agentleFS
samsational-codesamsational-code/business-skills/document-redline-diff/SKILL.md

document-redline-diff

Skill0 starsChanged 2 months agoMITView on GitHub

Produce a word-level redline of any recurring or revised official document against its prior version — a policy statement, terms of service, privacy policy, pricing page, regulatory guidance, company guidance or outlook, contract draft, press release, product policy, SLA, employee handbook, central bank statement — then classify every substantive change by direction (stricter/looser, expanding/retreating, more/less committed, favorable/ unfavorable to the reader) with one line of reasoning each, summarize any accompanying numbers or projections, and test whether the observed reaction matches what the text actually says. For legal, compliance, procurement, policy, communications, product, investor-relations, and strategy teams — anyone who has to explain "what changed" under time pressure. Trigger on phrases like "diff these two versions," "what changed from the last one," "redline this against the prior," "is this stricter or looser," "what did they take out," "compare the old and new terms," or when a user pastes two versions of the same document. Requires both versions to be provided — it never reconstructs a document from memory. Do NOT use for summarizing a single document or speech. Do NOT use to extract the full obligations package from a contract — use the contract-obligations-extractor skill. Do NOT use for a presentation or earnings call transcript — use the presentation-transcript-extractor skill. Not legal or investment advice.

The file

We split this skill into the rules it gives an agent

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

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈2

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 0

    Very specific

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

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Keep the whole brief to roughly one page excluding the redline.

    line 120
  2. Distinctive:

    "Ambiguous" is a valid classification: Do not force a directional call on a change that genuinely cuts both ways — a manufactured signal is …

    line 143

≈ “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

Document Redline Diff

Everyone reads the new version. The edge is in knowing precisely what changed, what each change signals, and whether the first reaction — from the market, the press, customers, or the room — is responding to the text or to a headline. This produces the redline, a change-by-change directional classification, a summary of any accompanying numbers, and a read on whether the reaction and the text agree — fast enough to be useful before the call, the hearing, or the negotiation starts.

Why this skill exists

Recurring official documents are written to be compared. Regulators, central banks, platforms, insurers, and companies reuse language deliberately, and the signal is carried almost entirely by what was added, removed, or reworded — a dropped adjective, a reordered risk sentence, a shift from "we will" to "we may," a definition that quietly widened. Reading the new version alone tells you the position; reading it against the prior tells you the change in position, which is what people react to.

Doing that comparison by hand under time pressure is exactly the kind of task where humans are slow and error-prone: the diffs are small, the language is repetitive, and the first ninety seconds are noisy. The mechanical part — find every difference, miss nothing — is what this automates, so the human spends the time on the judgment part.

Inputs

Required:

  • The new version, in full, pasted or attached
  • The prior version, in full, pasted or attached

Both are genuinely required. See the guardrails — this skill does not reconstruct documents from memory.

Optional but high-value:

  • Any accompanying numbers, new and prior: a pricing table, a fee schedule, guidance ranges, SLA thresholds, economic projections, headcount or budget figures
  • The reaction so far, with timestamps: price or share moves, customer support volume, press coverage, social response, partner or counterparty pushback, internal escalations
  • What was expected going in — what commentators, customers, or the negotiating team anticipated
  • Prior context on whether a change was telegraphed: earlier drafts, prior meeting notes, speeches, blog posts, or a counterparty's stated asks
  • Which side the user is on — the author, the counterparty, a customer, a regulated entity, an investor — because "favorable" depends on the reader
  • The specific question the user cares about

Missing-input behavior: If only one version is provided, ask for the other — do not proceed. If accompanying numbers are not provided, skip that section explicitly rather than describing figures from memory. If no reaction data is provided, run everything except the text-versus-reaction section and say it was skipped; the language analysis stands on its own. If the user's side is unclear and the direction of a change depends on it, ask once; otherwise classify from the perspective of the party the document is addressed to.

How it works

  1. Produce the word-level redline first, mechanically, before interpreting anything. Section by section, showing additions, deletions, and rewordings in place. Completeness matters more than elegance here — a missed deletion is the failure mode.
  2. Separate substantive changes from housekeeping. Date and version updates, rotating names and signatories, renumbering, formatting, and routine reference changes are noted and set aside so they do not dilute the signal.
  3. Pick the classification axis for this document type and state it. Central bank statements are hawkish/dovish; terms of service and policies are stricter/looser for the user; regulatory guidance is expanding/retreating in scope; company guidance is raised/lowered/unchanged; contract drafts are more/less favorable to the user's side; press releases are more/less committed. Then classify each substantive change on that axis, with one line of reasoning. Ambiguous is a legitimate and frequently correct classification — a change that cuts both ways should be labeled as such rather than forced.
  4. Read the structural signals that live above the word level: the order in which topics are discussed, the summary or risk-balance sentence, the conditionality of any commitment or forward-looking statement, the scope of key definitions, who bears each obligation, and any change in how the author describes its own stance or confidence.
  5. Note authorship, approval, and effective date: who issued or signed, any recorded vote or dissent, whether approval composition changed, and when the new version takes effect.
  6. Summarize accompanying numbers if provided: each figure's prior and new value, the change, and the movement of any distribution or range (not just the midpoint — a midpoint that holds while the range widens is itself the story).
  7. Deliver the verdict: net direction relative to the prior version, and the single most significant change. Commit to a call.
  8. Compare text against reaction. Given the observed reaction, assess whether it is consistent with what the text actually says, and name the plausible alternative explanations if it is not — a surprise in the accompanying numbers, a leak that was already priced in, a headline that misread the text, positioning or sentiment unrelated to the document, or anticipation of what will be said next.

Output format

# [Document type] Redline — [new version date] vs. [prior version date]
Issuer: [ ] | Decision or headline action: [ ] | Approval / vote: [ ] | Accompanying numbers: [provided / not provided]
Classification axis: [e.g. stricter/looser for the user; hawkish/dovish; favorable/unfavorable to our side]

## Verdict
Net [direction] versus the prior version.
The single most significant change: [one sentence].
[2–3 sentences of support.]

## Redline
[Section-by-section, showing every difference in place.
Additions and deletions marked clearly. Housekeeping changes separated out.]

## Change-by-change classification
| # | Change | Direction | Reasoning |
|---|---|---|---|
| 1 | "[old]" → "[new]" | [Stricter / Looser / Ambiguous] | [one line] |

Housekeeping (no signal): [list]

## Structural signals
- Summary or risk-balance sentence: [what changed, or "unchanged"]
- Order and emphasis of topics: [ ]
- Conditionality of commitments and forward-looking language: [ ]
- Scope of key definitions and who bears each obligation: [ ]
- Description of the author's own stance or confidence: [ ]

## Authorship, approval, and effective date
[Who issued it, any vote or dissent and direction, whether composition changed, when it takes effect.]

## Accompanying numbers (if provided)
| Figure | Prior | New | Change | Range / distribution note |
|---|---|---|---|---|
| [e.g. price tier, fee, guidance midpoint, SLA threshold, projection] | | | | |
Distribution: [where the mass moved, beyond the headline figure]

## Text vs. reaction (if reaction data provided)
| Signal | Observed | Consistent with the text? |
|---|---|---|
Assessment: [is the reaction tracking the document, the accompanying numbers,
or something else — and what the alternative explanations are]

## Questions the redline raises
3–5 specific questions that only the author, the presenter, or the
counterparty can resolve — for the Q&A, the press conference, the account
call, or the next negotiation round.

Keep the whole brief to roughly one page excluding the redline.

Reading revised language

Recurring patterns worth checking on every document. These are conventions, not rules — verify against the actual text each time.

  • Conditionality shifts carry the commitment. "Additional measures may be appropriate" → "the extent of any additional measures" moves from whether to how much. "We will notify you" → "we may notify you" removes an obligation without deleting the sentence.
  • Adjective removal is a signal. Dropping "elevated," "strong," "reasonable," or "promptly" from a description is usually deliberate and usually the point.
  • The summary or risk-balance sentence is the most reliable single indicator. In a policy statement it is the balance-of-risks line; in a contract it is the limitation-of-liability or termination clause; in guidance it is the sentence that states what the issuer expects.
  • Order matters. Which topic gets discussed first, and which is described with more qualification.
  • Intensity modifiers are load-bearing. "Attentive to" vs. "highly attentive to"; "commercially reasonable efforts" vs. "best efforts"; "may" vs. "will" vs. "shall."
  • Explicit statements about the author's own confidence — "we have gained greater confidence," "we are not yet satisfied," "we remain confident in the full-year outlook" — are among the strongest tells.
  • Trigger specificity that becomes more or less precise about which data, event, or condition activates a provision.
  • Definitions widen quietly. A change to "Personal Data," "Affiliate," "Confidential Information," or "Service" can move more obligation than any operative clause.
  • Sections that rarely change — balance sheet paragraphs, governing law, indemnities, dispute resolution — rarely change incidentally when they do.
  • Dissents and split approvals are information about the distribution of views inside the issuer, and a new dissenter or a flipped direction can matter more than the text.
  • Headline-holding, range-moving numbers: the midpoint is the headline, but a widening or shifting range tells you about conviction.

Guardrails

  • Never reconstruct a document from memory. Both versions must be provided by the user. Recurring official language is reproduced with near-identical wording version to version, which makes a plausible-sounding fabrication both easy to generate and very hard to spot — and a fabricated diff is worse than no diff. If a version is not provided, stop and ask.
  • The redline is mechanical; the classification is interpretation. Keep them visually separate. The reader must be able to check the redline against the source without wading through commentary.
  • Quote exactly. Old and new language reproduced verbatim, in quotation marks, with no paraphrase presented as a quote.
  • "Ambiguous" is a valid classification. Do not force a directional call on a change that genuinely cuts both ways — a manufactured signal is worse than an acknowledged wash.
  • Distinguish the document from the commentary around it. The document is the issuer's collective, considered language; the press conference, the executive's remarks, the account manager's email, and the blog post are not, and are frequently where the real reaction comes from. Never attribute commentary to the document itself.
  • Reaction data is as-provided, timestamped, and never assumed. Do not describe a reaction that was not given to you, and do not extrapolate from partial data.
  • Not legal or investment advice. This produces a reading of a document and a comparison to observed reaction. It does not recommend trades, positioning, a legal strategy, whether to sign, or a forecast of the next version. Contract and regulatory readings should go to counsel before anyone acts on them.
  • Confirm the date and version you are analyzing from the documents themselves rather than assuming the current one.

Quality checks

  • Both versions provided by the user — nothing reconstructed
  • Classification axis stated and appropriate to the document type and the user's side
  • Redline covers every section, including unchanged ones marked as unchanged
  • Housekeeping changes separated from substantive ones
  • Every substantive change classified with a one-line reason
  • Summary or risk-balance sentence addressed explicitly, even if unchanged
  • Authorship, approval, dissents, and effective date recorded, with any change in composition noted
  • Accompanying-numbers section either populated from provided figures or explicitly skipped
  • Verdict commits to a direction and names one most-significant change
  • Text-vs-reaction section uses only provided reaction data, or is marked skipped
  • Closing questions are specific to the diff, not generic

Example prompts

"New and prior versions of the platform's developer terms pasted below. Diff them, classify each change as stricter or looser for us as an API customer, and tell me whether the developer-forum outrage is about something that's actually in the text."

"Both central bank statements pasted — today's and the prior meeting's. Projections are out too, medians attached. Diff them, classify the changes, and tell me if this is hawkish. Front-end yields are up, the dollar is up, and equities are down in the first ten minutes."

"Attached: the regulator's updated guidance and the version it replaces. What expanded, what retreated, and what did they take out of the section on third-party oversight? No reaction data — just the text."

"Vendor sent back a revised MSA draft. Here's what we sent and what came back. Redline it from our side, flag every change to the definitions and the liability cap, and give me the questions for the next call."

Works well with

Pair with weekly-recap-writer when the redline needs to become client- or team-facing commentary, with contract-obligations-extractor when the diff of a contract draft reveals that the full obligations package needs abstracting, and with red-team-premortem when a reading of the change is driving a decision and needs to be attacked before it is acted on.

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.