Checklist

Impact tiers are qualitative starting priorities, not measured savings; impact varies by billing plan and workload. Confirm cost and quality against your baseline.

Before you start a session

• [High] New task → new chat. Side question → /fork.
• [High] Only the tools/MCP servers this task needs are enabled.
• [High] Model chosen deliberately (auto by default); thinking effort left at default.
• [Medium] Noisy paths excluded (.gitignore, files.exclude, search.exclude).
• [High] Prompt names the files/errors/constraints and the success criteria.
• [Medium] Prompt ends with an output limit (Code only. / 3 bullets max.); one task per prompt.
• [Medium] Standing instructions and rules are written in English, terse, and de-duplicated.

During

• [Medium] Use #-mentions for known files; #codebase only when you need search.
• [High] Complex change? Plan first, then implement; delegate bounded pieces to cheaper subagents.
• [High] Ambiguous task? Use Ask/Plan before Agent; give a clear “done” criterion.
• [High] Session getting long and still needed? /compact with a focus hint.
• [Medium] Bulk/polling work → run a script, hand the model the summary.

Weekly review

• [Low] Open the Copilot status dashboard: monthly allowance used?
• [Medium] Find your three most expensive sessions (context-window popover). What made them large?
• [Low] Run /chronicle:cost-tips if available; otherwise inspect Agent Debug Logs Summary and Cache Explorer.
• [High] Prune instructions and tools you never needed (MCP audit).

When in doubt

Measure (page 2), change one thing, measure again, keep it only if quality held.