I built the workflows.They built the product.

Start with a business problem worth solving

A founder can know their product inside out and still struggle to decide how to promote it. Positioning, audiences, channels and campaign content depend on one another. A collection of disconnected AI tools leaves the founder doing the integration.

I directed market and competitor research to examine those gaps, alternative tools, pricing and the risks of delegating marketing to AI. That work shaped the product strategy: carry the business context from analysis into campaign preparation, with a founder able to review the result before delivery. Research became a basis for deciding what to build and where human control mattered.

Two systems: the product and the way it gets built

I am the sole developer and product lead behind this build. I own the idea, strategy, UX, workflow architecture and acceptance decisions. AI delivers the implementation under my direction. As a UX Director and AI DesOps Lead, my contribution is turning expertise into an operating system for delivery.

I constructed two connected systems. Inside Launcherry, AI workflows turn a product URL or brief into marketing analysis and campaign drafts. Around the build, my development harness gives coding agents the context, skills, tasks and checks needed to produce reviewable work. The product is the tangible result of that approach.

Construct workflows that carry context forward

I designed the generation work as connected stages with defined inputs, outputs and checks. Business facts and marketing analysis inform campaign preparation; channel requirements shape the deliverable; review determines what can move forward. Each stage has a purpose beyond producing another answer.

This is the core of my AI workflow orchestration: decide what context an agent needs, what it must return, how that result will be evaluated and what happens when it fails. The current development implementation includes staged campaign generation, fact grounding and bounded repair. I retain responsibility for the decisions the system makes possible.

Build reusable skills from researched expertise

Generic marketing instructions were not enough. My channel review found that platform guidance was missing from parts of the generation workflow and that some drafts repeated internal goal labels instead of giving readers a useful next step. I used those findings to direct a reusable, model-agnostic skill library.

I researched the requirements of individual platforms and delivery types, then translated them into 16 channel-specific skills and one shared core for planning, writing and asset direction. Shared standards stay consistent; an organic post and an ad receive different guidance. The skills are connected to the generation code and checked against an authoring standard, with separate evaluation criteria to assess the outputs. This library is implemented in local development; availability in the public product depends on the release.

A development harness built for evidence

I built my working harness around AI coding tools: persistent project knowledge, reusable instructions, scoped implementation tasks, acceptance criteria, independent critique and verification gates. Agents receive a defined piece of work and a way to establish whether it is complete. I review the evidence and decide what is ready to integrate.

The supporting tooling includes a repeatable demo environment, browser checks, regression tests and evaluation runners that exercise the actual generation pipeline. I also built a bridge for iterating on real-model output before a production-provider check. I measure output quality against the product requirements and assess production cost, latency and caching separately.

Build → challenge → repair → verify

I organised development as a feedback loop. AI implements a bounded change; critics challenge the result; tests and browser checks expose failures; targeted fixes return through the same gates. Findings feed back into the code, skills and project knowledge so the next iteration starts with a stronger foundation. Improvement is directed and verified, with me owning acceptance.

For example, I rejected cutting generated ad copy to fit platform limits: a technically valid fragment can still be unusable marketing. I directed changes to generation and validation so overlong copy is rewritten within the requirements, then checked the behaviour through regression coverage and real-model evaluation. The loop improves the deliverable, not just the number of passing checks.

Design operations and business economics are part of the system

I connected research, skill maintenance, UX, implementation and quality review into one delivery practice. That is design operations applied to AI: turn judgment into reusable standards, give work a clear route through review and preserve the knowledge needed for the next change.

I also treated AI usage as an operating cost. When production evidence showed that repeated generation work was not benefiting from caching, I directed a correction and checked actual cache reuse. In two measured generation-repair calls, 3,879 of 3,932 input tokens came from cache: 99% reuse, rounded. That connects workflow architecture to unit economics. I optimise the process through observable behaviour rather than assuming a newer model or a longer prompt will make it better.

A working product in under two months

53 calendar days separate the first recorded implementation on 4 July and the verified working deployment checkpoint on 26 August 2026. At that checkpoint, the web and API were deployed, application and data-service health checks passed, and the recorded unit/configuration suite reported 4,309 passes. That is the documented delivery milestone behind this case.

Launcherry connects product input, marketing analysis, campaign review and export in an implemented web/API journey. Approval makes a campaign ready for export; download and supported platform delivery are separate choices. Launcherry cannot automatically activate paid media. I made that boundary part of the product experience and its implementation.

Development continued beyond that milestone. The 3 October integration checkpoint records 10,010 passing unit tests across the core, web and API suites, alongside build and browser checks. The API suite also records 71 skips and one todo. These are dated verification results for the local build; their value is in the behaviours they check, including approval, billing, generation and export.

The result demonstrates my ability to construct AI workflows, build the harness around them and carry the work through product and operational decisions. The public product is available to explore, while newer skills and workflow improvements continue in local development. This is the capability I bring to AI development and business process optimisation: connect domain expertise, reliable execution and a usable product.

See the product at launcherry.me ↗︎

Explore the AI delivery system behind Launcherry

Seven articles on the decisions, methods and evidence behind the build. Start with the development harness, or follow the part of the work you want to examine.

Building an AI Development Harness: From Brief to Verified Product
How I frame work, provide current context and connect AI implementation to critique and behavioural checks.

Building Reusable AI Agent Skills from Domain Expertise
How platform research becomes guidance that generation actually consumes, with separate output assessment.

AI Workflow Orchestration: From Analysis to Campaign
How business context moves from analysis into campaign preparation, with checks at the handoffs.

Evaluating AI Outputs: When Passing Validation Is Not Enough
Why I rejected truncated ad copy, and what structural checks and real-model evaluation each establish.

Human Approval in AI Products: Designing the Boundary Before Delivery
How review, approval and export form separate decisions before supported delivery.

Prompt Caching in Practice: Diagnosing Writes Without Reuse
The investigation behind 99% prompt-token reuse in two measured repair calls.

Building a SaaS with AI: From Product Idea to Working Deployment
What I owned, what AI implemented and what the 53-day deployment checkpoint proves.