context-budget

Context Budget

Everything that is always on costs tokens on every turn: project memory, every installed skill's name and description, and every connected MCP server's tool schemas. Steer agents by placing each piece of guidance at the cheapest layer that still fires when needed.

Decision matrix

Layer	Loads when	Best for	Cost
Always-on memory (CLAUDE.md / AGENTS.md)	Every turn	A few durable, always-relevant rules	High (persistent)
Path-scoped rule	When files matching a glob are touched	Conventions specific to an area, such as backend/**	Medium
Skill (SKILL.md)	On demand when its description matches	Large how-to knowledge you do not always need	Low always-on (just name and description); full body only when triggered
Subagent	When delegated	Isolatable, parallel, or large-context work you want off the main thread	Low (separate context window)
MCP server	Its tools are offered every turn (schemas always-on); the call runs on demand	A live external system or action the repo and local CLIs can't provide — issues, a database, a browser, a SaaS API	Medium–high always-on (tool schemas) + network/auth
Hook	Deterministically on an event (PreToolUse, PostToolUse, UserPromptSubmit, Stop, SessionStart)	Enforcement that must not depend on the model remembering	Near-zero tokens; runs as code

How to choose

1. Is it enforcement that must always happen? Use a hook.
2. Is it always relevant and short? Put it in memory.
3. Is it specific to a path, package, or area? Use a path-scoped rule.
4. Is it large know-how needed only sometimes? Make it a skill with references.
5. Does it need isolation, parallelism, or a large scratch context? Delegate to a subagent.
6. Does it need a live external system or action — not local, not static? Connect an MCP server, but only the ones you use (their tool schemas tax every turn).

Keep the budget small

• Move detail out of CLAUDE.md / AGENTS.md into skills and references/ files so the full text loads only on demand.
• Prefer one well-scoped skill with progressive-disclosure references over many tiny skills whose descriptions all stay always-on.
• Delete stale memory instead of layering exceptions on top of it.
• Put deterministic checks in hooks, not prose the model has to remember.
• Connect only the MCP servers a task actually uses — each server's tool schemas sit in context every turn, like skill descriptions.
• Keep path-scoped rules crisp: local conventions, not full design docs.

More detail

• references/decision-matrix.md explains the layers, tradeoffs, and anti-patterns.
• references/mcp-as-context.md covers when to reach for an MCP server vs a skill, CLI, or subagent, and the always-on cost of its tool schemas.
• references/examples.md gives worked placement examples.
• ../../examples/ contains inert rule and hook templates to copy into your own repository.

Portability

The placement model works across GitHub Copilot / AGENTS.md-style hosts and Claude Code. Hook and rule mechanisms differ by host; examples/ uses Claude Code JSON as the concrete case while keeping the guidance host-neutral. Install with copilot plugin install context-steering@context-kit, apm install context-steering@context-kit, or /plugin install context-steering@context-kit.