A photo reaction game on a beamer. The manager starts a round, guests join via QR code. Guests see the photos on the beamer and spend their reaction budget (3 ❤️ + 3 😂 + 3 😮) during a timer. Top 3 photos per category are revealed at the end.

An agentic coding tool that operates in a loop:

Read → Plan → Act → Observe → Repeat

• Terminal (CLI)
• VS Code & IntelliJ
• Desktop app
• Web app
• Smartphone app
• Slack & GitHub Actions

• Conventions are strong
• Compile-time verification
• Testability is built-in
• Structure is predictable

Your stack is your advantage.

Experimentation is free.

Separate context window.

Your session stays lean.

"Research X with a subagent"

❯ A photo reaction game on a beamer. The manager starts a round, guests join via QR code. Guests see the photos on the beamer and spend their reaction budget (3 ❤️ + 3 😂 + 3 😮) during a timer. Top 3 photos per category are revealed at the end.

Claude hardcoded 6 cards per A4 sheet.

Reality: 1 card per A6 page.

The agent only knows what's in context.

Every rule needs to earn its place through a real failure.

# Critical Rules
- Compile after .kt changes:
  ./gradlew compileKotlin compileTestKotlin
- Verify before done: never mark a task complete without
  proving it works (run tests, demonstrate correctness)
- Multi-instance safe: no in-memory state for dedup/guards;
  use DB checks. AdvisoryLock.kt for distributed consensus

.claude/rules/
├── kotlin-conventions.md    → src/main/kotlin/**
├── viewcomponent.md         → src/main/kotlin/page/**
├── tailwind-daisyui.md      → src/**/*.html
├── testing.md               → src/test/**
└── payment.md               → src/**/payment/**

Each file loads only when Claude touches matching paths.

Claude takes notes across sessions — stored locally, not in git

/context shows where you're spending it.

/compact manual compression.

Skills load fresh context at the moment of action.

.claude/skills/commit/SKILL.md

---
name: commit
description: Group uncommitted changes into feature commits.
---
1. Run `git status` and `git diff` to understand all changes
2. Group changed files by logical feature
3. Stage and commit each group with action-oriented message
4. Repeat until all changes are committed

## Reactions
Each guest gets 3 ❤️ + 3 😂 + 3 😮
% confirm this is the correct cookie name
% can we prompt user to enable fullscreen?

Claude suggested 10 changes

Forced to justify: 8 were unnecessary.

You're still the engineer. Understand each decision.

7 phases implemented, 17 tests passing

But the feature was unreachable.

Nobody asked: "How does the user get here?"

jvmskills.com

• Curated AI coding skills for the JVM ecosystem
• Each skill teaches what Claude doesn't know
• Expert-authored, community-driven
• Spring Boot, jOOQ, PostgreSQL, Kotlin, testing

# git-guardrails.sh
DANGEROUS_PATTERNS=(
  "git push" "git reset --hard"
  "git clean -fd" "git branch -D"
  "git checkout \." "git restore \."
)

for pattern in "${DANGEROUS_PATTERNS[@]}"; do
  if echo "$COMMAND" | grep -qE "$pattern"; then
    echo "BLOCKED: '$COMMAND' matches '$pattern'" >&2
    exit 2
  fi
done

TEST_PATTERNS=(
  "gradlew test " "gradlew test$"
  "gradlew.*--tests"
)

for pattern in "${TEST_PATTERNS[@]}"; do
  if echo "$COMMAND" | grep -qE -- "$pattern"; then
    deny "Use the /test skill instead"
  fi
done

⏺ Bash(git stash && ./gradlew test --tests "*LoginTest")
  ⎿  BLOCKED: Use the /test skill instead of
     running gradlew test directly

I need to serialize participants to JSON.

⏺ Added 1 line
    3 + import com.fasterxml.jackson.databind.ObjectMapper

# detekt.yml
ForbiddenImport:
  imports:
    - value: 'com.fasterxml.jackson.databind.ObjectMapper'
      reason: 'Use tools.jackson.databind.json.JsonMapper
        instead (Jackson 3)'

That mistake can never happen again.

Anything you can put in a hook — put it in a hook.

An open standard for connecting AI tools to external data sources.

The IDE becomes an API that the agent calls.

MCP Steroid — the agent is the user, not the developer

• Describe, watch, redirect — real-time collaboration
• Most interactive mode — you stay in control

Scope grows. Plan before you type.

• Claude self-invokes or you trigger with Shift+Tab
• Explore, prototype, iterate — no upfront spec needed

When Claude keeps guessing wrong.
Requirements only exist in your head.

• Drive the design before code exists
• Pair on the spec — with domain experts, UX, product

When the “what” isn’t obvious — only the “why”

Plans work — until the context window doesn’t.

The agent isn’t a deterministic machine

One fresh agent per phase. No drift.

Interview → Spec → Plan → Phase 1 Agent → Phase 2 Agent → ...

"The day shift is thinking,
the night shift is Claude implementing"
— Matt Pocock

- [ ] RED: failing test
- [ ] GREEN: implement

TodoWrite([
  {content: "RED: failing test"},
  {content: "GREEN: implement"}
])

Instructions become tracked state.

Same branch, same prompt — two reviewers.

Different models. Different blind spots.

Each commit contains a feature and its tests

Engineering quality doesn't disappear when AI writes code.
It migrates to specs, tests, constraints, and risk management.
— Thoughtworks, Deer Valley 2026

When implementation is cheap,
the bottleneck moves to judgment.

• What to build
• Whether it's right
• Owning the outcome

Building bad software has never been easier.

curl -fsSL https://claude.ai/install.sh | bash

Observe what goes wrong. Add a rule. Repeat.

Everything from this talk

jvmskills.com/spring-io-2026

OPEN TO WORK

Consulting · Full Time

Spring I/O 2026 · Barcelona