ai in
practice . . .

AI should be involved in every artifact you create — not as a shortcut, but as a collaborator that raises the floor of what's possible.

Vol. I — 2026

operations replaced

No. 01

Design Engineering

Meeting notes → prototype changes in a single session

Meeting notes → PRD → design brief → handoff was a multi-day cycle before a single Figma frame moved. Built a pipeline that ingests raw transcripts and outputs component-level prompts that map directly to prototype changes. A customer offhandedly mentioned bulk editing — the feature was prototyped and in review by end of day.

Claude API Notion Figma Make
1
Raw meeting transcript ingested from Notion
↓
2
Claude extracts design intent, maps feedback to components
↓
3
Structured prompt output → Figma prototype updated same session

Engineering / QA

Rage click → bug fix → PR, without a human in the loop

Pendo rage clicks and user-flagged bugs sat in a queue until someone manually triaged, wrote a ticket, and tracked down the right engineer. Built an agent that monitors Pendo for anomalies, classifies them as bugs, posts a diagnostic summary to Slack, generates a targeted fix, and opens a PR assigned to the feature's code owner — all before the next standup.

Claude API Pendo Slack GitHub API
09:14:02 pendo rage_click detected — /dashboard/filters, 14 events in 8s
09:14:03 agent classifying… bug confirmed (filter state not persisting on nav)
09:14:05 slack #eng-alerts — diagnostic posted, root cause attached
09:14:09 agent fix generated — useEffect dependency array corrected
09:14:11 github PR #482 opened → assigned to @mwilson (filters owner)

Program Management

One agent, three update formats, every stakeholder covered

Status updates were written three times a week for three different audiences — engineering standups, leadership newsletters, investor decks — each pulling from the same sources but rewritten from scratch. Built an agent that queries Jira, Slack, and meeting notes once, then generates all three: a technical wins/blockers brief posted to Notion, a monthly narrative for the leadership newsletter, and slide copy pushed directly into the quarterly investor deck.

Claude API Jira Slack Notion Google Slides API
1
Agent queries Jira, Slack, and meeting notes to aggregate weekly signal
↓
2
Claude generates three artifacts: technical brief, leadership newsletter section, investor slide copy — each tuned to its audience
↓
3
Outputs posted to Notion DB, newsletter draft, and slides updated in place

Product Strategy

Lucy — the customer advocate agent that tells you what not to build

AI makes it easy to ship more. The harder problem is knowing when not to. Built Lucy, an agent you can summon in Slack or point at a project brief — she pulls customer activity data and runs analysis on the idea before a line of code is written. Real case: we were considering a v2 grouping feature. Adoption of v1 was low. Lucy flagged it. We deprioritized. The roadmap got sharper.

Claude API Slack Pendo Notion

Without Lucy

Ship intuition-led features, discover adoption gaps post-launch
→

Lucy

Reads brief or Slack thread → queries customer activity → surfaces adoption signal and recommendation
→

With Lucy

v2 grouping deprioritized before sprint — roadmap time reclaimed

Vol. I — 2026

takes

No. 02
take_01.txt

01

The meaningful product moment isn't when users gain value — it's when they stop questioning outputs and begin trusting the system with responsibility.
take_02.txt

02

In AI products, trust rivals traditional metrics. Users evaluate reliability and reversibility. If they don't trust the system, they won't delegate to it.
take_03.txt

03

AI's real value emerges when it enables action, not just awareness. The gap between insight and doing is where most tools fail.
take_04.txt

04

Capability doesn't justify implementation. Trust is built through recovery, boundaries, and predictability — not peak performance.
take_05.txt

05

AI adoption requires designing for emotional states — curiosity, skepticism, anxiety. Trust is emotional. We need to build for that.
take_06.txt

06

Minor tweaks in copy, flow, or onboarding reveal hidden user assumptions. A three-word change in a notification doubled click-through.