lease-audit
Skill0 starsChanged 59 days agoMITView on GitHub
Audit a telecom lease document chain — an original ground or rooftop lease plus its amendments, usually as PDFs — and produce the deal file the rest of the toolkit runs on, together with a human-readable audit report. Use when asked to audit, review, or work up a cell site lease, to build a rent history from a lease chain, to extract defined terms from a lease, or to decide whether a site is worth pursuing. Use it again to RE-AUDIT where a missing instrument has since been found, supplying the previous audit as a baseline. Stage 1 of the lease negotiation pipeline.
In this repository4 other files it gives its agents.
- amendment-draftingSkill
- deal-intakeSkill
- negotiation-playbookSkill
- redline-reviewSkill
The file
Show its rules and text
We split this skill into the rules it gives an agent
17 rules. About 0 are common, 8 are distinctive, and 9 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈8
Distinctive · estimate
This author's own choices. Worth a look.
9
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 8
Read a lease and its amendments.
- Very specific: line 21
ORDER MATTERS, and the deal file comes first: Write it, then hand it to a human for review in review.html — which lists your low-confidence …
- Distinctive: line 84
Read these. They are the contract you are writing to:
- Very specific: line 118
Was the found instrument already known to be missing?: If it was on audit.sources[] as extraction: unavailable, the first audit knew it …
- Very specific: line 135
Never invent a value: If a page is missing, illegible, or the document is silent, that is a finding — absent or ambiguous with a note.
- Very specific: line 144
Distinguish stated from computed: Every rent figure is one or the other.
- Distinctive: line 196
Do not assume the obvious name.
- Distinctive: line 243
The escalation anniversary is not the amendment date: If an amendment resets the rent on 1 October but leaves the escalation clause intact, …
- Distinctive: line 250
A replacement escalator does not stack on the old one: When an amendment deletes and replaces an escalation clause, the new rate …
- Very specific: line 281
Leave history.payments empty unless a payment record was supplied.
- Distinctive: line 297
Read from the tenant side — the library is scored that way.
- Very specific: line 330
The report: Follow fixtures/lease-chain/expected-audit.md:
- Distinctive: line 332
Document chain — stated titles and chronological order side by side
- Distinctive: line 362
Run against it and compare.
- Very specific: line 370
When built, the method is fixed: run src/lease-engine.js over rentSchedule to produce the expected schedule and diff against actuals.
- Very specific: line 375
Database corroboration: No access yet. When it arrives, records join audit.sources[] as a corroborating source.
- Very specific: line 405
A re-audit of a reviewed file inherits its corrections: Read audit.corrections[] before you re-read the documents: a reviewer who found …
≈ “Common” is an estimate: at least 1 in 5 of the 372,429 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.
Files
Lease audit
Read a lease and its amendments. Establish what the documents actually say, decide whether the site is worth pursuing, and emit two things:
| Output | What it is |
<site>.deal.json | The deal file. Authoritative. Every downstream tool reads it. |
<site>-audit.md | The report a human reviews. Prose, with the arithmetic shown. |
| The financial model | financials.html reads the deal file and produces it — a summary a director reads, the schedule from commencement to end of life, what each instrument did to the money, and what does not add up. Exports .xlsx for finance to reconcile a ledger against. |
| A completed audit form | Where leadership asked for one. The form is an input, not something you design — whoever asks the question sets its shape, and the negotiator has no say in it. You fill it from the deal file, every answer traced to its instrument and section. tools/auditform.html does it deterministically and is the definition of how a field resolves; forms/standard-audit-form.json is a shipped example, there so the toolkit has all its pieces, and is meant to be replaced by the real one. |
The first two, every time. The JSON is what the tools consume; the report is what makes the JSON reviewable; neither substitutes for the other. The last two are generated on demand from the first.
ORDER MATTERS, and the deal file comes first. Write it, then hand it to a human for review in
review.html — which lists your low-confidence readings, anything you took off a degraded scan, and
the arithmetic everything downstream compounds — and only then write the report and generate the
spreadsheet and the form.
The reason is that only the deal file is authoritative. The spreadsheet and the form are derived and
regenerate in seconds, so a correction costs nothing there. The report does not. It is prose you
wrote, nothing regenerates it, and a figure corrected after you wrote it leaves the narrative
disagreeing with the file it describes — silently, because no check compares them. Writing it last
costs you nothing and removes that entirely.
The financial model only works if you fill rent.components[] — see step 3. The form is only
produced where somebody asked for one.
A form is an input as well as an output. Where a form arrives with the instruction, read it first: its fields tell you which subjects the audit must establish before anything else, and a question it asks that the chain cannot answer is a document to go and find rather than a field to leave empty. Where the form is prose, a Word file or a PDF rather than a form definition, answer it in its own shape — the JSON form is how the toolkit fills one mechanically, not the only form that counts.
The most common subjects are term, rent, escalator, termination rights, sublease and assignment,
access, and utilities, which is what the shipped example covers. Every one of them resolves from
audit.terms, so record those entries carefully even when the full report would have covered the
same ground in prose.
Where an approved template was supplied
Measure against it, not against the clause library. The library is sixty-two clauses, every one
status: draft and provenance: researched. A gap against it means the lease differs from language
somebody wrote for us; a gap against the approved form means it differs from what the business has
agreed it will sign, which is a far stronger thing to put in front of a landlord.
template.html does the mechanical part and writes audit.baseline. Record it, because nothing
downstream can otherwise tell which kind of gap it is reading.
The lease controls, and the template's vocabulary is not a standard it failed to meet. A form saying "Landlord" where the lease says "Lessor", or "Premises" where it says "Leased Premises", is two documents using different words. The one that governs is the lease, and editing borrowed language to the lease's vocabulary is ordinary drafting rather than a finding. Do not report it as one.
What is worth reporting is a term of this lease that is not fit to rely on: never defined, two terms in force that do not reach the same set, redefined with nothing written down, or recorded without a section anybody can check. That would be the finding with no template in the room.
Where such a term has to be defined or redefined, then the form matters: it is language the
business has already approved, which makes it the obvious wording to adopt. audit.baseline
records those terms and the form's word for each, and nothing else — a word that merely differs is
not recorded, because recording it would send a drafter to change words that do not need changing.
A template will usually arrive as a Word document or a PDF, and it will have blanks in it. It is a form, so rent, dates, party names and notice addresses are missing on purpose. Do not read a blank as the form's position: "rent of $______ per month" says the form leaves rent to be filled, not that the form is silent on rent. The blanks are themselves useful — each is a value somebody has to supply before an amendment can be drafted from the form.
A PDF of a form is a print of a document that exists in Word. Ask for the Word original; a scan has no text to read at all, and a text layer without a Unicode map extracts as characters that are not the ones on the page. Where only the PDF exists, compare by reading it and say so in the report rather than recording a baseline nothing checked.
Before anything else
Read these. They are the contract you are writing to:
schema/deal.schema.json— the deal file format. Field descriptions carry the reasoning.clauses/clauses.json— the clause library and itstenantBenefitscores.docs/audit-terms.md— the term registry and what triage means.fixtures/lease-chain/expected-audit.md— a worked example of the report format. Follow its shape.
And where they were supplied with the instruction: the approved template (see below), the audit form to be completed, and the previous audit where this is a re-audit.
Re-auditing
There is no separate procedure. A re-audit is this one, run again over a fuller set of documents. What the previous audit adds is a single check, and it is worth more than it looks: account for every difference between the two.
Diff the old deal file against the new one and put each change in one of three groups.
Explained by the new document. A rent row moved because the found amendment reset the rent; an option count changed because it added periods. Expected, and the reason to have re-audited. Say so.
An improvement in what is known. A low confidence became high; an audit.sources[] entry
stops being extraction: unavailable; an ambiguous term is now found. This is the measure of
what the found document actually bought, and it belongs in the report because somebody asked for
the re-audit and is entitled to know whether it was worth it.
Unexplained. A figure, term or finding that moved and the new document does not account for. This is the one to chase. It means one of the two audits is wrong, and nobody would ever notice without the previous one in front of them — the new audit reads as complete on its own. Do not quietly adopt the new value because it is newer. Work out which is right, say which, and say why.
Two further things to record:
- Was the found instrument already known to be missing? If it was on
audit.sources[]asextraction: unavailable, the first audit knew it existed and said so, and the chain was assembled correctly. If it was not, the first audit did not know — which says something about how the stack was gathered and is worth a line. - Does the triage change? A found instrument that removes a critical gap, or adds one, changes the recommendation. Restate it either way rather than leaving the reader to compare two verdicts.
Carry forward audit.baseline, and re-run template.html where a new approved form is the reason
for the re-audit. A new form or a new audit form on their own are not re-audits: the lease has
not changed, only the baseline or the question has, and template.html and
tools/auditform.html answer those against the existing deal file.
Hard rules
These are not style preferences. Each one exists because breaking it produces a confident wrong answer that survives review.
Never invent a value. If a page is missing, illegible, or the document is silent, that is a
finding — absent or ambiguous with a note. It is never a reason to supply what the document
"probably" says. An audit that says unknown is complete; one that guesses is worse than useless
because it looks finished.
Never carry a superseded value forward through an illegible amendment. If an amendment changes
a figure and you cannot read the new one, the answer is ambiguous at confidence: low — not
the old figure. This is the most common failure and it is silent.
Distinguish stated from computed. Every rent figure is one or the other. stated means it
appears in a document. computed means you derived it by applying an escalator. Long computed
chains inherit every assumption in the derivation; label them and say so.
Show the arithmetic. The report must let a reviewer check the rent history by hand. Give the
closed-form check (1,200 × 1.03⁵ = 1,391.1289), not just the answer.
Record how reliably you read it. confidence: high | medium | low on findings, and
extraction / legibility on sources. A degraded scan is where you are most likely to produce a
plausible wrong reading. Marking it low is not failure — it tells the reviewer where to look.
And where you could not read a word, say so IN THE TEXT. [?] for a word or figure,
a bracketed note for a passage, a banner for a page — the convention is in
docs/document-format.md, and conformed.html refuses to present a copy as finished while any
source carries one. Never substitute a plausible figure. A gap stops a reader and a guess
does not, and the notice period nobody could read is the number the whole option calendar
counts back from.
And say which instrument each source IS, with documentRef on every supplied audit.sources[]
entry. ref is a filename and the chain uses ids; without the link between them, a finding read off a
half-legible page cannot be identified as one. On the Example Hill chain that link moved three
findings from high into the review queue — they had been read confidently, from the one document
with an illegible page, and nothing said so. An unavailable source may leave it blank: there is no
reading to qualify.
Never write client content into the repository. Deal files, quoted lease text and source PDFs
are confidential. Write outputs to deals/ (gitignored) or a path outside the repo. fixtures/
is for invented material only. Set confidential: true on any real deal file.
Do not judge the market. Triage financials is unknown unless you have been given market
evidence. Stage 2 (market.html) establishes that. Guessing "above market" from a rent figure
invents the finding the market test exists to produce.
Procedure
1. Inventory the documents
Build documentChain[] and audit.sources[].
Record the stated title and the chronological position separately. They disagree in practice — a document titling itself "First Amendment" may be chronologically second, and more than one document in a chain may claim the same title. Order by date, not by title. Every later cross-reference depends on getting this right.
Note anything referenced in the chain but not supplied — a recital naming an amendment you do not
have is a gap in the audit, recorded as extraction: unavailable.
2. Extract defined terms
Fill definedTerms with provenance: the term as actually used, which document defined it,
which section, and whether it exists at all.
Do not assume the obvious name. Rent is frequently not called "Rent" — "Lease Fee", "Monthly Rent", "Rental" are all common, and introducing "Rent" into an amendment when the lease says something else creates two payment obligations on one document.
Mark a term absent when the lease never defines it. This matters more than it looks: library
clauses use {{PropertyName}} and {{EquipmentName}} without defining them, so if the lease
never defined them, those clauses refer to nothing. boilerplate.capitalized-terms does not rescue
this — it points at the Agreement, and there is nothing there to point at.
Where the chain has TWO terms in force for one subject, record both. An amendment that
introduces a second name without deleting the first leaves both operative, and definedTerms has
one slot per subject — so put the term the next amendment should use in value and the others in
coexisting[], each with definedIn, section, usedIn (the sections and instruments still
using it) and relationship.
relationship is the substantive question, not a label. co-extensive means the two describe the
same thing and choosing between them is cosmetic. narrower, broader or unclear mean the
provisions are split between terms that reach different sets, and choosing is a drafting decision.
Example Hill: the original defines the "Communications Facility" as a category — tower, antennas, shelters, cabling, fencing and related improvements. The First Amendment defines the "Lessee Facilities" as an enumeration — everything on Exhibit C-1, exhaustive — and does not delete the first. Utilities, relocation, taxes, termination and removal still use the original term. The obligation to remove therefore reaches a wider set than the right to install, and a battery cabinet consented to in 2011 sits on neither exhibit.
Discarding one of the two because the schema has one slot is the failure this prevents: the discarded term is still what five sections of the lease say.
3. Build the rent history
history.rentSchedule[], forward from commencement. The careful part. Work through the chain in
date order and apply each instrument in turn.
Then fill rent.components[], and do not skip it. The schedule records a TOTAL for each period
and describes the split in prose, which is right for an audit trail and useless for arithmetic:
nothing can carry a total forward without knowing what it is made of. One entry per recurring
charge — name, current amount, rate, anniversary as MM-DD, whether it countsAsRent — and the
financial model projects from it to the end of the lease. Without it the model reports history and
refuses to project, which is the honest failure but not the useful one.
A charge the lease says is not rent still belongs here with countsAsRent: false. It is money
leaving the business, so a ledger shows it; it is not rent, so it is out of escalation and out of
anything a percentage or co-location share is calculated on.
Five things go wrong here:
The escalation anniversary is not the amendment date. If an amendment resets the rent on 1 October but leaves the escalation clause intact, escalation continues on the original anniversary. Escalating from the amendment date puts every later figure wrong.
A premises expansion is not a rent event. Additional area at no additional fee changes nothing in the schedule. Only create a row where the payable amount actually changes.
A replacement escalator does not stack on the old one. When an amendment deletes and replaces an escalation clause, the new rate supersedes. Applying both compounds an error that grows for the rest of the term.
Options accumulate unless the document says otherwise. An amendment adding two extension terms to four existing ones gives six, not two.
Rounding is usually unspecified. Rounding each step and compounding unrounded give different
answers over a long chain. State which you used in history.reconciliation.method. Where the
difference is material, say so.
A lease past its last renewal does not stop having a rent history. If every option has been
exercised or elapsed and the chain supplies nothing further, the schedule does not end at the last
contractual date — the site is still there and money is still moving. Continue the schedule with
reason: "holdover" and say in the note what it rests on, then record the holdover terms as the
term.holdover entry in audit.terms — a registry id that is a KEY in that open map, not a path
under term, which is closed — and carry them into termsAnalysis as a risk. A lease with no
holdover provision at all is the case that matters: nothing states what governs the tenancy, and
that is a finding, not a gap in the audit.
Two things follow, and both are easy to get wrong:
- Whether the escalator still runs is a reading of the document, not an assumption you may
make silently. Some leases hold the last rate; some keep escalating; most say nothing. Say
which you applied and why, and mark every holdover row
computed. - Do not model holdover in the comparator. It already handles this: give it the last
contractual date as
current.termEndwithbaselineMode: continue, and it flags every payment after that date asassumed— the do-nothing baseline then reads as resting on a continuation assumption rather than an obligation, which is exactly what it is. The audit and the history are where holdover is recorded; the comparator needs nothing added for it.
Leave history.payments empty unless a payment record was supplied. Its absence means the
over/under-payment question is unanswered, not answered "no". Never populate it from the
contractual schedule — that would make reconciliation compare a thing to itself.
4. Record the terms
Fill audit.terms{} against the registry in docs/audit-terms.md, each with status,
confidence, documentRef, sectionRef, and verbatim where a reviewer would want the exact
words. Unknown ids warn rather than fail — add one if a lease addresses something the registry
does not cover, and say so in the report.
5. Analyse the terms against the clause library
audit.termsAnalysis. For each library clause that could apply to this site, compare what the
lease has.
Read from the tenant side — the library is scored that way. Site facts move the score: a rooftop clause is irrelevant on a tower, and interference matters far more on a shared site.
Two states are easy to miss and both matter:
present-better. The lease beats the library. You are looking for gaps, so you will find
gaps; a provision more favourable than the library's own position is a thing to defend, and
raising it in negotiation risks losing it. Flag these explicitly and consider
priority: do-not-raise.
unclear. The document is ambiguous. That is a finding, not a blank.
The gap is the case. A lease missing several clauses scored +3 is worth pursuing on terms
even where the rent is at market, which is why triage treats terms and financials independently.
6. Triage
audit.triage. The audit's first job is deciding whether stages 2–8 happen at all, and most
sites should stop here. A lease at market with clean terms is a completed audit, not a failed
one.
financials—unknownunless given market evidence.terms—problematicon its own is sufficient to pursue.recommendation—pursue,monitor, orno-action.monitormeans nothing now but something later, usually an NNR or option deadline approaching.rationale— always. A recommendation nobody can review is not a recommendation.
7. Emit both outputs
The deal file. Validate it: every enum against the schema, every date real (2027-13-45
parses into a real Date and is still wrong), current.baseRent equal to the last rentSchedule
row, finalTermEnd not before currentTermEnd. Set stage: "audit" and confidential: true.
The report. Follow fixtures/lease-chain/expected-audit.md:
- Document chain — stated titles and chronological order side by side
- Rent history — the table, with the arithmetic checks
- Defined terms — with the absent ones called out
- Terms analysis — the gaps, and anything the lease does better
- Low-confidence findings — what you could not read, and what it affects
- Term and options
- Triage and rationale
Lead the report with anything that changes the recommendation. A reviewer should be able to stop after the first screen and know whether to proceed.
Before you finish
- Every rent figure labelled
statedorcomputed, with the arithmetic shown - Chronological order derived from dates, not titles
- Defined terms carry provenance; absent ones marked absent
- Nothing illegible silently replaced by a superseded value
- If the last term has elapsed, the schedule continues with
reason: "holdover"and the holdover terms — or their absence — are recorded as a finding financialsisunknownunless market evidence was suppliedpresent-betterfindings identified, not just gaps- Deal file validates;
confidential: true; written outside the repo or todeals/ - No client content anywhere under version control
- Re-audit only: every difference from the previous audit accounted for — explained by the new document, an improvement in what is known, or unexplained and chased down. An unexplained change adopted silently means one of the two audits is wrong and nobody has been told
Testing
fixtures/lease-chain/ is a synthetic chain with a known-correct answer in expected-audit.md.
Run against it and compare. Differences are either an agent bug or a fixture that needs
correcting — both worth knowing.
It exercises reasoning, not extraction: the files are text, and production input is PDF.
Not built yet
Payment reconciliation. history.payments and history.reconciliation are shaped but unused.
When built, the method is fixed: run src/lease-engine.js over rentSchedule to produce the
expected schedule and diff against actuals. The same engine that prices proposals — a second
implementation of escalation would disagree with the first, and the disagreement would be mistaken
for a variance.
Database corroboration. No access yet. When it arrives, records join audit.sources[] as a
corroborating source. An audit performed from documents alone is complete, not provisional — the
database confirms or contradicts, it does not supply what the documents already say.
A mechanical diff of two deal files. Re-auditing compares the previous audit by reading. A tool that listed what moved between two deal files would let the reviewer work from a list rather than from two documents side by side, and would make the unexplained-change check hard to skip. A convenience rather than a gap: everything it would surface is readable today.
What the review writes back, and what you should expect to see
Stage 1.5 runs review.html over your work. It reads three fields you set and
writes four you should recognise when a reviewed file comes back to you:
| Field | Who writes it | What it means |
confidence on a finding | you | how reliably you READ it — the review orders the queue by this |
verbatim on a finding | you | the exact words, so a reviewer checks without reopening the PDF |
audit.sources[].documentRef | you | which chain instrument each source IS |
audit.reviewedBy / reviewedOn | the reviewer | an agent-produced audit is not finished until this is filled |
audit.confirmed[] | the reviewer | readings a person checked and agreed with |
audit.corrections[] | the reviewer | readings a person changed, each keeping what you read |
A low you were honest about is cheaper than a high a reviewer has to catch.
The queue is ordered low → degraded source → ambiguous → medium, so your
own confidence is what decides where a human looks first — and documentRef is
what lets a finding read off a half-legible page be identified as one at all.
Without it, ref is a filename and the chain uses ids, and the link reaches
nobody.
A re-audit of a reviewed file inherits its corrections. Read
audit.corrections[] before you re-read the documents: a reviewer who found
page 2 legible in the recorded copy has told you something you could not see,
and re-reading the bad scan and overwriting them silently is the one way this
gets worse rather than better.
When a document arrives later
You are not the only writer. An assignment found after you ran, legal's answer
to an open question, finance's rent roll — those go through deal-intake,
which records them as changes rather than corrections where the world moved
rather than you misreading.
Where such a document changes the rent history, the term arithmetic or the
option count, the right answer is a re-audit with the current file as the
baseline, not a patch. deal-intake will say so; act on it.
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
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.

