Skills I use, free to install.

Two layers of bad habit go. The first is classical bloat, meaning passive voice, abstract subjects and Latinate padding, which is the Orwell and Gowers tradition. The second is the recognisable AI dialect: banned vocabulary, the reflex rule of three, hedge stacks, the summary that opens by restating the question.

What separates it from the many prompts that do this is that it has been reconciled with a working house style rather than imposed over one. Four upstream rules collided with standing instructions, so the skill says which four, what happened to each, and why. A rule that contradicts your own style guide does not get followed, it gets followed inconsistently, which is worse than not having it.

Three modes: audit flags and suggests, rewrite returns clean prose, edit fixes a named file in place.

You have been reading this sentence all year: 86 editing jobs, four Word add-ins, the same instruction to each, six binary checks per run.

Four facts, four different roles. A sample size, the things tested, a control, and how the results were judged. The sentence flattens all of it into a list and leaves you to work out which is which. It has not removed the information, it has removed the instructions for assembling it, which is exactly why it reads as confident and tells you so little. The rhetorical name for the habit is asyndetic parataxis, meaning the conjunctions are dropped and every clause is laid side by side as an equal.

Fixing it needs something a model can actually check itself against, because "write more clearly" is not an instruction it can act on. This gives it seven identifiable features, the strongest being whether the sentence could be turned back into several complete clauses simply by restoring the relationships between its parts. It also says what NOT to flag, so a real list stays a list and a headline stays a headline.

Three modes: audit flags and suggests, rewrite returns repaired prose, edit fixes a named file in place.

Ordinary research treats the report as the deliverable, so the evidence dies with the session and the whole thing has to be trusted at once. This inverts it. The folder is built first, the agents write into it, and the report is derived from it afterward.

It fans out by discovery route rather than by topic, because routes have different blind spots while topics share one. It saves every source whole before writing a word about it. And it has a separate agent open each number’s own URL and try to refute it, because the finder has already decided the claim is true, so asking the finder to check it is asking somebody to disagree with themselves.

You have a draft and you have the file: the executed leases and amendments, the rent roll, the T-12 and prior-year operating statements, the PCA, the Phase I, the survey, the PSA, and the engagement letter. Somewhere between reading all of that and writing the report, a number changed, a date slipped, or a sentence came forward from the last job on the same street.

Point this at both. It reads the draft and sorts every factual statement into three piles. What ties out to a document in the file, cited. What conflicts with one, with both passages quoted so you can see which is right. And what has nothing behind it, which is the pile that matters, because a conclusion your workfile cannot support is still your problem long after the fee is paid.

It tells you what is missing from the file before it tells you anything else, so a clean result never just means it had nothing to check against. And it will not take the rent roll's word for a lease term. A rent roll is somebody's summary, so the lease is the evidence and the rent roll becomes a second thing to check.

It does not touch your numbers, your adjustments or your opinion of value. It finds the disagreement and hands it to you.

A clause gets pasted in from another deal and brings its font with it. A heading is retyped and loses its spacing. One run of a defined term ends up a point smaller than its neighbours. None of it is visible while you read for meaning, and all of it is visible to whoever receives the document.

The part that makes this work is unglamorous: it reads the styles before it reads the text. A .docx is a zip, and almost all the formatting lives in styles.xml rather than beside the words. Measured on a plain fourteen-paragraph agreement, the text part carried exactly one font size, one colour, one spacing and one indent, while the styles part carried 26 sizes, 267 colours, 197 spacings and 164 styles linked by 158 inheritance references. An agent that reads only the text part finds nothing and tells you the document is clean, which is the worst answer available because it is indistinguishable from a careful pass.

So it resolves each paragraph's effective formatting down the chain, from the document defaults through the style and its parents to anything set directly, and only then compares like with like. On a test document with four deliberate defects planted in it, following this method found all four, including a heading whose spacing only reads as an outlier once you know the other headings inherit 10 pt from a style that never appears beside them.

It also knows what to leave alone. Italics on a case name and a block quotation set a point smaller are choices, not drift, and two examples at a level are too few to call either one wrong.

A federal solicitation carries its own rulebook by reference, and the parts that go wrong are the ordinary ones: who may actually bind the Government, whether SAM registration and the Unique Entity ID are stated, whether the commercial-item provisions 52.212-1 through 52.212-5 are all there or only some of them, whether a set-aside names its NAICS code and size standard, and whether the due date in one place matches the due date in another.

The FAR text ships inside the skill. Parts 1, 2 and 12, the 52.212, 52.204 and 52.219 clause families, and the Revolutionary FAR Overhaul deviation guide are all bundled, so the check quotes the rule instead of recalling it. The dollar thresholds come out of 2.101 rather than out of memory, because a superseded figure is the commonest error in this area.

It also says what it did not check. FAR Part 19 is named rather than carried, so a question turning on set-aside procedure comes back as unchecked with the source named.

Federal award money brings its own procurement rules with it, down every tier of subaward until a document says it stops. The questions that decide a finding are unglamorous: is this vendor a contractor or a subrecipient, which of the eleven procurement sections applies, does the contract actually carry the Appendix II provisions, and has anyone checked the exclusion lists.

The regulation ships inside the skill. Subparts A through D, the procurement standards 200.317 to 200.327, Appendix II, and the government-wide award terms in parts 25, 170, 175, 180, 182, 183 and 184 are all bundled. So is 2 CFR 1201, which is the one that catches people, because DOT money removes the state exception at 200.317.

Subparts E and F, the cost principles and the single audit, are not bundled, and a question that turns on either comes back named as unchecked.

The cyber clauses reach almost every defense acquisition that is not solely commercially available off-the-shelf, and the trigger is the data rather than the money, so 252.204-7012 has no dollar threshold at all. The CMMC clauses add a money element on top. Getting this wrong is expensive in both directions, because an unnecessary CMMC level costs the bidder and a missing clause costs the buyer.

The clause and requirement text ships inside the skill, including 252.204-7012, the assessment and CMMC clauses, DFARS subparts 204.73 and 204.75, FAR 52.204-21, 32 CFR 170 and both revisions of NIST SP 800-171.

Both revisions are bundled deliberately, because confusing them is the commonest error here. NIST has withdrawn Revision 2 and serves it as an archived document, yet Revision 2 is still the operative version for CMMC, since 32 CFR 170.2 incorporates it by reference. So Revision 2 answers what an assessor will check and Revision 3 answers what NIST currently says, and the check names which one it is quoting every time.

An airport sponsor that accepts an Airport Improvement Program grant takes on obligations lasting the useful life of the facility, up to twenty years, with no time limit at all on the exclusive-rights and revenue assurances. The trigger is the money rather than the fact of being an airport, so the first thing this establishes is whether federal money is actually in the transaction.

The assurance and provision text ships inside the skill, meaning all forty Airport Sponsor Assurances, the required contract provisions for AIP contracts with the applicability of each, 49 USC 47107 behind them, and AC 150/5100-14E for qualifications-based consultant selection. The 2 CFR 200 procurement standards are bundled too, because grant money brings those as well.

The DBE and ACDBE programmes at 49 CFR 23 and 26 are not bundled, and neither is the AIP Handbook, so a point turning on a DBE goal calculation comes back named as unchecked with its source.

Chapter 287 Part I binds a unit of the executive branch of Florida state government, and it expressly excludes the universities and colleges. A county, city, school district or independent special district is not an agency under Part I at all, and only a few sections reach it, chiefly the Consultants' Competitive Negotiation Act, public entity crimes and scrutinized companies. So the first thing this check settles is whether the chapter binds the issuer, because applying the Part I thresholds to a body they do not reach is the error that wastes the most time.

The statute and rule text ships inside the skill, meaning the whole of Chapter 287 split in two at s. 287.059, the Department of Management Services purchasing rule at F.A.C. 60A-1 which sets the procedure the statute only requires, the s. 120.57(3) protest clock and bond, and the public meetings and records provisions.

The Legislature amends this chapter most years, so every file carries its edition and the check says when a bundled copy may be behind the publisher.

A scrivener’s error is a mistake in the writing rather than in the deal. Nobody intended “Thirty (60) days”. It survives review because the people reading are reading for meaning, and meaning is exactly what carries you past a wrong numeral.

It runs as two passes, because doing both at once produces a long list of plausible things and a reviewer who stops reading. Detection scans for seven specific error types. Adjudication then decides which are real, assigns one of five priority levels, and writes the corrected text only where the fix is unambiguous.

The parts that took the longest to get right are the guards. An importance gate before anything is called high or critical, because a reader who finds one inflated “critical” discounts every finding after it. A rule that unfilled placeholders are never high, since a blank is a to-do rather than a defect. And a self-check that catches the failure this job actually has, which is a correction that quietly fixes a second thing: spacing elsewhere in the paragraph, a hyphenation, a capitalisation nobody asked about.

Definitions drift because they are edited in one place and used in forty. A term gets renamed in the definitions section, a schedule arrives from another deal, a recital written before the definition existed uses the term anyway. None of it is visible while reading for meaning.

Two things separate this from asking a model to “check the definitions”. First, a search that comes back empty is not evidence of absence, so nothing is called undefined until its plural, possessive, acronym and collective forms have been checked too. Second, an explicit list of what never counts: proper names, statutes, headings, schedule labels, and words capitalised only because they start a sentence.

It also carries a completeness section, which exists because every rule guarding against a false positive pushes toward saying nothing. A check that reports three findings in a sixty-page contract has usually been silent rather than careful, and only one of those two failures looks like rigour.

A rent roll is somebody else's summary of the leases, prepared by the party with the most to gain from it reading well. The income approach, the underwriting and the offer all sit on top of it.

This goes to the executed documents. It orders each tenancy's amendments so the surviving term is the one that controls, pulls rent steps, escalations, abatements, expense structure, options and everything else with a number attached, and cites the clause for anything a reader might dispute.

Then it reconciles against the rent roll line by line and reports the direction of every difference, because a pattern across several tenancies is the finding rather than any single row.