Build your own lease counsel.

What you'll end up with

Your agent's baseline isn't generic real-estate law — it's your deal history. Before any setup magic, gather the raw material into one folder:

• →Executed leases — every one you can find, both sides of the table.
• →Your marked-up drafts — .docx files with tracked changes are gold. Your redlines are your standard; twenty years of margin comments encode positions you couldn't list from memory.
• →Attorney email threads — export the back-and-forth. The negotiation reasoning lives there.
• →LOIs and proposals — yours and the ones you received.

Don't worry about volume — that's the point. When we built ours, the email export was ~4,000 messages. Nobody read them. Claude scored each one (points for a law-firm domain, points for lease terms in the subject, a point per lease term in the body; keep anything scoring 4+) and distilled the pile to 753 threads that actually mattered, then mined those into positions. A weekend of compute, zero weekends of yours.

# a corpus folder your agent can grow into
lease-counsel/
├── corpus/
│   ├── leases/          # executed leases (.pdf, .docx)
│   ├── redlines/        # your marked-up drafts (tracked changes)
│   ├── emails/          # exported attorney threads (.eml, .mbox)
│   └── lois/            # letters of intent, proposals
└── kb/
    ├── landlord-rules.md   # your standard, one side of the table
    ├── tenant-rules.md     # your standard, other side
    └── sources.md          # where every position came from

One discipline to insist on from day one: every position traces to a source. Have the agent keep a sources file where each rule cites the interview answer, the specific email, or the specific redline it came from — and have it list what it couldn't read (scanned PDFs, ancient .doc files) so you know the gaps instead of assuming coverage.

The corpus shows what you did. The interview captures why — and what you'd never accept, which no document trail records. This is the step people skip, and it's the one that makes the agent yours instead of a generic lease checker. Text or voice both work; a good session is 20–30 minutes of pointed questions and scenarios.

Run separate rounds wearing each hat — as landlord and as tenant — because the entire analysis flips with the role. Expect questions like:

Two governing rules to give the agent, verbatim:

Structure matters more than clever prompting. The pattern that works: a thin persona file that says how to think and when to act, pointing at a deep knowledge base that holds what's true and can grow. For the lease counsel, that looks like:

# the persona is small; the brain is separate and grows
~/.claude/agents/lease-counsel.md    # who it is, two modes, when to ask

lease-counsel/kb/
├── landlord-rules.md      # DRAFT mode: maximize owner protection
├── tenant-rules.md        # REVIEW mode: maximize tenant protection
├── state-law-guardrails.md  # what your state won't let you do
├── negotiation-playbook.md  # ask → pushback → fallback → trade
└── calibration-log.md     # dated interview answers + corrections

Mode detection is explicit: the agent works out from the documents whether you're the landlord or the tenant, and if it's genuinely ambiguous it asks exactly one question — “landlord or tenant?” — because everything downstream flips on the answer.

Then comes the part that feels like cheating: the due-diligence fan-out. On an acquisition, you don't want one agent reading the data room front to back — you want a team working it in parallel, each specialist returning a structured result into one memo:

• →Title & survey reader — easements, encroachments, exceptions worth losing sleep over.
• →Environmental summarizer — digests the Phase I, flags RECs and data gaps.
• →Zoning checker — current use vs. permitted use, nonconforming rights, parking ratios.
• →Lease-abstract extractor — abstracts every lease in the rent roll into one comparable table: term, options, escalations, CAM, co-tenancy, termination rights.
• →Financial normalizer — reconciles the offering memo's numbers against the actual T-12 and rent roll, and tells you where they disagree.

For a demo of the underlying fan-out mechanic (subagents working in parallel), see the agent-teams video on the main guide.

An agent that produces a different deliverable every run is a toy. Give yours a fixed contract — ours produces the same four artifacts on every single lease review:

1. 1A ranked issue memo with a risk rating — each issue quotes the offending clause, or flags it as ABSENT — should be present.
2. 2A paste-ready redline — real tracked changes in the counterparty's actual .docx, not a list of suggestions. You open it in Word and it looks like your lawyer already went through it.
3. 3A negotiation playbook — for each ask: the expected pushback, your fallback, and what you'd trade it for.
4. 4A missing-clause checklist — everything your standard says should be in this lease and isn't.

Then go one step further: hand over the actual spreadsheets and templates you already use — your diligence tracker, your lease abstract form, your LOI and proposal templates. Claude fills them in per deal from the documents. The deliverable stops being “AI output” and becomes your paperwork, completed, waiting for your review.

Where this goes

Once the team works, make it a habit, not a project. Two structures carry it:

Drop-zones per entity. Give each LLC or property its own inbox folder. A lease lands in entities/main-street-llc/inbox/, the review lands next to it in reviews/. New entity, copy the template folder. Your filing system becomes the workflow.

The calibration loop. Every time you override the agent — “no, I concede the guaranty when the tenant posts a bigger security deposit” — the correction gets appended, dated, to the calibration log and folded into the rules. Review #50 is meaningfully sharper than review #1, because it has absorbed 49 of your corrections. That compounding is the entire reason to build this instead of pasting leases into a chatbot.