Your press desk has an agent now.

I. The ordinary world

You have a story. The press doesn't know you exist.

You're a founder. You shipped a product, raised a round, hit a number. The reporters who cover your sector don't have you on their radar. The PR firm quoted you $12,000/month and assigned a 24-year-old. So you do it yourself: cold emails, frantic googling, three drafts of a release that's still wrong on the math.

And the moment a real reporter does email — usually on a Friday with a 6pm deadline — you have ninety minutes to figure out who they are, what they've covered, and whether your answer makes you look like a serious operator or someone in over their head.

II. The call

An inquiry lands. The clock starts.

Who is Kara? What has she written? Which of your portfolio cos fit the story? When did your team last talk to her? Whose intros worked before? You have all of that — somewhere. Across Gmail, HubSpot, three Asana projects, a Slack channel, your CRM, your head.

III. The refusal

This is when most teams ship a half-baked reply.

Or worse: ignore it, lose the relationship, watch the story run without you. Or worst: copy-paste a stock response into ChatGPT, ship something that quotes a number you can't substantiate, and end up with retraction emails for the next quarter.

IV. The mentor arrives

Hi, I'm Penn.

I run your press desk. The inquiry from Kara is already in your Asana inbox by the time you've finished your coffee. So is the dossier on Kara — who she is, what she covers, your team's prior history with her, the best intro path in your network. So is a pre-built shortlist of portfolio companies that match her story.

I drafted three response options. One short, one with a quote, one declining. All three end with an AI-disclosure footer (it's a professional practice, not a confession). None of them claim a number I can't substantiate. None of them name a quote I haven't attributed.

When you approve one, I'll send it through Postmark with an idempotency key, log the dispatch in a tamper-evident audit chain, and update Kara's relationship card. If the send hits any per-domain cap I have, I refuse. If the journalist has opted out, I refuse. If I'm uncertain, I ask.

V. Tests, allies, enemies

What Penn refuses to do.

VI. The reward

What you get back.

• Hours, not days. Inbound inquiries get a content-neutral ack within a minute, a fully-contextualized Asana task within five, and the team typically responds within an hour because the prep work is already done.
• Relationships, not transactions. Penn remembers everything. Every interaction goes into the journalist's card. The next time Kara writes, the relationship-radar dossier knows that you intro'd her to Apptronik in March.
• An evidence trail your counsel will love. Every classifier verdict, every refusal, every send, every retraction — HMAC-chained, R2-object-locked, daily-checkpoint-signed. When a journalist asserts "you said X" and you didn't, you have receipts.
• No accidental blocklist. Per-domain send caps, deliverability monitoring, suppression-list propagation, hard-bounce cooldown. The catastrophic failure is being blocklisted — Penn is paranoid about not making volume the success metric.

VII. The return

How Penn works.

Mail to press@yourcompany.com routes through Postmark and lands at Penn within seconds. A Haiku classifier reads the subject + body, extracts the deadline (Pacific-timezone-aware), and matches the sender against the journalist corpus. Relationship Radar surfaces prior history; pitch.vc surfaces matching portfolio companies. Your Asana inbox receives one task per inquiry with everything inline — urgency-prefixed, deadline-aware, evidence-linked.

For outbound, you draft a release. Penn auto-classifies tier (1/2/3 based on sender / topic / target list). Tier-3 sends require TOTP step-up. Every approved send goes through an atomic transactional Worker — re-validates the approval token, increments the per-domain cap, dispatches via Postmark with idempotency, and emits a signed audit event. Mode B (persona-driven) reaction-predicts journalist response using a corpus-grounded multipov persona before you commit.

Penn is built on the Agent First framework. Every action is a typed tool. Every state change is audited. Every refusal is explainable.