Claude Code's 1% context cap: Vercel plugin alone eats ~20K tokens, silently hiding your skills
thatroblennon · x · 2026-09-04
Rob Lennon published a detailed PSA on context budgets in agent harnesses:
- Claude Code allocates only 1% of context (8,000 tokens) and Codex 2% for system instructions, AGENTS.md/CLAUDE.md files, and every installed skill's name and description.
- When over budget, the app silently strips descriptions from least-used skills, leaving just names — the model then can't tell what they do and mostly ignores them unless called by name.
- Plugins are token hogs: Gmail's integration ships 29 MCP tools (12,000 tokens); Vercel's official plugin packs 32 MCP tools, 33 skills, 5 commands, 3 agents (17,000 tokens) plus a 2,700-token injected CLAUDE.md.
- A few skills plus a couple of plugins can push your whole setup past the cap, and since budgets are organized by usage frequency, newly installed tools are the first to go missing.
He shares a megaprompt that makes your agent audit the skill listing, report how far over budget it is, and apply fixes; his Personal OS update also slims skill descriptions and auto-triggers a context audit.
More from coding & agent
- Loop Engineering: a 4-loop framework for building AI agents that improve over time — MaryamMiradi · 2026-09-04
- XState author is porting state machines to Effect, with two Effect-native machine libraries emerging — DavidKPiano · 2026-09-04
- Vibe Coding Gets You 80% There: Stream Engineering Director on the Last 20% — amos_gyamfi · 2026-09-04
- Open-source remote MCP server for Obsidian/git Markdown vaults, 5-min deploy — rootException · 2026-09-04
- Agent cost benchmark on 114 tasks: money goes to retrieval, not reasoning — Coworker_ai · 2026-09-04
- LLM eval pipeline flaw: unsupported tool-call responses could score as perfectly stable — docybo · 2026-09-04