10 Claude Code Power Tips You Haven't Seen on Twitter

Most Claude Code tip threads are the same ten things: use --continue, write a CLAUDE.md, don't paste secrets. Fine. Useful the first time you read them. Not useful the tenth.

Here are ten tips I've accumulated over a year of heavy Claude Code use that I haven't seen on Twitter. They're not flashy. They're the kind of thing you'd figure out in month four if you paid attention. Steal them.

1. Rewind mid-session instead of restarting

Most people, when an agent goes sideways, hit Ctrl+C and start over. Waste of context. Use the rewind features: scroll back, identify the message before the agent went wrong, and resume from there with a corrected prompt.

In Claude Code this looks like editing a prior message rather than starting a new session. You keep all the context the agent already has — its understanding of the repo, the files it's read — and just swap out the bad direction.

Saves tokens, saves wall-clock time, and produces better results because the agent doesn't have to re-learn the repo.

2. Use a CLAUDE.md per subdirectory

The top-level CLAUDE.md is table stakes. What people miss is that you can have per-subdirectory CLAUDE.md files that get picked up when the agent is working in that subdirectory.

Useful for:

• src/backend/CLAUDE.md: backend-specific conventions, auth model, error handling.
• src/frontend/CLAUDE.md: component conventions, state management rules.
• tests/CLAUDE.md: testing framework, mocking patterns, coverage targets.

The agent only loads what it needs. Context budget stays manageable. See the Claude Code integration docs for the file-loading rules.

3. Write skills for things you repeat

If you've pasted the same prompt three times in a week, it's a skill. Turn it into a .claude/skills/<name>.md file and call it by name. Skills compose — one skill can reference another — and they're project-scoped.

Examples from my own repos:

• review-pr: my exact review checklist.
• write-tests: conventions for new test files.
• release-notes: how to draft release notes from a commit range.

Skills have a learning curve, but the dividend is real. Full coverage in the Claude Skills workflow post.

4. Pin the model per pane, not per session

Claude Code lets you switch models mid-session, but I've found it's better to dedicate panes to specific models and keep them consistent.

Why: when the model changes, the personality changes, and you spend cognitive effort adjusting. A pane that's always Sonnet feels different from one that's always Opus, and you stop confusing yourself.

In SpaceSpider, the per-pane CLI assignment persists, so this is trivial. See the grid layouts docs.

5. Use the shell pane to run the same diff through multiple checkers

Instead of asking the agent to "run tests, check types, lint," do it yourself in a shell pane with a one-line wrapper:

pnpm test && pnpm typecheck && pnpm lint

Why: the agent running tests costs tokens. You running tests costs nothing. The agent reads the result from the shell pane (if you paste it) or just trusts your summary.

This is anti-agentic in spirit, but it saves money and it's faster. The agent doesn't need to do everything. See cost-optimize AI coding for more of this.

6. Feed errors in, not file paths

When you get a test failure or stack trace, don't tell the agent "the test is failing in foo.test.ts." Paste the full error. The agent will read the error, read the right files itself, and fix it faster than if you hand-held it to the file.

Counter-intuitively, giving more information lets the agent do less work, because you've eliminated the reconnaissance step.

7. Use --print mode for scripted workflows

Claude Code's non-interactive mode is an under-used feature. You can pipe a prompt in and get the answer out, scriptable, composable.

Example — generate release notes from a diff:

git log --oneline main..HEAD | \
  claude --print "Turn these commits into release notes" > notes.md

Example — batch-generate docstrings:

for file in src/**/*.ts; do
  claude --print "Add JSDoc to $file. Output the modified file."
done

Use with care — these run without interaction, so your prompt needs to be precise.

8. Put the agent's output in your PR description

Have the agent draft the PR description before you open the PR. Prompt: "Summarize the changes in this branch for a PR description. Include a 'what changed' section and a 'test plan' section."

The agent has better short-term recall of what it did than you do. The output needs a once-over, but it saves five minutes per PR and produces more consistent descriptions than humans write.

9. Run two Claude panes on the same branch

Not the same worktree — the same branch, in two panes, one Opus and one Sonnet. Opus is doing the work, Sonnet is watching.

When Opus finishes a change, I ask Sonnet in the other pane: "Review what Opus just did on this branch. Anything missed?" Sonnet is cheap, it's reading fresh, and it often catches things Opus glossed over.

This is a poor man's multi-model code review with just one CLI and two models.

10. Keep an "explain it" pane for your own learning

One of my panes is dedicated to "explain what just happened." When the agent does something clever or non-obvious, I copy the change into the explain-it pane and ask "why did you make this choice?"

The answers are sometimes revelatory. The agent has reasons, and articulating them teaches me things. This is where I've learned the most about TypeScript edge cases, database transaction semantics, and distributed systems patterns — not from the code, from the post-hoc explanations.

It's also a check: if the agent's explanation doesn't make sense, the code is probably wrong.

Summary table

A quick scan of the ten:

Setup-level stuff that helps everything

Two meta-tips that amplify the list:

Use a grid terminal. Most of these tips assume you can see multiple panes at once. If you're in a single terminal tab, tips 4, 9, and 10 collapse into "switch tabs a lot," which is friction. Get a grid. See why grid terminals beat tabs.

Define your space once. Once you've got the layout and per-pane CLI assignments you like, define it as a space so you're not rebuilding it every morning. Getting started walks through the setup, and the parallel AI agents use case covers common space templates.

Key takeaways

The power-user moves with Claude Code are not secret features — they're habits. Rewinding, per-subdir context files, skills, pane pinning, shell-run tests, and structured self-review add up to roughly the same thing: do less work for better results.

Pick two or three of the ten to try this week. Most of them become second nature within a few sessions and stack with each other. The tenth — the explain-it pane — is the one that keeps paying dividends long after the others feel automatic.