Context Steward — AI skill for Claude Code

context-steward

**Load skills dynamically. Learn what works. Works with any LLM.**

MCP server for Claude Desktop, Claude Code, and Cursor.

---

The problem

You have 10 skill files. Your agent loads all of them into the system prompt. That's 15,000 tokens burned before the task starts — whether the agent reads them or not.

On a 32K model, that's 47% of your context gone. And you have no idea which skills actually help.

How it works

Lazy loading

Skills are MCP tools. Zero skill content in the initial prompt.

When your agent hits a task, it calls:

load_skills({ task: "refactor the auth module" })

context-steward finds the relevant skill, returns it with a `contextId`, and content enters context only when needed — right before generation. The `contextId` is a handle: pass it back later to link outcome feedback to the specific skill that was used.

Feedback loop

After the task, report what happened — not a score:

report_outcome({
  contextId: "abc123",
  signal: "praised",
  notes: "Clean decomposition, all types correct, user said 'perfect'"
})

**Signals** are observable conversation events:

Signal	When to use	Derived score
praised	User explicitly said good/great/perfect	0.95
used_as_is	User accepted and moved to the next topic	0.70
revised	User asked for specific changes	0.40
rejected	User said no, start over, dismissed output	0.15
redone_by_user	User did it themselves after seeing the attempt	0.10

Why signals instead of scores? Because a model scoring its own work is unreliable — it will always be generous with itself. Signals are binary observations: did the user accept it or not? Did they ask for changes or not? No subjectivity in the observation.

The "derived score" column is a deterministic mapping used only for aggregation and ranking. It's not a judgment; it's a sort key.

Over time, skills accumulate signal history:

$ context-stewa