Reference

Prompts, skills and subagents

The five prompts the connector puts in Claude’s menu, printed in full — and the plugin that adds twenty broker skills, four subagents, two commands and a set of guardrails.

The connector publishes five prompts. In Claude they sit in the + menu under Brella OS; in Claude Code they are available as prompts from the server. Each is a starting point for a real piece of broker work, and each is written to read before it writes and to stop for you before anything leaves the agency.

Today's work

todays_work

What needs you today in Brella OS — renewals by runway, overdue items, your reminders — as a short plan.

Using Brella OS, show me what needs me today. Start with get_workspace_context and get_today_work, and list my reminders due today. Give me a short prioritised plan: anything overdue first, renewals by how much runway is left, then the records to complete. For each item say the one next action and which policy it is on. Do not change anything yet.

Prepare a submission

prepare_submission

Take a policy to carrier-ready: close the readiness gaps, then generate the ACORD forms and the Schedule of Values.

Using Brella OS, get "Cypress Creek Self Storage LLC" ready to send to carriers. Run get_readiness first. Fill every gap you can from the record, the documents on the policy and the property data service, and tell me where each fact came from. Ask me for anything only I or the insured can answer — never estimate a limit, a value, a date or a class code. Then generate the ACORD forms and the Schedule of Values, and finish with what is still blank on each document.

Review a renewal

renewal_review

Everything that should happen before a policy renews: exposure changes, claims, the E&O re-ask, documents to request, reminders.

Using Brella OS, review the renewal of "Magnolia Kitchen & Oyster Bar". Read the policy, its activity, claims, documents and coverage advisory record. Tell me what has changed since the last term, whether the coverages need to be re-offered and recorded for the new term, which documents to request from the insured (updated schedule of values, loss runs), and the dates that matter. Propose reminders and any message to the insured, and wait for my go-ahead before sending anything.

Certificate request

certificate_request

Check a certificate holder's requirements against the policy and the recorded endorsement, then issue it.

Using Brella OS, handle a certificate of insurance for "Greenway Property Management LLC (additional insured, primary and non-contributory)" on "Bayou City Apartments LLC". Check the holder's requirements (additional insured, waiver of subrogation, primary and non-contributory) against what the policy carries and the endorsement on record, add the holder if they are new (resolve the address first), and show me exactly what the certificate will say before issuing it. Flag any requirement the policy cannot back.

E&O check

eo_check

The coverage advisory record for a policy this term: what was offered, what was declined, what is undocumented.

Using Brella OS, check the coverage advisory record of "Hill Country Office Plaza LP" for the current term. Tell me which coverages are still unanswered, which declines are missing their reason, and which declines still need a signed rejection on file. Draft what should be recorded, and ask me before recording anything.

The policy names in these prompts are the demo book’s. A prompt sends the text above with the name you give it; nothing about it is fixed, and editing one is expected.

The Brella OS plugin#

The plugin is the connector plus the workflows an agency actually runs, for Claude Code and Cowork. Two commands install it:

claude plugin marketplace add brellapartners/claude-plugin
claude plugin install brella-os@brella-partners

Its source is public, at brellapartners/claude-plugin, so what it tells Claude to do can be read before it is installed. The full setup is on connect Claude.

/brella-os:today: Today’s work in this agency, ranked, as a short plan.
/brella-os:submission: Take a named policy to carrier-ready.

Skills#

Twenty skills, one per broker workflow. A skill is guidance Claude loads when the work matches it — how a submission is built, what a certificate may and may not say, what a renewal needs at ninety days — written so that the tool calls it makes are the right ones in the right order.

Placing business

commercial-submission: Take a lead or policy to a carrier-ready submission: readiness first, then the ACORD forms and the SOV.
property-schedule-sov: Build a property schedule an underwriter will trust, then generate the Schedule of Values from it.
workers-comp-worksheet: Build the comp worksheet behind the ACORD 130, monopolistic and independent-bureau states included.
general-liability-exposure: Record or review general liability: the limits, one hazard row per location, the class code and the basis.
surplus-lines-placement: Place and document E&S business — diligent effort, the home state, the filings that follow.
quote-comparison-proposal: Compare quotes against each other and against the expiring policy, and flag every reduction in coverage.

Servicing the book

renewal-playbook: Run renewals on a 120/90/60/30 clock so nothing lapses and every term is re-underwritten.
policy-change-request: Capture a mid-term change completely and track it to endorsement, through a person.
certificate-of-insurance: Handle certificate requests without implying coverage the policy does not give.
claims-first-notice: Take a first notice of loss to every carrier that might respond, without deciding coverage.
loss-run-analysis: Turn carrier loss runs into an underwriter-ready summary: frequency, severity, aging, trend.
policy-check: Check an issued policy or declarations page against what was bound, before a claim finds the difference.
daily-work: Turn the Today board and your reminders into a short, ranked plan for the day.

Evidence and compliance

coverage-review: Find the exposures an account has that its coverage does not answer, and offer each one on the record.
eo-advisory-record: Keep the coverage advisory record — what was offered each term, and what the insured decided.
eo-safe-drafting: Draft anything a customer will read so it cannot be taken as a coverage promise or a binder.
producer-licensing-compliance: Answer “can this producer do this, here” from the licensing record, with its basis.

Working the tools

using-brella-os: How Brella OS models an agency, and the rules every connector tool call follows.
connect: Walks a member through connecting and confirms who Claude is acting as.
troubleshooting-connection: Diagnoses an authentication, permission or “not in your book” error and names the fix.

Every tool name a skill mentions is pinned to the connector’s registry by a test, so a skill cannot tell Claude to call something that does not exist.

Subagents#

Four subagents, for work that means reading a lot and handing back one table. Three of them read only; the fourth edits a record only with your go-ahead.

book-auditor: An E&O and compliance sweep of the whole book: certificates claiming an endorsement nobody recorded, advisories nobody answered, GL decisions nobody made, licences at risk, surplus lines obligations with no deadline. Reads only.
renewal-strategist: Sweeps the renewal window account by account and returns one table: what changed, what must be re-offered this term, which documents to request, renew or remarket, and the reminders to set. Reads only.
submission-builder: Takes one policy to carrier-ready — readiness, then every gap it can fill from the record, the documents and the public property record — and edits the record only with the person’s go-ahead.
loss-run-analyst: Normalises the loss runs on a policy into one table per policy year per line, then frequency, severity, large losses, open-claim aging and the trend. Reads only.

Guardrails#

Three hooks, which run only on Brella OS tool calls, make no network call of their own, and never deny a call. They ask, and they flag.

confirm-outward: Before any outward or irreversible tool, makes Claude Code ask you — and flags risky coverage language on its way out. Its list is pinned to the connector’s own outward set by a test.
address-cache: Remembers, in a private temporary file, the addresses lookup_address resolved this session.
address-check: Flags an address about to be written to a record that was never looked up. Addresses come from the property data service, not from a person’s typing.

Why an address hook exists

An address typed from memory is how a building ends up on the wrong policy and a certificate goes to the wrong holder. Addresses come from lookup_address, which resolves what somebody said into a real, complete one; the hook flags an address about to be written that never went through it.

Writing your own#

Name the policy. The insured’s name or the BRL- reference. Claude searches the book first, and a reference is unambiguous where a name may not be.
Ask it to read first. “Run the readiness check first” or “read the activity before you answer” costs one call and changes what comes back.
Say what it may not do. “Don’t send anything” or “draft it for me to review” is understood, and the outward tools would have asked anyway.
Never ask it to estimate. A limit, a value, a date or a class code that nobody recorded should come back as a question, not as a number. The skills say so; your prompt can too.

Data handling

Troubleshooting