Brendan Graetz • @bguiz • Hard parts: Building with Gen AI

Hard parts: Building with Gen AI

Brendan Graetz
2026/06

Brendan Graetz • @bguiz • Hard parts: Building with Gen AI

$ whoami

  • Web2 software engineer
  • ETH Singapore 2018
  • DApps Dev Club
  • Rootstock
  • Hedera
  • Injective
Brendan Graetz • @bguiz • Hard parts: Building with Gen AI

Topics

  • What is actually hard now
  • Product planning
  • Implementation planning
  • Building
  • Reviewing
  • Demo
Brendan Graetz • @bguiz • Hard parts: Building with Gen AI

What Is Actually Hard

Hard parts: Building with Gen AI

Brendan Graetz • @bguiz • Hard parts: Building with Gen AI

What Used To Be Hard

StageMain burden
PlanningLight
CodingHeavy
ReviewingLight

Ask yourself

  • What takes the most time?
  • What requires the most expertise?
Brendan Graetz • @bguiz • Hard parts: Building with Gen AI

What Is Hard Now

StageMain burden
PlanningHeavy
CodingCheap
ReviewingHeavy

Reframe

  • Coding got dramatically cheaper
  • Planning and review became the bottlenecks
  • This talk focuses on the planning side
Brendan Graetz • @bguiz • Hard parts: Building with Gen AI

Recap: What Is Actually Hard

  • Software work still has planning, coding, and review
  • The effort curve has inverted
  • Coding is no longer the scarcest part
  • Planning and reviewing now need more attention
Brendan Graetz • @bguiz • Hard parts: Building with Gen AI

Product Planning

Hard parts: Building with Gen AI

Brendan Graetz • @bguiz • Hard parts: Building with Gen AI

Planning = ?

  • What the product itself is or does
  • How to implement the product
Brendan Graetz • @bguiz • Hard parts: Building with Gen AI

Planning Inputs and Outputs

StageInputOutput
(1) Product definitionIdeaDetailed product description
(2) Implementation planningDetailed product descriptionImplementation specifications
Brendan Graetz • @bguiz • Hard parts: Building with Gen AI

Input: The Idea

  • Problem + (probable) solution
  • Extremely high level
  • Usually just 5-10 points
Brendan Graetz • @bguiz • Hard parts: Building with Gen AI

Output: Product Description

  • Terms and definitions
  • Requirements x shapes
  • PRD
  • Breadboard
This is where the rough idea becomes something concrete enough to build.
Brendan Graetz • @bguiz • Hard parts: Building with Gen AI

From Idea To Product Description

  • Start with existing agent skills
  • Modify their behavior with custom prompts
  • Chain them into a repeatable flow
  • Manage context windows, model choice, and effort level

Note: I'm still learning too!

Brendan Graetz • @bguiz • Hard parts: Building with Gen AI

Product Planning Skill Stack

  • /grill-with-docs
  • /to-prd
  • /shaping
  • /breadboard
  • /grill-with-docs again
All combined with custom prompts.
Brendan Graetz • @bguiz • Hard parts: Building with Gen AI

Recap: Product Planning

  • Planning has clear inputs and outputs
  • Stage one turns an idea into a detailed product description
  • Agent skills can be composed into that workflow
  • Prompting and orchestration matter as much as the skills themselves
Brendan Graetz • @bguiz • Hard parts: Building with Gen AI

Implementation Planning

Hard parts: Building with Gen AI

Brendan Graetz • @bguiz • Hard parts: Building with Gen AI

Stage Two

StageInputOutput
(1) Product definitionIdeaDetailed product description
(2) Implementation planningDetailed product descriptionImplementation specifications
Brendan Graetz • @bguiz • Hard parts: Building with Gen AI

Input: Product Description

  • Terms and definitions
  • Requirements x shapes
  • PRD
  • Architecture
  • Breadboard
Brendan Graetz • @bguiz • Hard parts: Building with Gen AI

Output: Implementation Specs

  • Slices
  • Architecture
  • Test plan
  • Issue tickets
  • Epic tickets
Brendan Graetz • @bguiz • Hard parts: Building with Gen AI

From Product Description To Specs

  • Reuse proven agent skills
  • Adjust behavior with custom prompts
  • Chain them into a mechanical workflow
  • Manage context, model, and effort intentionally
Brendan Graetz • @bguiz • Hard parts: Building with Gen AI

Implementation Planning Skill Stack

  • /shaping
  • /grill-with-docs
  • /grill-with-docs again
  • /to-issues
Again: these work best when wrapped in custom prompts and orchestration.
Brendan Graetz • @bguiz • Hard parts: Building with Gen AI

Recap: Implementation Planning

  • Stage two converts product description into build-ready specifications
  • The workflow is more mechanical than stage one
  • It is often the planning stage engineers already recognize
  • Skill composition still matters
Brendan Graetz • @bguiz • Hard parts: Building with Gen AI

Building + Reviewing

Hard parts: Building with Gen AI

Brendan Graetz • @bguiz • Hard parts: Building with Gen AI

Not Today

  • Building
    • Skills
      • /tdd (@mattpocock)
      • /test-driven-development (@obra)
      • /build (@addyosmani)
    • Orchestrators
      • /goal
      • ralph, openGSD, oh-my-codex, etc
  • Reviewing
    • (still researching)
Watch this space.
Brendan Graetz • @bguiz • Hard parts: Building with Gen AI

Recap: Building + Reviewing

  • We intentionally skipped the deep dive
  • The important point is that planning comes first
  • Build and review workflows still need iteration
Brendan Graetz • @bguiz • Hard parts: Building with Gen AI

Reflections

Hard parts: Building with Gen AI

Brendan Graetz • @bguiz • Hard parts: Building with Gen AI

My journey + screw ups

  • Find "off the shelf" agent skills
  • Run them
  • Get insta-specs
  • Orchestrator go BRRRR
  • Middling result
Brendan Graetz • @bguiz • Hard parts: Building with Gen AI

My goal + revised goal

  • Initial goal
    • Use agent skills to one-shot great specs
    • Orchestrator grinds to one-shot build
  • Realisation
    • WTF did I flip from agile back to waterfall?
    • Don't build the thing; build the script that builds the thing
  • Current goal
    • No one-shots allowed: "Human in the loop"
    • Properly grill: "I am clear about what I'm building"
    • Orchestrator grinds to build vertical slices
Brendan Graetz • @bguiz • Hard parts: Building with Gen AI

Intended takeaways

  • One-shot builds are good for:
    • Simple greenfield projects
    • Single features in existing complex projects
    • Everything else: Nope!
  • Don't take the human out of the loop
    • Thoughtfully select insertion points
  • Most important:
    • Mindset: Self-clarity
    • The hard parts: Before and after coding
Brendan Graetz • @bguiz • Hard parts: Building with Gen AI

Skills which guide you

Brendan Graetz • @bguiz • Hard parts: Building with Gen AI

Demo

Hard parts: Building with Gen AI

Brendan Graetz • @bguiz • Hard parts: Building with Gen AI

Demo Recap

  • Live walkthrough completed
  • The workflow is the point, not just the output
  • Planning quality determines build quality
Brendan Graetz • @bguiz • Hard parts: Building with Gen AI

Q&A

  • Ask about planning workflows
  • Ask about skill composition
  • Ask about building with gen AI in practice
Brendan Graetz • @bguiz • Hard parts: Building with Gen AI

After The Session

  • Drop questions in the recording comments
  • Reach out directly if you want to compare workflows
  • I'm especially interested in better review approaches
Brendan Graetz • @bguiz • Hard parts: Building with Gen AI

Hard parts: Building with Gen AI

Thanks

2026/06

Brendan Graetz • @bguiz • Hard parts: Building with Gen AI