01
What I explored, built & changed
Agent Customization, Skills & Context Optimization
prompt-pack
-
Explored agent patterns from the
Agentic AI Foundation
and built an
orchestrator workflow that pairs a
test-oracle writing acceptance tests upfront
with a blind-implementer coding against them,
sight unseen.
-
Created reusable agent instructions and packaged helpful
day-to-day routines with
Agent Skills
(
SKILL.md) to cut down repetitive context, keep
agents focused, and save tokens.
-
Hooked up agents with Model Context Protocol (MCP) servers
(GitHub, Figma, Miro, Chrome DevTools) to give them
practical superpowers in UI design, code reviews, and
browser testing.
Multimodal AI & Schema Integration
exan
-
Built an image-scanning and OCR analysis app powered by
multimodal LLMs.
-
Crafted prompt structures and clean data contracts that
bridge provider outputs with deterministic application
schemas for reliable text extraction and similarity scoring.
Local LLM Benchmarking for Agentic Development
ai4se-local-llm-guide
-
Experimented with open-weight multimodal models on local
machines to find the sweet spot: figuring out what tasks
(like quick ideation, drafting, and boilerplate) work great
locally versus when it makes sense to bring in frontier
cloud models.
02
What I learned — challenges & limitations
Key takeaways
-
Giving agents well-structured context (using
SKILL.md and tailored
.agent.md setups) makes their reasoning much
more consistent while keeping prompts lean.
-
Separating test generation from code implementation keeps
everything honest — preventing hallucinations and keeping
coding goals clear.
Challenges
-
Guiding autonomous code generation smoothly without getting
stuck in feedback loops or straying from the original
intent.
-
Bridging the gap between creative multimodal outputs and
strict, predictable software schemas.
Limitations
-
Local models can still struggle with large context windows,
reliable tool calling, and complex multi-file reasoning
compared to cloud frontier models.
03
Why I want to join the workshop
-
I've really enjoyed following Daniel Meppiel's
Agentic SDLC Handbook, and I'd love to take these
hands-on experiments to the next level by turning them into
dependable, production-ready workflows for our team.
-
I would be gladto meet up with Daniel and other practitioners
in Düsseldorf to swap ideas, discuss real-world multi-agent
coordination, and explore practical quality gates and governance
together.
04
How I would contribute & help colleagues afterward
-
Host lunch-and-learns and interactive sessions to help teammates
level up from simple autocomplete to structured, test-driven
agent workflows.
-
Put together clear, practical guidelines on when to run models
locally versus leveraging cloud agents for the best balance of
speed, cost, and privacy.