed", which invites the reader to assume both.

2. Scan by number and by content

  • Search section numbers and the citation of the older law, then open every hit and read its heading. Section numbers repeat across titles: a hit on Sec. 904 may be an unrelated provision in another title while the section in question sits elsewhere. A number-only match publishes a false finding.
  • Follow with a content scan for the named things — frozen capability lists, program elements, defined terms. Zero hits for those names across the whole text is the strongest negative available.
  • State the negative as bounded: "in the fetched EH/RS text", never "in the bill", and never as a procedural forecast. "No renewal clause in either chamber's text" is a finding; "renewal can only come through conference" is a prediction the text does not contain.
  • scripts/scan_bill_text.py performs this pass (section hits, citation windows, keyword counts) on a downloaded HTML or text file; run it before writing any absence claim.

3. Getting an artifact a plain client cannot fetch

  • A 403 to curl is a statement about your client or the edge protection, not about the document. Load the URL in a real browser before concluding the source is gone, empty, or unpublishable.
  • When the numbers live in a chart or an embedded image, read the rendered figure: screenshot the element and read it as an image, then record that you read a figure and not the report's table.
  • Check a chart reading by internal arithmetic. Sum the component bars and compare with the total the figure prints. Disagreement means your transcription is wrong, not the source — re-read before quoting.
  • Keep a per-figure status in the evidence note: figure readable / table unreadable = partially blocked. A legible chart never licenses citing the body text or the data table.
  • When every path on a domain returns the same block page, that is an unresolved item. Split the question and record each part separately: listing/index state, access control on a gated list, availability of the original file elsewhere or in a mirror, and whether the document was submitted or is publishable. One unreachable index answers none of them, and "channel closed" is not a finding about a blocked page.

4. Caliber ladder for counts

authorized → actual/staffed → filled/hired; positions are not people, and neither is money. Name the caliber in the same sentence as the number. When sources give different baselines, list them side by side with their enumerators instead of reconciling: the disagreement is usually itself the finding, and a single reconciled number launders it away.

5. Propose the change, do not make it

When a page has a reviewer or another writer, a draft edit is not a proposal — it is an unauthorized landing. Prepare it outside the tree:

  1. Copy the target into scratch; apply edits with an assertion that each anchor string occurs exactly once, so a silent no-op or a double replacement fails loudly instead of shipping.
  2. Emit diff -u --label a/<repo-path> --label b/<repo-path> <original> <candidate> so the reviewer can apply it verbatim.
  3. Validate before sending: frontmatter parses, headings and links intact, and git status --porcelain empty as proof the real tree was untouched.
  4. In the covering note, say what each hunk replaces and what it preserves, so acceptance can happen at hunk granularity.

On a dated page, annotate rather than rewrite: keep the historical sentence and its date, attach a dated correction marker, and state what the new evidence can and cannot compare with the old figure.

Pitfalls

  • Treating one print's silence as the measure's silence, then reporting it as "not in the bill".
  • Matching a section number without reading its heading; the number repeats across titles.
  • Escalating a bounded negative into a procedural conclusion about what later stages can or must do.
  • Declaring a source unavailable because a plain fetch was blocked, when a browser renders it fine.
  • Quoting a chart total without summing its components; the transcription error becomes a cited fact.
  • Letting a readable figure stand in for the table or body text it does not contain.
  • Reconciling two staffing or budget baselines into one number instead of naming both calibers.
  • Editing a shared page to demonstrate a proposal, which converts review into an accident.

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.