principle-guard-the-context-window
Guard the context window. Apply when the context fills up: large outputs, long files, repeated reads or planning a fan-out.
- License 许可
- MIT
- 最近更新
- 2026-11-15
- Stars
- ★ 2
- Forks
- 0
01 前言 Front Matter
- name
- principle-guard-the-context-window
- license
- MIT
- metadata
- { "source": "pstack principle-guard-the-context-window, MIT, Copyright (c) 2026 Lauren Tan, github.com/cursor/plugins, adapted; checked against pstack 0.15.36 at commit 0bb6f37" }
- description
- Guard the context window. Apply when the context fills up: large outputs, long files, repeated reads or planning a fan-out.
03 SKILL.md 正文
<!-- Generated by scripts/build.mjs from the sources; edit the source, then run npm run build. -->
Guard the context window
The context window is finite, and you cannot renew it within a session. Each token must be worth its cost.
Why. When the context overflows, reasoning gets worse, compression loses detail, and progress stops.
What to do.
- Isolate large payloads. Send long outputs, screenshots and large documents to subagents. The main context gets summaries, not the raw data.
- Keep content that you use often inline. A template or reference that each use needs belongs in the skill file, not in a separate file that costs a read each time.
- Size the phases and limit the scope. Limit the files in each phase, set turn budgets, and count the cost of each mechanism.
04 同仓库更多 Skills
writing-standard
Write at 80% of Simplified Technical English. Apply to any text an agent writes for a person: messages, reports, specs, findings, reviews, plans.
sage
The state tool of sage mode: records tasks, routes, agent runs, findings, verdicts by head SHA and the user's gates in ~/.claude/sage/<project>/. Use in sage mode for every change of a task's state, and before any merge. Claude Code only.
sage-help
Help with sage: how to set it up, switch sage mode and autopilot, read the board, answer the chief's questions, pick a skill or a principle, fix a run that went wrong, and find where sage keeps its files. The user types /sage:sage-help (or /sage-help) with a question. Load it only when the user types it; do not load it on your own. Claude Code only.
report
The report that every sage agent ends with: STATUS, RESULT, EVIDENCE, FINDINGS, QUESTIONS, NOT VERIFIED, BRANCH. Use at the end of any task that the sage chief of staff gave you. Claude Code only.
remote-control
Set up Remote Control, so that the user can follow, steer and start Claude Code sessions from the Claude app on a phone. It turns Remote Control on for every new session, and it can run a server in a folder (such as ~/workspace) that starts at each login. Use when the user asks to reach sessions or start new ones from a phone, or to set up, check or undo Remote Control. Claude Code only.
readme-guide
Write a project README. Apply when you write, change or review a README of any project, in any language.