Skip to main content

Tips & tricks

Things that are not in the mental model but change the day.

Context​

  • /compact mid-task, /clear between tasks. A fresh context beats a compacted one carrying a dead task's residue.
  • Sessions living above 150k tokens cost more on every turn, cached or not.
  • Write big output to a file, read the slice. Never dump a full test log or a 200-page extraction into the conversation.
  • .claudeignore keeps lockfiles and build output out from the start. Cheapest win available.

Git worktrees​

Run several sessions on the same repo without them fighting over the working tree:

git worktree add ../repo-feature-x feature-x

One agent per worktree, each with its own branch and its own checkout. Docs.

Permissions​

Pre-allow the read-only commands you approve twenty times a day:

"permissions": {
  "allow": ["Bash(git status:*)", "Bash(git diff:*)", "Bash(ls:*)", "Bash(rg:*)"],
  "deny":  ["Read(./.env)", "Read(./**/*.pem)"]
}

deny is the important half. Fill it before allow.

Plan mode​

Plan mode for anything 3+ steps or architectural. If something goes sideways, stop and re-plan rather than pushing. After two consecutive tool failures, change approach entirely.

Reviewing a diff​

git diff | claude works, but /review or the reviewer subagent is better: they read the surrounding code, which is what turns a confident-wrong finding into a real one.

When it goes wrong​

Symptoms and what they usually mean:

SymptomUsually
Edits the same file 3+ timesIt never read the whole file. Tell it to read, plan, then make one edit
Keeps retrying the same failing approachTwo failures is the signal to change strategy — say so
Ignores a correctionMake it quote back what you asked before proceeding
Wanders the codebaseGive it a starting file, or dispatch explorer

Those four are worth putting in CLAUDE.md — they are in this kit's already.

Official​