Claude Code Power Tips

(Straight from Boris Cherny's thread – the guy who built it)

Boris (@bcherny) dropped an insane thread about landing 259 fully AI-written PRs in 30 days. The replies are gold for anyone going deep with Claude Code. Here's the latest power moves he's sharing.

1. Session Forking: Branch your conversations at any point #

Just landed: easy conversation forking to experiment without losing your main thread (perfect when compaction threatens a good state or you're juggling multiple ideas).

• Command: claude --continue --fork-session
• Then in the new session: Press Esc twice to rewind to any earlier message and branch from there
• Boris built this on the spot after a feature request – now it's live for splitting tasks or trying alternatives safely

2. Stop Hooks: Unlock true multi-hour/day autonomy #

Claude loves to "stop" after a tool call or when it thinks it's done. Stop hooks fix that by automatically resuming – turning short bursts into uninterrupted marathons.

• Boris's longest run: ~42 hours straight
• Result: Hundreds of PRs with zero manual intervention mid-session
• Best starter: The official ralph-wiggum plugin – it keeps poking Claude until real completion criteria are met
• Repo: https://github.com/anthropics/claude-plugins-official/tree/main/plugins/ralph-wiggum
• Pro move: Define hard exit conditions (all tests green, coverage thresholds, etc.) and let it iterate until they're satisfied

3. Ditch dangerous mode forever – permissions + sandboxes are better #

Boris straight-up never uses --dangerously-skip-permissions.

Instead:

• Fine-tune /permissions → always allow safe stuff (git, ls, grep), ask or deny the risky ones
• Drop into /sandbox when you want isolation
• Outcome: Claude runs way more autonomously without babysitting every command, while staying safe

4. 100% AI-generated code at scale is real (and high quality) #

The bar doesn't drop just because it's AI-written. Same rules: tests, clean architecture, reviews.

How Boris pulls it off:

• Opus 4.5 is excellent at respecting existing code and user intent
• Strong iteration loops via stop hooks
• Kick off most tasks in Plan mode for better structure
• Heavy reliance on verification (tests, coverage, automated reviews)

5. Keep your context window healthy on big projects #

Long sessions eat tokens fast. Stay on top of it:

• Spam /context to see what's bloating the window (subagent logs, MCP tools, etc.)
• Recent updates suppress noisy output and auto-reload docs after compaction
• Track usage locally – cumulative stats are coming soon

Boris goes on to share Claude Code: Best practices for agentic coding alot of great tips to work through and incorporate into your work flow.

• Previous: Sparse Attention