Unofficial guide. Not affiliated with OpenAI, Anthropic, GitHub, or Claude.
Agent Token Budget is a polished, copy-friendly playbook for keeping AI coding agents fast, focused, and safer. It turns the messy problem of "why did this agent burn so much context?" into a repeatable workflow: scope first, load less, keep security boundaries visible, and archive heavy sessions only after handoffs exist.
Built by 2023Anita.
- A beautiful GitHub Pages site that explains token leaks as a visual system.
- Ready-to-copy
AGENTS.mdrules for Codex-style workflows. - Prompt templates for safe read-only analysis, browser reading, repo scanning, config audits, and maintenance reports.
- Original diagrams for token leakage, context budgeting, safe browser boundaries, and project replication.
- Brand and trademark guardrails so the project can reference AI ecosystems without pretending to be official.
git clone https://github.com/2023Anita/agent-token-budget.git
cd agent-token-budget
open site/index.htmlCopy the templates you need:
cp templates/AGENTS.md ~/.codex/AGENTS.md
mkdir -p ~/.codex/prompts
cp templates/prompts/*.md ~/.codex/prompts/Do not copy auth.json, API keys, cookies, access tokens, or private project files into this repository or any third-party webpage.
| Leak | Symptom | Fix |
|---|---|---|
| Unbounded repo scans | The agent reads everything before deciding what matters. | Read entrypoints, docs, graph indexes, then narrow to 1-2 modules. |
| Long stale sessions | Old context keeps hitchhiking into new work. | Summarize phases, create handoffs, start fresh when the goal changes. |
| Always-on skills | Skills load even when unrelated. | Trigger skills only when named or clearly needed. |
| MCP schema sprawl | Many tools make every request heavier. | Keep default MCPs minimal; enable high-power tools per task. |
| Browser side-channel leaks | Private pages get summarized by web assistants. | Read visible content locally; avoid third-party AI sidebars for private data. |
| Safety boundary drift | Convenience slowly normalizes risky actions. | Confirm before upload, delete, submit, push, deploy, or transmit sensitive data. |
| Missing handoffs | Cleanup risks losing project memory. | Write handoff docs before archiving sessions or worktrees. |
只读分析，不改文件，不调用外部提交/上传/删除。
先读 graphify/README/入口文件，给我范围建议，再决定是否深入。
只读取可见页面，不点击提交，不输入敏感信息，不调用网页 AI 总结私有内容。
只读检查配置和凭证引用，不输出密钥值，只列风险和修复建议。
- AGENTS.md
- Safe read prompt
- Repo scan prompt
- Browser safe prompt
- Codex safety audit prompt
- Codex maintenance report prompt
The visual language is a Japanese anime hand-drawn engineering notebook: delicate ink lines, watercolor texture, cyan for safe flow, amber for caution, and clear safety boundaries. The diagrams are intentionally original and do not reproduce OpenAI, Anthropic, Claude, or GitHub logos.
See Figma brief for the full design direction.
OpenAI, ChatGPT, Codex, Anthropic, Claude, GitHub, and related marks belong to their respective owners. This project is an independent educational guide. See BRAND_NOTICE.md.
Contributions should preserve the core contract: less context, clearer boundaries, and no secret leakage. Use small pull requests, avoid brand confusion, and include screenshots when changing the site.
