Promptless for Agent Instructions
Promptless for Agent Instructions governs the skills, subagents, hooks, commands, rules, and MCP configs your agents run on. It collects evidence from real agent sessions, finds where those instructions fail, and opens reviewed pull requests that fix them.
The instruction-drift problem
Section titled “The instruction-drift problem”Teams keep moving operational knowledge into agent-read files: skills, subagent definitions, hooks, commands, rules, MCP configs, and the AGENTS.md or CLAUDE.md context an agent loads. Once that knowledge steers real work, it is infrastructure, and it drifts like any other infrastructure. The result is instruction debt, sometimes called skill slop: a growing pile of instructions that quietly degrade the agents that read them.
The drift takes recurring shapes:
- Stale: the procedure moved on, but the instruction still describes the old one.
- Missing: a new workflow has no instruction behind it.
- Contradictory: two skills push an agent in opposite directions.
- Dead: an instruction points at a command, tool, or path that no longer exists.
- Unowned: no one is accountable for keeping the instruction correct.
- Untestable: nothing tells you whether the instruction still works.
- Unenforced: the instruction is there, and agents ignore it anyway.
- Undistributed: the canonical instruction never reaches the agent that needs it.
- Non-portable: the instruction assumes one agent runtime or format and breaks elsewhere.
- Risky: the instruction shifts permissions, autonomy, or escalation behavior.
These failures compound. Agents thrash, repeat mistakes you already corrected, and take actions you did not intend. Meanwhile, the people who own the instructions have no evidence pointing at the instruction actually at fault.
Who it’s for
Section titled “Who it’s for”Promptless for Agent Instructions serves the people accountable for an agent fleet across a company. That includes AI-governance leads, platform and AI-engineering teams, and engineering leadership. The payoff scales with the fleet, so it matters most once a team runs dozens of skills across many agents and hosts.
You may already run continuous, evidence-driven scanning over code and infrastructure (the posture that Semgrep, Wiz, Snyk, or Vanta take). This applies the same discipline to agent instructions. Findings come backed by real session evidence, not a one-off audit.
How the loop works
Section titled “How the loop works”You author and organize your instructions in an Instruction Hub, a versioned Git repository. The pig toolchain compiles that hub into target-native plugin packages, and your fleet installs them. On each host, a runtime enrolls with a worker you deploy and streams native session traces to it. Inside the worker, the Friction Analyzer studies those sessions and records findings, each backed by append-only evidence. Findings become GitHub issues, and high-confidence findings open remediation pull requests against your hub. A human reviews and merges the fix. That merge triggers the recompile and redistribution that closes the loop.
For the full end-to-end path, see How it works.