Claude CodeHelsinki

Part 1

Claude News

What shipped, what is worth your attention.

Part 2

Speakers

Four talks on Claude beyond coding.

Part 3

Networking

Drinks, demos, conversations.

Multi-session workspace

• Sidebar with every active and recent session
• Drag-and-drop pane layout
• Each session gets its own Git worktree
• Move between repos as results arrive

Tools in the app

• Integrated terminal and file editor
• Rebuilt diff viewer for large changesets
• HTML and PDF preview
• Side chats with Cmd+; without derailing the main thread

What it does

• Reads the repo, drafts CLAUDE.md
• Suggests skills, agents, hooks
• Cuts onboarding friction for new contributors

What is new

• Applies the smaller-CLAUDE.md principles from meetup #2
• Skips what is discoverable
• Keeps non-obvious conventions and gotchas

Where it is safe

• Read-only operations
• Sandboxed work
• Local-only experimentation

Where you want explicit approval

• Writes to shared infra
• Force-pushes
• Dependency changes

Examples

• Nightly dependency sweeps
• Weekly dead-code reports
• Recurring CHANGELOG drafts

What changed

• Peak window: 5–11am PT / 1–7pm GMT weekdays
• 5-hour sessions burn down faster in that window
• Weekly limits unchanged
• ~7% of users will hit caps they wouldn't have before
• Enterprise plans unaffected (dedicated capacity)

How to plan around it

• Shift long runs to off-peak hours
• Prefer Remote Schedule for token-heavy batch work
• Capacity expanded outside peak, no net loss

The three issues

• Reasoning effort: default lowered Mar 4, reverted Apr 7. Now xhigh on Opus 4.7
• Session-clearing bug: shipped Mar 26, fixed Apr 10. Made Claude seem forgetful
• Verbosity prompt: added Apr 16, reverted Apr 20. Hurt coding quality

What you get

• Secure sandboxing
• Built-in tools
• SSE streaming
• Container configuration via API

Pattern

• Fast executor handles the bulk of the work
• Advisor provides strategic guidance mid-generation
• Long-horizon agent runs benefit most

Result

• Approaches advisor-solo quality
• At executor cost
• Useful for agent harnesses with long sessions

What changed

• Same input now consumes 1.0 to 1.35× more tokens
• The factor depends on what you feed it
• Max output and context limits unchanged

What to do

• Re-tune budget-sensitive prompts
• Watch the bill on first week of migration
• Migration notes published with the launch

Why it matters

• The gap between the model that ships and the model that exists is now public
• Used to surface old, subtle vulnerabilities at scale
• The precedent for partner-only releases starts here

Non-coding angle

Recurring work owned by a Cowork task instead of a calendar reminder. Marketing digests, sales rituals, support triage, weekly KPI rollups.

Supported providers

• Amazon Bedrock
• Google Cloud Vertex AI
• Azure AI Foundry
• Any compatible LLM gateway

Why it matters

• Inference stays on infra you control
• For regulated or data-residency-constrained orgs
• Configured via MDM (Jamf, Intune, GPO)
• End users cannot override the admin profile

Use cases

• Approve a sales summary on the train
• Retry a failed nightly run from bed
• Trigger an ad-hoc digest before a meeting

Why it matters

• Ops, PMs, founders ship dashboards in five minutes
• No BI tool required
• Shareable links, no extra tools

Use cases

• Pull data from three sites, build a comparison
• Drive a CRM through a sales playbook
• Click through a recurring report

For designers

• A collaborator that speaks Figma, not code
• Iterate on screens in the tool you already use
• Reference images become structured input