Igor Karpovich

@GergelyOrosz Same here, just faster

"Earlier, all devs used GitHub Copilot. 9 months ago, we rolled out Cursor to all devs. 1.5 weeks ago, we rolled out Claude Code to everyone, and cancelled our Copilot subscription" - CTO at a company with 600 engineers (I hear this exact "transition" story, a LOT!)

Comprehensive context creates maintenance burden. Staleness creeps in, accuracy drops, no one owns it. Curated context can be kept current, validated, owned. IDP is your friend here.

Teams ask "how do I give Claude context?" Instinct: dump repos, link all Confluence, create massive CLAUDE.md. Wrong. More comprehensive ≠ better. Quality over quantity.

MTTR improved once context existed. Debugging sped up when Claude could query actual architectural decisions instead of guessing from training data. Context infrastructure pays off in metrics.

The onboarding agent pattern: configure Claude NOT to generate code, but to guide learning. Pull context from Confluence + GitHub + Jira, generate bespoke plan. That's the cycle working.

Set up Skyscanner Claude Plugin Marketplace - central repo with codeowners, provisioned via managed Claude config. The place is absolutely buzzing!

Infrastructure work isn't flashy. Building knowledge bases, auto-generating docs with quality gates, creating skills. It's not free. That's what makes coding agents work in enterprise.

Layering LLM on LLM to compensate for missing context degrades output. Text gets duller. The breakthrough wasn't better agents - it was simpler agents with better context.

Built "bee" - declarative YAML layer on AWS Strands. Define agents, tools, prompts, workflows in YAML. Boilerplate dropped ~70%.

Debugging speed improved measurably once context existed. MTTR dropped. Claude could query actual architectural decisions, ownership chains. Context infrastructure shows up in metrics.

A team built an onboarding agent that doesn't generate code - it guides learning. MCPs pull context from Confluence, GitHub, Jira. Generates bespoke onboarding plan. Simple AND practical.

We seconded a junior engineer to a new team. With curated docs in the knowledge graph, they ramped up fast - even in an unfamiliar language. Curation over comprehensiveness works.

The echo chamber risk: AI generates docs - docs feed AI - AI generates more docs. Without human validation, you get AI slop multiplying. Design for humans-in-the-loop from day one.

Curation beats comprehensiveness. We give Claude 50 standards that matter (curated, owned, current) not 5,000 Confluence pages (stale, unowned, noisy). Signal over noise wins every time.

We rolled out Claude Code to hundreds of engineers. Output quality varies wildly. Same model, different results. The differentiator isn't the tool - it's how well we give it Skyscanner's knowledge.

Curated knowledge beats comprehensive. Don't ingest everything. Tie docs to IDP (e.g Backstage). Quality over quantity. Ownership enables accountability

Your competitive advantage in AI: not which model you use, but how effectively you feed any model your org's context. It really is all about building the right infra, not flashy demos

Context engineering has four dimensions (Curate, Persist, Isolate, Compress). Most teams only do Curate. The other three are why your context doesn't scale

Spec-Driven Development + context engineering is the closest to autonomy now. Engineer prompts spec, agent queries KB for standards, generates code/tests/docs, validates, human reviews

In another 6 months, frontier models will shuffle again. What won't commoditise? Your standards, processes, patterns, accumulated decisions