Iris
Iris is Zimac's Communications Expert. She owns the loop from intent → message → experiment → evidence → next iteration: find the one promise, cut whatever obscures it, carry it across channels without losing its identity, then decide whether the result deserves to ship. Her standard is simple: focus the signal, cut the drag, prove the result.
Exact, cultured, unsentimental — and willing to be a little daring. Iris keeps the author's nerve while cutting fog, committee language, and borrowed hype. She treats frameworks as diagnostics, not costumes, and measures claims an LLM should never eyeball: readability, significance, attribution, and economics are computed in code. Her answer is a decision with evidence, not a cloud of tasteful adjectives.
❂ Aperture
Locks audience, action, and the single promise before polishing a word.
❂ The Cut
Removes drag and generic sheen while preserving facts, identity, and the author's nerve.
❂ Prism
Adapts one message across email, chat, social, web, and press without identity drift.
❂ Readout
Turns real campaign data and deterministic math into a clear ship, hold, or kill call.
Who Iris is#
Iris combines an editor's taste with an experimentalist's discipline. Give her an announcement, launch page, campaign, or result that needs a verdict and she separates signal from polish and evidence from noise. She leads with the consequential call ("the ask is buried"; "the lift is not significant yet"), shows the score, count, p-value, or structure behind it, and makes the next move obvious. She writes for the audience rather than for applause — and would rather hold an inconclusive test than manufacture certainty.
Her signature moves#
- Aperture — establish the audience, desired action, and single promise. If those three do not resolve, Iris fixes the brief before decorating the copy.
- The Cut — remove drag: hedges, passive construction, clichés, filler, buried verbs, and generic assistant sheen. Facts and voice stay intact.
- Prism — refract one core message into channel-native versions, changing length and cadence without changing the claim.
- Readout — join observed results to deterministic campaign science and return a decision: SHIP, HOLD, or KILL, plus the evidence and next test.
Working with Iris#
Hand Iris a draft, a campaign, or a result that needs a verdict. For words, give her the audience and desired action when you know them. For an experiment, give her the variants, exposure, conversions, and the decision threshold. When one missing fact would materially change the work, she asks one sharp question rather than inventing context.
Writing refinement#
Share a draft and Iris does two distinct things with it. First, a deterministic clarity review: she scores the writing across readability and the habits that dull it — Flesch reading-ease and grade level, sentence-length variance (the rhythm that makes prose readable instead of a wall), passive voice, hedges, clichés, filler, and nominalizations (the buried verbs that make a sentence sound like a policy). Each is counted, not eyeballed, so she can point at the exact phrase and tell you what to cut rather than say it "reads a little heavy."
we believe, somewhat, in general), and one 44-word sentence carrying three ideas. Split it, cut the hedges, and you land around grade 9 — where an all-hands note should be."Second, she is the revision editor in the Review Studio. You read a document and leave margin comments on the passages you want changed — "too long," "soften this," "say what we actually did" — and Iris rewrites only those sections, staying inside the document's own voice and structure. It's an edit, not a reply: she returns revised prose that drops back into place, never a chat message about your prose. She preserves names, numbers, and claims; if requested detail is missing, she leaves an explicit TBD rather than fabricating it.
It should be noted that a degradation in latency was experienced by a subset of users.Iris returns, in the doc's voice:
Some users saw slower responses for about an hour. — same fact, half the words, the passive and the nominalization gone.
Mass communications#
When a message has to reach a lot of people — an announcement, a newsletter, a team broadcast, an exec update — Iris takes the one thing you want to say and shapes it for how each audience will actually receive it. She adapts one message to each channel, with the length and tone the channel expects:
- Email — a subject line that earns the open and a body that front-loads the ask; she'll score the subject line and the CTA and tell you which of two is stronger and why.
- Slack — short, skimmable, one clear action, the tone a team channel expects rather than a press release pasted into chat.
- Blog — room to explain, with the structure a reader can follow and a headline that sets the promise.
- LinkedIn — a hook in the first line, the human angle up top, the detail below the fold.
- Press — tighter, claim-careful, the who/what/why a reader outside the company needs.
She also segments the audience: from people-profile attributes she splits one broadcast into the versions each group should get, so the migration notice that matters to on-call engineers doesn't read the same as the one going to the exec sponsors. She consults Joy for the voice and people profiles that keep the tone consistent and the segments real, and pulls the facts a message needs — dates, numbers, what actually shipped — from Quill's memory rather than approximating them.
Marketing#
For positioning and copy, Iris reads the way a marketer does but backs it with computed structure, not vibes. Paste a landing block, a feature announcement, a one-liner, or a whole page and she runs it through the checks that separate copy that converts from copy that just sounds nice:
- Value-proposition clarity — is there a single, legible promise, or does the reader have to assemble it themselves?
- Feature-vs-benefit balance — how much of the copy describes what it is versus what it does for the reader, with the ratio named.
- Structure detection — whether the copy follows a real persuasive shape (AIDA, PAS) or wanders, and where the arc breaks.
- Claim & hype density — it flags unsubstantiated superlatives ("the most powerful," "seamless," "revolutionary") that a reader discounts on sight, so you swap adjectives for evidence.
- Funnel-stage fit — whether the copy matches where the reader is — awareness, consideration, or decision — because a decision-stage CTA on an awareness-stage reader falls flat.
Because each read is computed from the text, she tells you the hype-density number and the exact superlatives, the feature/benefit split, the stage the copy is actually written for — not a plausible-sounding paraphrase you can't act on.
Campaign science#
This is what turns Iris from an editor into an operator. Marketing decisions hide a lot of math an LLM should never do in its head — significance tests, attribution splits, lifetime-value discounting — so Iris runs them in deterministic code and relays the exact result. Paste the numbers (or pull them from a connected channel below) and ask for a decision.
Experiments — is this a real win?
Give her two variants' visitors and conversions and she runs a two-proportion z-test: the relative lift with a confidence interval, the p-value, and a plain verdict — ship, don't ship, or not yet. When it's not yet, she tells you the sample size and days you still need (from your baseline rate, the minimum lift worth detecting, and your traffic), so "keep running" comes with a finish line. She'll also flag a sample-ratio mismatch — a broken split that quietly invalidates the whole test — and give a Bayesian read (the probability B beats A) when you want the intuitive number instead of the frequentist one.
Funnels, retention & pacing
Hand her the counts at each funnel stage and she computes per-step and cumulative conversion, calls out the bottleneck (the step bleeding the most), and will simulate what fixing it is worth before you invest — "a 20% lift on signup→activated adds 340 paying customers a year." She reads a cohort-retention curve for where it stabilizes and a projected average lifetime, and checks budget pacing mid-flight — on track, or about to burn out early.
Attribution — who actually earned the credit
From conversion paths she compares the five rule-based attribution models side by side (first-touch, last-touch, linear, position-based, time-decay) and names the channels whose credit swings between them — the contested ones. For the defensible answer she computes exact Shapley values, the game-theoretically fair split, so credit doesn't just follow whichever model flatters a channel. And for the honest question — did the campaign cause conversions or just harvest them? — she runs a holdout / geo-lift incrementality test with significance, incremental conversions, and iCPA / iROAS.
Unit economics
She closes the loop on money: customer lifetime value (discounted, from ARPU, margin and churn), CAC payback and the LTV:CAC ratio graded against the standard bars, breakeven and target ROAS from your margin, and a budget allocator that reallocates spend across channels under diminishing returns to maximize modeled revenue. So "should we spend more here?" gets a number, not a shrug.
Email deliverability
Because a campaign that lands in spam never had a chance, Iris analyzes the plumbing: she parses a pasted SPF record for the notorious 10-lookup limit and its all-qualifier, grades a DMARC record's enforcement posture, scores list health against real inbox-provider thresholds (Gmail's 0.3% complaint limit, the 2% bounce bar), and lays out an IP/domain warmup schedule to a target volume. Like Daphne's security tools, these read what you paste — they never probe a live mail server.
Live channel data#
Iris doesn't have to wait for you to paste numbers. Connect an email-marketing platform in Settings and she reads it directly — read-only, so she measures and plans but never sends or edits on your behalf. Then the campaign-science tools run on your actual results instead of a hypothetical.
- Mailchimp — your audiences (size, engagement, list rating), recent campaigns, and the full per-campaign report (opens, clicks, bounces, unsubscribes, abuse reports). Add a Mailchimp API key to the Mailchimp card in Settings.
- HubSpot — your marketing emails, each email's statistics (delivered, opens, clicks, bounces, spam reports and rates), plus recent contacts and lifecycle properties for contextual inspection. Contact lookup is not a total segment-size report. Add a private-app token to the HubSpot Marketing card in Settings.
The payoff is the handoff between the two: Iris pulls a campaign's real send data, then feeds the sent/bounce/complaint counts into her list-health check, or two campaigns' rates into the significance test — a verdict grounded in what actually happened, in one pass. Every pull is fenced as external data and saved to your artifacts trail like any other integration.
March Product Update campaign: 48,200 sent, 0.62% bounce, 0.04% complaints, 31% opens. List health: 100/100 — clean send, no deliverability risk. Its 3.1% click rate beats February's 2.4% at p = 0.008, so the shorter subject line is a real improvement — keep it."Boundaries, studios & lessons#
Iris owns words, audience, and positioning — not the legal read. She is not a legal reviewer: anything contractual, regulatory, or compliance-related she defers plainly rather than approve, and when a claim needs a compliance or security check before it goes out she pulls in Daphne instead of waving it through. You can send a draft into Review Studio with a one-click chip, save recurring output formats, and correct her — "keep exec updates under 150 words", "never use the word seamless" — so the preference becomes a durable lesson. Night Shift can recheck one connected Mailchimp or HubSpot campaign report as its numbers settle; it is not a general web-page copy monitor.