Skip to main content
HUMΛN
Developer
Developer

Resources and prompts — ground the agent without pasting the KB

HUMΛN Team··11 min·Technical (Developers)

Context windows are not a strategy.

Teams paste half the Knowledge Base into the system prompt, watch quality collapse, and blame the model. HUMΛN is the trust layer between people and AI — MCP resources and prompts exist so agents pull the right grounding when needed (docs, schemas, patterns) under the policies HumanOS enforces, instead of carrying the entire Canon like a guilt backpack.

The hook: the 40k-token “brief”

A builder dumps architecture markdown into the MCP client “so the agent knows HUMΛN.” The next tools/call is slower, vaguer, and somehow more confident about things that are not shipped.

Resource URIs reverse the default: start thin; fetch on demand.

What shows up on the wire

Typical HUMΛN resource families (exact URIs evolve with packs — treat these as the map, not a frozen catalog):

  • human://docs/… — public developer docs and AI articles
  • human://kb/… — governed KB slices the delegation is allowed to see
  • human://schemas/… — shapes agents should not invent
  • human://patterns/… — executable patterns, not vibes

Prompts such as design-delegation-scope and review templates turn recurring craft (scopes, PR checks) into named, reusable instructions instead of re-prompting folklore every session.

So that tools stay sharp

When the agent needs Canon language, it reads a resource. When it needs to propose scopes, it loads a prompt. When it needs to act, it tools/calls with a narrow delegation. Separation of grounding and acting is the craft.

Scroll-stopper: If your agent’s memory is “we pasted the wiki,” you do not have grounding — you have a landfill with embeddings.

Pair with Canon check

Resources without Canon gates still drift. Use this essay with Canon in the loop: pull the doc, edit, human.canon.check, then publish through the real surfaces.

Go deeper