Daily Brief
_01Opens every cockpit with the day framed in money: the numbers computed in SQL, the model writing only the bullets. Explicit permission to skip included.
Supernova's agents are workers, not chat windows: they read the real corpus (proposals, signed agreements, threads, recorded calls), save structured artifacts with verbatim provenance quotes, and write every move to a ledger you can open: model calls, tool calls, tokens, cost. An agent that cannot show its work does not ship here.
THE SALES FLOOR
Cheap, fast drafts on every desk. Propose, never act: apart from two audited, human-clicked write-backs, nothing touches the CRM or an inbox on its own.
Opens every cockpit with the day framed in money: the numbers computed in SQL, the model writing only the bullets. Explicit permission to skip included.
Drafts the reply grounded in the real thread: a tool loop over the deal's emails, meetings, and recorded calls, then a fresh-context critic checks every claim before a human sees it. Nothing sends without you.
Writes the deal note from your last recorded call: what moved, what was said, what you owe them next.
Scores MEDDPICC with the evidence that earned each score: it reads the actual emails and transcripts through its tools, then quotes who said it and where. Never inflates.
Who is in the room, where the deal stands, what to ask. On your desk before the call, not after.
Extracts what each side promised from the thread and the calls, with provenance, and keeps it in front of you until it is kept or dismissed.
Researches the company on the live web: news, size, key people, buying signals. The first agent trusted with a real CRM write-back, one deliberate click, fully audited.
THE DELIVERY LINE
The engines behind the nine-stage ratchet: requirements, personas, the data model, roles, design, pages, mocks, tests, the stack. Every artifact versioned, every claim quoting the corpus, every gate held by a named human.
Pages through the full proposal and signed agreement with read tools, searches, dedupes against what exists, then saves each use case with its stories and tests in one transaction. A verifier rejects any requirement whose quote is not found in the corpus.
Researches the client product's real cast: users, buyers, approvers. Pains marked EVIDENCED when the corpus says so, RESEARCHED when the web filled the silence. Never your own staff dressed up as users.
Derives the product's data model from approved use cases: entities, fields, PII flags, state machines that cite the business rule gating each transition.
What the product grants: per-role permissions in plain language, naming the ontology entities they touch, personas mapped to each.
Every page the product needs: purpose, owner role, primary actions, the entities shown, the stories served. Consolidates ruthlessly; an honest small map beats a page per whim.
Codifies the client product's design language: a DESIGN.md humans read and scoped tokens machines render, previewed on fifty real components before anyone commits.
Renders one working mock per page definition, wearing the approved design tokens. The customer sees their product before a line of production code exists.
Writes the Playwright specs that prove the acceptance criteria, shipped into the project repo so the gates can one day turn green on their own.
Reads your feedback and answers two questions before anything mutates: what did you mean, and what does it touch. Understanding plus blast radius, then you decide.
Proposes the delivery stack layer by layer with a rationale a client could read, quoting the corpus verbatim where it names a technology. Approved by the tech lead; shipped into the repo as TECH-STACK.md.
Agents propose; named humans approve at every gate. Every run is a session on the ledger: what it read, what it saved, what it cost. Two agents still in the shop: Meeting Minutes and What's Next.