Applying CLAUDE.md, Skills, MCP, etc. indiscriminately burns through tokens very quickly. (Keep checking with /context and /usage.)
In the previous era the "code" itself was the asset, but in the AI era the "test cases" are the asset. -> SQLite:Testing#How SQLite is tested
Register skills or agents that need to be used in common on a marketplace.
Break work down with subagents. In this case, context is not shared in Agent 2 Agent. -> Since the context is not connected, tokens are saved, but the intermediate process is cut off, so answer quality drops somewhat.
Use several subagents regularly
code-simplifier - simplifies code after Claude finishes its work
verify-app - includes detailed instructions for end-to-end testing of Claude Code
Agent Teams (CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS is in beta as of March 5, 2026) In this case, context is shared in A2A.
Operating-related content can also be registered as a subagent.
Use Claude Hooks for static checks right after Write and Edit.
Use a Notification to announce task completion.
If you need straight-through work without questions, use a Background task (&) -> this requires a certain amount of trust to have been built.
When using multiple sessions, either split them with tmux or use a related GUI Application.
Make active use of file references (@), Commands (/), etc.
Make active use of Frontmetter to avoid quality degradation caused by the LLM's trailing-token recognition problem.
Provide content that doesn't really need to be referenced as links.
Don't make changes mid-run; fix the plan and run again.
If you don't like the result, don't keep repeating code change requests.
Revise the plan document, then run again in a new session.
Changing direction during the execution phase makes quality collapse quickly.
Teach style based on examples.
For security, operations, and long-term maintenance code, you must understand the whole code. LLMs have the characteristic problem that comprehension drops as the context size grows.
Don't ask the AI to change scripts and code you can simply run yourself. -> The indiscriminate stdout/stderr after asking is the culprit that burns tokens.
Use tests, linters, and builds as a feedback loop
Commit when each task is done, so you can roll back per single Q&A.
Split work into extremely small units
For repeatable work, use AI to build scripts/tools that do that work. -> What you ask the AI for as output is "the actions needed to reach the goal" rather than "the goal".
Measure the quality of AI results by "consistency of the output".
Keep asking the AI for "structural completeness" in your requests.
If you're short on tokens, pick the model to suit the situation. -> Opus for coding, sonnet for writing commit messages, Haiku for simple translation