context-steering¶
Place guidance at the cheapest layer that still fires
A skill-only teaching plugin for deciding where guidance should live — always-on memory, path-scoped rules, on-demand skills, delegated subagents, MCP servers, or deterministic hooks — without turning every turn into a context-budget tax.
Install¶
Components¶
| Component | What it is |
|---|---|
context-budget skill |
A decision matrix for placing guidance in always-on memory, path-scoped rules, skills, subagents, MCP servers, or hooks while keeping the always-on budget small. |
examples/ directory |
Inert, copy-paste templates for path-scoped rules and Claude Code hook JSON — documentation examples, not active hooks or rules. |
The idea¶
Project memory (CLAUDE.md / AGENTS.md) and installed skill descriptions are
always in the agent's working set. Every always-on token competes with the actual
task, so put guidance where it costs the least:
| Layer | Use for | Cost |
|---|---|---|
| Always-on memory | short, durable, always-relevant rules | highest (always loaded) |
| Path-scoped rules | conventions for a specific area of the tree | loaded when that path is touched |
| On-demand skills | bulky know-how pulled in by task | loaded when relevant |
| Subagents | independent investigations in their own context | isolated |
| Hooks | deterministic enforcement | runs outside the model |
Ships nothing active
context-steering installs no live hooks or rules. The examples/ are
templates you copy and adapt.
At a glance¶
| Category | steering |
| Provides | 1 skill + inert examples |
| Dependencies | none |
| License | MIT |