Skip to content
John HiemstraCreative

Product Engineering

What does it even mean?

AI-assisted development is radically changing software engineering flows and roles, though as a hybrid problem-solver, I've been more or less working this way my whole digital career. I've contributed as facilitator, researcher, designer, architect, developer, tester and manager — whatever my team needed to succeed at the time.

With AI, I'm just more the product owner than ever before, especially in my own house.

House Rules

  • Ship quality fastDeliver aligned, useable work as fast as possible for feedback and alignment, repeat and refine on cadence against established measures.
  • Measurable goalsEstablish early and constantly deliver on what matters to customer most.
  • Where code is enoughDeterministic code does what must be done the same every time. Models run only where judgment pays, so products use less compute & incur less cost.
  • Agents in harnessUse LLMs only with clearly bounded roles, consistent skills/context and clearly defined outcomes set for every token spent.

Learned in enterprise

If that sounds like nothing new, it's because building on-brand custom software on large teams of humans teaches many efficiencies that also align with creating efficient, reliable LLM agents & systems.

Here's where I learned mine.

  • Global accountingCollaborative real-time accounting appGreenfield web app built to serve 500,000+ concurrent users. I was the first front-end hire and led one phase of delivery for a team of more than 20.
  • Regulated healthcare0-1 in four weeksMy first AI-native build: a client-facing testing tool, made with an agent team and a person reviewing every step, shipped in four weeks to unblock a flagship rollout.
  • Healthcare ITMission critical modernizationI led a team of 12 moving live, patient care product from on-prem to AWS, legacy code to latest frameworks and test suites with zero customer disruption.
  • Real-time collaborationPandemic pivotWhen COVID closed core-business IRL training, I helped drive our product's pivot to a virtual classroom experience, helping company maintain value delivery & revenue through lockdowns.

AI Case Studies

  1. Agents in harness

    Remember your lessons.

    Individually defined agents also share agency skills and policies: what 'ready' and 'done' mean, language and coding standards, and more. All work moves forward through gates managed by agents or human review & curation. Every lesson learned flows back into collected agency wisdom for the next work to be done. Simple 'one-way' enterprise UI engineering pattern:

    Shared component libraries: one button under every stepSign-in form, product page, shopping cart and checkout form, one after another. One button component reaches all four, and feedback from any of them goes back into the component: improve once, ship everywhere.Shared UI Component PatternSIGN-IN FORMPRODUCT PAGESHOPPING CARTCHECKOUT FORMBUTTON COMPONENTImprove once, ship everywhere Shared component libraries: one button under every stepSign-in form, product page, shopping cart and checkout form, one after another. One button component reaches all four, and feedback from any of them goes back into the component: improve once, ship everywhere.Shared component librariesSIGN-IN FORMPRODUCT PAGESHOPPING CARTCHECKOUT FORMBUTTONCOMPONENTImprove once, ship everywhere

    One-way architecture like this is a common state-management pattern; now applied to building an agent harness that delivers a complete piece of work:

    Work moves forward, lessons flow back to policies and skillsWork moves forward: plan, build, review, ship, live. Four roles each point to the phases they work in: all four plan, two build, two review, one ships and one runs it live. Policies and skills feed every role, and lessons from live flow back into both.Agent Harness PatternPLANBUILDREVIEWSHIPLIVEROLEROLEROLEROLEPOLICIESSKILLSLessons flow back to policies and skills Work moves forward, lessons flow back to policies and skillsWork moves forward: plan, build, review, ship, live. Four roles each point to the phases they work in: all four plan, two build, two review, one ships and one runs it live. Policies and skills feed every role, and lessons from live flow back into both.Agent Harness PatternPLANBUILDREVIEWSHIPLIVEROLEROLEROLEROLEPOLICIESSKILLSLessons flow back

    How big is this agency?

    25
    roles
    126
    policies
    72
    skills
    13
    gates to ship

  2. Ship quality fast

    Work products use single source of truth.

    Our in-house communications use the same flows we build for clients so turnaround is quick, clear and consistent. Established brand fundamentals drive every new work product, from visual design to brand voice. We build master artifacts in HTML/SVG and CSS so we can rapidly iterate and deliver expressive, brand-aligned docs in any desired format for review, approval and implementation.

    Compose once in HTML, deliver in any formatThe brand kit and your input go to agents that draft in HTML. I review and edit by hand in the same file. The one source becomes a page on your site, a PDF, a Word document, an email or a carousel.BRAND RULESNEW ASKHTML SOURCEHUMAN REVIEWWEB PAGES.PDF & .DOCPRINT & DISPLAYEMAILCAROUSEL Compose once in HTML, deliver in any formatThe brand kit and your input go to agents that draft in HTML. I review and edit by hand in the same file. The one source becomes a page on your site, a PDF, a Word document, an email or a carousel.NEW ASKYOUR INPUTAGENTS DRAFT HTMLI REVIEW AND EDITPAGE ON YOUR SITEPDFWORD DOCUMENTEMAILCAROUSEL
    • The top of a John Hiemstra Creative proposal for Ember & Rye, a demo supper club: masthead, title and scope of workJHC Proposals HTML/PDF
    • Page two of a John Hiemstra Creative sitemap for Ember & Rye, a demo supper club: numbered pillar pages with their child pages and countsJHC Design Documents HTML/PDF
    • The first slide of a John Hiemstra Creative carousel: Tests check the code. Who checks the tests?JHC Social Carousels HTML/PDF
    • Speaker page for Rockstar Blends' founder: hero with a quote, a photo and a speaker kit downloadHeather Reilly - Speaker Kit HTML/PDF
    • Rockstar Blends' wholesale introduction sheet: the four blends listed, then an open wooden display box of bottles beside The Introduction, three of each blend, 12 bottles and a display boxRockstar Blends - Wholesale HTML/PDF
    • Rockstar Blends' tent at Cannadelic, 2026: the branded canopy, tall product banners and a candlelit merch tableRockstar Blends - Live Events PDF/Print

  3. Measurable goals

    $148.00/month to $8.50

    Our first ecommerce solutions for Rockstar Blends wanted a faster customer experience, an easier administrative backend and a lower monthly bill, setting a clear roadmap for ecommerce development in Tailor, our house style website builder. We shipped first live version in September without any business disruption.

    One year of the shop's platform bill, before and nowTwo straight lines over twelve months on one scale from zero to 2,000 dollars. Before, BigCommerce and DropInBlog: 1,776 dollars a year. Now, Tailor on a dedicated VPS: 102 dollars a year. The shaded gap between them, 1,674 dollars a year, is saved.Annual spending change$1,674SAVED$1,776BigCommerce + DropInBlog$102Tailor, dedicated VPS$0$1,000$2,000StartMonth 6Month 12 One year of the shop's platform bill, before and nowTwo straight lines over twelve months on one scale from zero to 2,000 dollars. Before, BigCommerce and DropInBlog: 1,776 dollars a year. Now, Tailor on a dedicated VPS: 102 dollars a year. The shaded gap between them, 1,674 dollars a year, is saved.Annual spending change$1,674SAVED$1,776BigCommerce$102Tailor, VPS$0$1,000$2,000StartMonth 6Month 12
    Performance on a phone, against big media sites, on Google's own ratingsLighthouse mobile, three runs each, November 2, 2026 (BigCommerce: June 2026). CNN: score 43, 10.8 seconds; ESPN: score 37, 7.4 seconds; BBC: score 53, 4.3 seconds; Rockstar Blends: score 70; Rockstar Blends on Tailor: score 96, 2.6 seconds; johnhiemstra.com on Tailor: score 100, 1.4 seconds. Google rates a score of 90 to 100 good, 50 to 89 needs improvement and under 50 poor; and the largest content on screen in 2.5 seconds or less good, up to 4 seconds needs improvement, past 4 poor.Lighthouse performance score on a phonePoorNeeds workGoodCNN43ESPN37BBC53Rockstar Blends (BigCommerce, June)70Rockstar Blends on Tailor96johnhiemstra.com on Tailor1000204050608090100Largest Contentful Paint, secondsGoodNeeds workPoorCNN10.8 sESPN7.4 sBBC4.3 sRockstar Blends on Tailor2.6 sjohnhiemstra.com on Tailor1.4 s0 s2.5 s4 s6 s8 s10 s12 s Performance on a phone, against big media sites, on Google's own ratingsLighthouse mobile, three runs each, November 2, 2026 (BigCommerce: June 2026). CNN: score 43, 10.8 seconds; ESPN: score 37, 7.4 seconds; BBC: score 53, 4.3 seconds; Rockstar Blends: score 70; Rockstar Blends on Tailor: score 96, 2.6 seconds; johnhiemstra.com on Tailor: score 100, 1.4 seconds. Google rates a score of 90 to 100 good, 50 to 89 needs improvement and under 50 poor; and the largest content on screen in 2.5 seconds or less good, up to 4 seconds needs improvement, past 4 poor.Lighthouse performance score on a phonePoorNeeds workGoodCNN43ESPN37BBC53Rockstar Blends (BigCommerce, June)70Rockstar Blends on Tailor96johnhiemstra.com on Tailor10005090100Largest Contentful Paint, secondsGoodNeeds workPoorCNN10.8 sESPN7.4 sBBC4.3 sRockstar Blends on Tailor2.6 sjohnhiemstra.com on Tailor1.4 s0 s2.5 s4 s8 s12 s
    Lighthouse 12.8.2, mobile, three runs each, November 2, 2026; the BigCommerce 70 is PageSpeed Insights, June 12, 2026. Bands are Google's own ratings.

  4. Where code is enough.

    90% less post-production time.

    Established podcast took 4-5 hours of human time to review, prepare search-optimized summary for submission to podcast platforms and craft optimized blog post with highlight quotes. Social content and reels took another 4-5 hours per episode. New "podcast pipeline" application drafts every episode's titles, summary, blog post, quote cards and reels from automated transcripts. Human review & adjustment time down to around 30 minutes per episode from 'stop recording' to 'published on Apple, Spotify, blog and social'.

    One person's hours per podcast episode, before and now, hour by hourOn one scale from zero to ten hours. Before: 4 to 5 hours of episode work, then five reels at about an hour each, of which two or three ship. Now: about 30 minutes for the episode and its reels.BeforeNowEpisode: 4 to 5 h5 reels, 1 h eachAbout 30 min, episode and reels0 h1 h2 h3 h4 h5 h6 h7 h8 h9 h10 h One person's hours per podcast episode, before and now, hour by hourOn one scale from zero to ten hours. Before: 4 to 5 hours of episode work, then five reels at about an hour each, of which two or three ship. Now: about 30 minutes for the episode and its reels.BeforeNowEpisode: 4 to 5 h5 reels, 1 h eachAbout 30 min, episode and reels0 h2 h4 h6 h8 h10 h

    Application flow:

    How an episode gets made and publishedTwo phases. Produce: an audio file goes in, code does the deterministic part, the LLM analyses it and generates into the templates, and a person reviews. Publish: the person adds images and submits, code and the LLM generate the publishing content, the person reviews, and it publishes. Between the phases sits the base: templates and scripts for the code, policies and skills for the LLM. Every review feeds its lessons back into that base.TEMPLATESSCRIPTSPOLICIESSKILLSProducePublishAUDIO FILECODELLMREVIEWADD IMAGESCODELLMREVIEWPUBLISHIterateImprove How an episode gets made and publishedTwo phases. Produce: an audio file goes in, code does the deterministic part, the LLM analyses it and generates into the templates, and a person reviews. Publish: the person adds images and submits, code and the LLM generate the publishing content, the person reviews, and it publishes. Between the phases sits the base: templates and scripts for the code, policies and skills for the LLM. Every review feeds its lessons back into that base.ProduceTEMPLATESSCRIPTSPOLICIESSKILLSPublishAUDIO FILECODELLMREVIEWADD IMAGESCODELLMREVIEWPUBLISHIterateImprove

  5. Measurable outcomes

    8 hours of raw video.

    Rockstar Blends founder co-hosted an all-day online summit live with multiple panels, speakers and topics - leaving her one 8 hour video file to convert into communications and social content. We quickly adapted our podcast capabilities to create a longform video editor that prepares full length videos for publication, and then creates search-optimized short videos and social content. Code transcribes, agents find the speakers, the search terms and complete thoughts for code to structure and present for human review. In phase 2, code generates reels from approved phase 1 structured data, and agents generate summary texts, captions, alt and hash tags. On final creative approval, loader ships selected videos to YouTube with search optimized titles, captions, etc.

    One person's hours for one summit's video, by hand and with the video editorOn one scale from zero to 30 hours, with a line at each 8-hour day. By hand: at least 6 hours of analysis, search optimisation, writing and YouTube publishing, then 19 reels at an hour or more each, at least 25 hours in all, past three days of work. With the video editor, the final run takes 5 hours: the compressed long-form video, the blog post and 19 reels with credits and captions.By handVideo editorAnalysis, writing: 6 h+19 reels, an hour or more each25 h+5 h: long-form video, blog post, 19 reels0 h5 h10 h15 h20 h25 h30 h One person's hours for one summit's video, by hand and with the video editorOn one scale from zero to 30 hours, with a line at each 8-hour day. By hand: at least 6 hours of analysis, search optimisation, writing and YouTube publishing, then 19 reels at an hour or more each, at least 25 hours in all, past three days of work. With the video editor, the final run takes 5 hours: the compressed long-form video, the blog post and 19 reels with credits and captions.By handVideo editor6 h+19 reels, 1 h+ each5 h, final run0 h5 h10 h15 h20 h25 h30 h
    The video editor's final run: the compressed long-form video, the blog post and 19 reels with credits and captions, 5 hours of a person's time.
    Who does each step, transcript to YouTubeSeven steps in three lanes. Code makes the transcript. Agents analyse speakers and search and answer-engine terms, find complete thoughts, and draft highlight quotes and the episode summary. A person reviews and approves; revisions go back to any of the agents' steps. Then code renders the videos with captions and uploads them to YouTube.CodeAgentsPersonTRANSCRIPTSPEAKER &SEO/AEOANALYSISCOMPLETETHOUGHTSHIGHLIGHTQUOTES &EPISODESUMMARYREVIEW & APPROVERENDERWITHCAPTIONSUPLOAD TOYOUTUBERevisions Who does each step, transcript to YouTubeSeven steps in three lanes. Code makes the transcript. Agents analyse speakers and search and answer-engine terms, find complete thoughts, and draft highlight quotes and the episode summary. A person reviews and approves; revisions go back to any of the agents' steps. Then code renders the videos with captions and uploads them to YouTube.CodeAgentsPersonTRANSCRIPTSPEAKER &SEO/AEOANALYSISCOMPLETETHOUGHTSHIGHLIGHTQUOTES &EPISODESUMMARYREVIEW &APPROVERENDERWITHCAPTIONSUPLOAD TOYOUTUBERevisions

Questions

Who made this page?

I did, with the agent team it describes. It is a Tailor page, built from a template in this site's repo and published through Tailor's API.

Who should read this page?

Teams hiring for AI product, forward-deployed or product-engineering work, and businesses that want software built this way.

Where can I see live examples?

This site runs on Tailor. So does Rockstar Blends, our ecommerce product. Its podcast, posts and reels come out of the pipelines on this page.

Do you consult?

Yes. How it works is on the Consulting page.

Can I see the code?

The podcast pipeline is public at github.com/jhcreative/podcast-pipeline. The rest of my public work is at github.com/jhcreative.

Want to chat?

Book a time