28. Best practices

28.1 Always plan first, then execute

The read-only constraint of plan stops the AI from rushing into code with the wrong premise. Even for small tasks, enter plan → let the assistant list a plan in 5 min → exit plan and enter execute — overall efficiency is better.

28.2 Let the team own personas

Keep the assistant's system prompt in the team repo. New members joining pick up the team-consensus AI style automatically.

28.3 Set "Ask" on high-risk tools

For tools touching production, shared secrets, file deletion or pushing code, recommend setting them to "Ask" mode and using "same tool + same args" granularity so every call is human-confirmed.

28.4 Use history compaction

Don't open a new session just because the conversation is long — original context can still be expanded; "new session from zero" loses implicit knowledge.

28.5 Parallel is not always faster

Background sub-tasks are good for parallel reads (reading several files / searches concurrently), bad for parallel writes (conflicts on the same file). Parallelise reads; serialise writes.

28.6 Use companion messaging for long tasks

Long tests / scans can go to background; have the assistant push progress to your phone — leave your desk and do something else.

28.7 Use "hand of fate" for a sanity check

When a particular persona has you going in circles, swap in a totally different persona temporarily for an adversarial review — finds blind spots in the main persona.