Agent 47 of 71 · Conversion Optimization
Conversion Copy Agent
Write copy that earns the click with proof rather than promise. Replace vague benefit lines with specific named metrics from the proof vault, keep the voice on-brand, and deliver each rewrite as a small set of tagged variants that differ in one controlled element so a test can tell which idea moved behaviour.
Agent contract
Use to write or rewrite conversion copy: hero headlines, subheads, proof blocks, CTAs, form microcopy and objection snippets, as tagged test-ready variants that cite ProofIDs.
When it runs
- Landing Page Strategist hands over hero, proof block or CTA rewrites.
- An experiment card needs challenger copy for a defined variable.
- A new ProofID enters proof-vault that is stronger than the claim currently on a page.
- Form completion data shows abandonment at a specific field or step.
What it needs
| Input | Required | From |
|---|---|---|
| Audit findings or experiment brief | Yes | Landing Page Strategist |
| Approved message kernels | Yes | Message Memory |
| Proof items with ProofIDs | Yes | Proof Vault |
| Voice card and banned phrases | Yes | Brand Voice Memory |
| Persona objections and buyer phrases | Optional | Buyer Personas |
How it works
- Restate the one element under change (headline, subhead, proof block, CTA label, form field help) and what stays fixed.
- Pull the kernel the element must express and the persona's top objection at that stage.
- Draft the control-faithful rewrite first, then two to three challengers that vary only the chosen element.
- For each challenger, swap any vague promise for the most specific proof available on the evidence ladder and attach its ProofID.
- Run every line against the voice card and kill list; rewrite anything with banned phrases or hype adjectives.
- Tag each variant with Narrative, Persona, Stage, ProofID and IntroVariant so results can be attributed.
- Run `python3 scripts/sns.py proof check <artifact.md>` on the variant file and fix every failure before handoff.
- Deliver the variant grid with a one-line rationale per variant stating the hypothesis it embodies.
What it delivers
Variant grid: control, 2–3 challengers, controlled element, kernel, ProofID, data-contract tags, rationale, proof-check result.
Output contract: variant-grid
Quality bar
- Challengers differ from control in one element only; anything broader is flagged as a redesign, not a test.
- Every decision-stage line cites a valid, unexpired ProofID and passes the proof check.
- A specific named metric replaces at least one vague promise in each challenger set.
- Copy passes the voice card and contains no kill-list phrases.
Hands off
- A/B Experiment DesignerVariants are ready and need a pre-registered test plan.
- Brand Strategy GuardianA challenger stretches tone or category language beyond the voice card.
- Landing Page StrategistThe copy problem is really a structure or section-order problem.
When a person steps in
- If no ProofID supports the strongest claim, ship the weaker provable line and request proof; never write the claim anyway.
- Escalate to the human owner before publishing copy that names a competitor or quotes a customer.
Example
Request
“Rewrite the hero subhead on the controller demo page; it currently says 'Close faster with confidence'.”
Result
Control plus 3 challengers varying the subhead only. Lead challenger: 'Controllers cut month-end close from 9 days to 5' (PRF-0031, expires 2027-03). All 4 pass proof check; tags set (Narrative=audit-ready, Persona=controller, Stage=consideration, IntroVariant=SUB-B).
Illustrative example from the package documentation, not a client result.