Field Notes - Dec 3, '25

Executive Signals

Outcomes over internals: show time dividends with before-after, keep trajectory visible
Horizon-aware agents: local for sub-10-minute loops, cloud for hour-long chores
Demo before design: brand shell now, framework surgery only after interaction density
First mile brittleness : selector telemetry and scoped retries beat global backoffs
Timeouts expose truth : fail fast trims makespan while preserving success
CMS gravity vs control: editors favor WordPress, performance moats prefer Next.js

Executives remember outcome deltas, not internals. Frame automation as reclaimed cycles and reliability trends, anchored in realistic human baselines that include interruptions. Show a trajectory over checkpoints, not a one-day miracle.

Chart manual cycle time vs an overnight automation window
Sample 10–20 human runs at normal cadence to set baseline
Track rolling success rate and mean/median completion times

Hire For Operational Rigor

Strong operators can ship safely on a Tuesday and explain yesterday’s mistake. Screen for process fluency, consent-aware analytics, and rollback discipline. In interviews, a crisp postmortem with a next-day control is a green flag.

Require local → staging → prod with a rollback and a health check
Gate one GA4/GTM event by consent and prove post-consent firing
Ask for last prod incident, blast radius, and the control added next day

Set the AI Coding Policy That Rewards Verification

Declare AI-literate, verification-first as the bar. Tools are accelerants; reproducibility and tests are the contract. Make provenance explicit in PRs and keep secrets out of prompts.

Ban customer/secret data in prompts; require diffs with tests
Add an “AI used” checkbox and prompt snippet in PRs
Demand local reproducibility and passing tests before merge

Marketing

WordPress vs Next.js For Growth Sites

Default to WordPress when non-technical editors change copy weekly and plugins are leverage. Graduate to Next.js when performance, control, and a smaller attack surface are the moat. Hybrid works: headless WP for content ops, Next.js for guardrails.

Switch thresholds: P95 LCP ≤ 2.0s, TTI ≤ 1.5s, <10 client-side vendors
Use PRs and approvals for pixels regardless of platform
Prefer headless WP + Next.js when both speed and governance matter

Instrumentation As Compliance Drill

Your analytics stack is the client-side data surface. Treat funnels as consent-aware systems with change logs, server-side tagging where possible, and monitoring for third-party drift.

Enforce consent state → tag firing; no consent, no data
Maintain owner, purpose, and data fields per script; review on new vendors
Monitor CSP reports and alert on new or modified third-party scripts

AI-Guided Brand Alignment In A Single File

Use AI as a design assistant to align internal dashboards or microsites without a rebuild. Feed the brand guide and exact HTML/CSS, then request targeted, paste-and-refresh updates.

Upload brand tokens; provide a full-page screenshot and the current markup
Ask for a short “quick wins” list before a single updated HTML file
State hard constraints: preserve IDs/variables, no new build steps

Product

Demo-Driven Scope Control

For internal tools, ship the working flow and add minimal brand so non-technical reviewers perceive progress. Use the demo as a forcing function: secure a directional yes before architecture investments.

Lock scope to the demo-critical path; defer charts and net-new pages
Ask only: “Directionally correct?” and “One change before Friday?”
Convert design debt into post-demo tickets with owners and dates

Information Architecture Beats Pie Charts

Reduce time-to-judgment by fixing how information lands before adding graphs. Lead with triage views and artifact-first layouts, then logs.

On failure, surface the error summary at the top; on success, show the artifact first
Open “View Output” in a modal to avoid page hops
Use table states and badges for Queued, In-Progress, Done, Failed with brand colors

Make Internal Tools Pass The Impress Sniff Test

Auditors, buyers, and execs do see internal dashboards. A brand-consistent shell reduces questions about the engine and feels dependable.

Add a clear header, breathing room, table hierarchy, and status badges
Format dates, normalize copy, and standardize on one icon set
Include a crisp PR note summarizing visual deltas for reviewers

Engineering

Manage Agents Like Teammates, Time-Box By Horizon

Treat code agents like coworkers with scoped run windows and clear artifacts. Keep UI loops local for speed; kick off longer backend chores in the cloud while you’re in meetings. Ground changes with both code and screenshots.

Speak prompts; specify constraints and desired artifacts
Provide repo context and a screenshot; ask for plan → code → PR draft
Run sub-10-minute tasks locally; 30–90 minute jobs in cloud before meetings

Two-Day Dashboard Rule: Tailwind Now, React Later

When you have 48 hours to impress, avoid framework surgery. Add a utility CSS layer for brand polish and defer React until interaction density warrants it.

Ship polish with Tailwind or a lightweight CSS pass; compile to plain CSS if needed
Graduate to React at >3 interactive components or shared state
Optimize for a clean, on-brand read over a “perfect stack”

Kill Flaky Logins With Selector Telemetry And Targeted Retries

A brittle first mile needs surgical hardening, not global backoffs. Instrument selector choice, retry narrowly, and learn which selectors win.

Add 1–2 login-specific retries with short backoff; cap total attempt time
Promote highest-success selectors automatically; log structured failure reasons
Review login failures daily until sub-0.5% is sustained

Shorter Timeouts Shrink Batch Wall-Clock

Long timeouts hide tail failures and bloat makespan. Tune to reality, fail fast, and retry early while bounding cumulative time per item.

Set timeouts to roughly P95 duration plus a small buffer; re-evaluate weekly
Fail fast and retry early; track per-item cumulative time
Monitor makespan and tail jobs; drop timeouts again if success holds

Queue Purges Shouldn’t Kill Workers

Control-plane actions must be survivable. Treat purges as fault injection and architecture as recovery-first.

Replace “purge” with “drain and cancel” using idempotent tokens
Supervise workers with liveness/readiness probes and auto-restarts on exit codes
Add a chaos check in CI/CD: simulate queue clears during load and verify recovery