Build, Refine, Learn: How AI Is Changing How We Ship Product

Shipping a maintenance notice the AI way

Yesterday, our ops team scheduled downtime. Old way: open the visual editor, hunt for a template, tweak copy, mess with spacing, translate, wire the CTA, target it to the right audience and… publish.

The AI way:

1. Prompt: “Create a maintenance banner for July 30, 2–3am ET. Primary CTA ‘Understood’. Use the product theme. Also make a dark‑mode variant and French copy.”
2. AI builds the first draft in seconds: copy, layout, CTA wiring, variants.
3. You polish in the same editor: tweak spacing, swap an image, adjust target segment.
4. Publish & learn: ship, watch engagement, get an auto‑suggested B‑variant to improve CTR.

Same output, radically less thrash. And—critically—it’s reversible and reviewable: everything the agent does is visible as changes you can edit, approve, or roll back.

Why this matters for enterprise teams

• Speed with guardrails: Conversation gets you the 80% draft; the editor ensures brand, accessibility, and review flows.
• Switch, don’t stall: Move chat ↔ editor without losing structure. Start in one, finish in the other.
• Proof over vibes: Measure time‑to‑first‑publish (and to second/third) and show uplift from variant testing.
• Clear constraints win: A bounded set of components beats “anything goes” custom code for reliability, security, and scale.
  

To see how product growth squads are evolving their approach with AI, see our Growth Squad of One. 

What Pendo teams are seeing

“We’re designing for fast, safe creation—conversational to get started, Designer to finish with confidence.” — Alejandro Dao, PM, Pendo

Pendo has been exploring conversational creation with strong guardrails and a clean hand‑off into Designer. The north star: make the first draft easy, keep the last mile accountable. Teams want to:

• Create with natural language, then finalize in Designer with the same components and accessibility defaults.
• Suggest placements and targets using existing tags and analytics (not brittle on‑page heuristics).
• Treat AI as a time‑saver for the parts you’re not great at—writing, layout, localization—while keeping human judgment for what ships.

“AI should shorten the path to a great first draft, not replace the judgment needed to ship responsibly.” — Alejandro Dao, PM, Pendo

How we’re building it at Candu

Notes from Stu, Candu's CTO

We’re shipping this pragmatically, not perfectly—so teams get leverage now and we learn from real usage.

• Break work into tools: instead of a mega‑prompt, the agent calls small tools (create, edit, restyle, localize, place, etc.).
• Stream & verify: each action logs success/fail and retries within constraints. Users can see diffs before publishing.
• Eval as documentation: a lightweight suite checks output shapes and edge cases. Imperfect, but fast and honest.
• One source of truth: everything the agent creates is editable in the same editor—same components, same guardrails, same accessibility.
• Model pragmatism: we favor models that follow rules reliably; “smarter” isn’t helpful if it ignores schemas. Multi‑model is on deck.
• Safety beats vibes: brand themes and accessibility by default; no arbitrary JS; enterprise review paths stay intact.

TL;DR: Leverage AI to get to 80% in a few seconds, keep humans in the loop for the last mile, and use engagement data to decide what to focus on next.

The loop that compounds

1. Prompt a draft → 2) Polish in the editor → 3) Publish with targeting → 4) Observe engagement → 5) Auto‑propose improvements → 6) A/B and roll forward.

Over time, this loop reduces time‑to‑first‑publish (and time‑to‑third), raises baseline quality, and focuses teams on learning velocity.

A practical checklist 

• Start with one clear use case (e.g., maintenance banners, feature announcements).
• Define the allowed components and defaults (brand theme, accessibility).
• Decide what’s done in chat vs in the editor—and make the handover smooth. 
• Capture success/fail logs and show diffs of AI changes.
• Measure time‑to‑first‑publish and variant uplift as the primary success metrics.
• Keep human review before publish (at least initially). Earn trust, then automate more.

What this unlocks for product teams

• Faster time to first/second/third guide (or announcement, or embed).
• A single flow that blends AI speed with editor precision.
• Proactive suggestions on where to build next based on analytics.
• A shared canvas where PMs, designers, and engineers can each add value—without blocking one another.

About the authors

Jonathan Anderson — Co‑founder & CEO, Candu
Jonathan leads Candu, a no‑code editor for in‑product experiences. He’s focused on AI‑assisted creation that lets product teams ship, refine, and learn faster. Previously at LaunchDarkly, he cares about turning strategy into shipped UI.

Alejandro Dao — Product Manager, Pendo (Guides & AI)
Alejandro builds Pendo’s AI‑assisted guide creation and authoring workflows. His team explores conversational creation with strong guardrails and a clean hand‑off into Designer—so teams move fast without sacrificing enterprise standards.

If you’re experimenting with prompt‑to‑publish in your own stack, we’d love to compare notes.