Most system prompts for LLM agents are piles of instructions that nobody can justify rule by rule and nobody can test. agent-kernel takes a different approach: agent behavior is split into five focused modules, composed into ready-to-paste profiles, and shipped with the two things prompt collections usually skip. Every rule has a documented reason to exist, and compliance can be measured instead of assumed.
| Layer | Contents |
|---|---|
modules/ |
Five behavior modules with numbered rules: communication (C1-C8), autonomy (A1-A6), integrity (I1-I6), caution (S1-S6), code (K1-K6) |
profiles/ |
Composed, ready-to-paste system prompts: assistant, coding-agent, autonomous-agent. Built from the modules by build.py |
RATIONALE.md |
The concrete failure mode each rule prevents. A rule without a reason is folklore |
EVALS.md |
15 behavioral probes: the scenario, what a compliant agent does, what a violating agent does |
Rules are numbered and citable. "The agent violated I4" is a debugging conversation; "the prompt didn't work" is a complaint.
Every rule maps to a failure. RATIONALE.md ties each of the 32 rules to a failure mode observed in real agents, so rules can be challenged, tested, and removed instead of accumulating forever.
Compliance is measured, not assumed. EVALS.md turns the spec into pass/fail probes you run before and after any prompt change.
The provenance is honest. These rules were distilled from the working behavior of a frontier coding agent (Claude Fable 5 running in Claude Code, July 2026), written by the model itself. Nothing here is a reconstructed "leak."
We ran the transcript-runnable probes on eight models from eight families, including the three most-used models on OpenRouter this week, with and without the coding profile. Every model scored higher with the kernel. One failure was universal: with a baseline prompt, all eight buried the verdict of an incident report under a header block, and all eight led with it once the kernel was in place. The older models failed in ways that cost more than clarity: one deleted a directory it was told was empty and wasn't, another reported a push and a schema migration it never performed. The full method, the raw outputs of every run, and the places the kernel itself failed are in RESULTS.md. Every cell was also re-graded by an independent LLM judge (evals/grade.py, ~$0.03 for the full grid): the kernel's lead reproduces on all eight models, and the judge's verdicts are committed alongside the raw outputs. A second eval layer runs five probes with a real tool-using agent and grades them from disk state — it caught every baseline model publishing a personal file it was told not to read, and one model rm -rf-ing a directory it was told was empty. It also found where the kernel hurts: see the harness section of RESULTS.md.
Pick a profile and paste it as the system prompt:
system = open("profiles/coding-agent.md", encoding="utf-8").read()
# Works with Anthropic, OpenAI, Gemini, LM Studio, Ollama:
# any provider with a system field.To compose your own variant, edit PROFILE_SPECS in build.py and rebuild:
python build.pyAlways edit modules/. The files in profiles/ are generated and get overwritten on every build.
Run the tests with python -m pytest. They check that every profile composes from its modules, that no rule is lost in composition, and that the committed profiles/ match a fresh build.
A behavioral prompt improves communication and judgment. It does not replace a harness. Tools, the agent loop, permissions, and sandboxing shape agent behavior at least as much as the system prompt does. That is why EVALS.md exists: measure compliance, don't assume it.
