Skip to content

Power User Techniques

Qin Yu, 21 May 2026, updated 29 May 2026

Power User Techniques

Who this is for

Most techniques here apply to anyone using an AI agent interactively. A few — Collapse tool calls, Put stable prompt prefixes first, and Use structured tool outputs — are primarily relevant if you are building or scripting agent workflows via API. If you are a software engineer, ML engineer, or scientist using Copilot or Claude Code, you can skip those three and still get most of the value from this page.

Each of these carries trade-offs and rewards deliberate practice.

  • Think in code, not prose — when analysing large files or datasets, write a script rather than feeding raw content to the model. Scripts are deterministic, reusable, and produce compact output.
  • Use CLI tools when they are leaner — CLI outputs are often more compact and predictable than MCP wrappers or screenshots.
  • Compress shell output — shell output can be extremely verbose. Tools such as rtk-ai/rtk can post-process and compress output before it enters the context window. Some harnesses already compress or omit shell output.
  • Analyse usage regularly — use tools such as /context, /usage, /session, and /chronicle where available.
  • Collapse tool calls — multiple sequential tool calls inflate context. Plugins such as jsturtevant/copilot-codeact-plugin can batch tool calls to reduce per-turn overhead.
  • Put stable prompt prefixes first — for API workflows with prompt caching, invariant instructions should come before variable task data.
  • Use structured tool outputs — compact JSON, tables, and summaries are cheaper than raw logs.
  • Benchmark on representative tasks — different models respond differently to prompt structure, context ordering, and instruction style.

References: