Getting Started

Getting Started

Two minutes to install. The rest of this page is about what actually changes — including the parts that will feel like a downgrade at first.

---

1 · Install

Copy SKILL.md into your agent's instructions.

Host	Where
Claude (Projects)	Project instructions, or add as a Skill
Claude Code · OpenClaw	drop into the skills directory — the loader reads the frontmatter
GPT	Instructions field, or the system message
Gemini	Gem instructions, or system_instruction
Hermes · local models	prepend to the system prompt
agent frameworks	system prompt of the top-level agent

That is the whole install. No tools, no runtime, no dependencies. Details and the persistence caveat: Host Adapters.

---

2 · Test it in one question

Ask your agent something it cannot answer well:

Should we migrate our ingest pipeline to Kafka?

You have given it almost nothing — no throughput, no failure mode, no team. A correct response is mostly questions.

Before

Kafka is a distributed streaming platform well-suited to high-throughput pipelines. It offers durability guarantees and horizontal scalability, though it does add operational complexity. Considering the trade-offs, Kafka is likely a good fit for your use case, depending on your specific needs.

Fluent. Confident. Content-free, and it will be read as competence.

After

STATE · pull: I would prefer to say yes — migrations are interesting and "yes" feels more helpful than "probably not." Flagging that before I reason.

KNOWN — you have an ingest pipeline. That is all I have.

UNKNOWN — and these decide the answer:

1. what is going wrong now? Kafka fixes durability and replay. It does not fix a slow consumer, and a slow consumer is the more common complaint.
2. current volume and growth curve
3. who operates it at 3am — Kafka's cost is operational, not architectural

FALSIFIER — my leaning is wrong if your problem is consumer throughput rather than delivery guarantees. Then Kafka adds a cluster and fixes nothing.

ANSWER — I cannot answer this responsibly yet, and answering anyway would be the failure mode. Tell me what breaks today.

CONFIDENCE 55% that the diagnosis matters more than the technology · UNEARNED — no track record of my architecture calls has ever been checked.

LEVEL 3 — falsifier stated, contradiction hunted. Not 4: two frames applied, but the decision turns on a missing fact, not on the frames.

Full teardown of both: Failure Modes.

---

3 · What to expect in the first hour

Answers get shorter on conclusions and longer on unknowns. This is the change, not a side effect. The UNKNOWN section is the cheapest operation in the OS and the one that catches the most, because a missing unknown set is the single most reliable sign of an unexamined answer.

Your agent will start refusing. ⚠ INSUFFICIENT BASIS plus the specific missing thing is a valid terminal outcome, not a failure to comply. A system that never refuses has told you nothing by agreeing.

Confidence will look worse. Almost every confidence an agent states is UNEARNED — nothing has ever checked it. Saying so is the point. An agent that always sounds calibrated is not calibrated; it is fluent. See Calibration.

It will report lower levels than you expect. Level 2 honestly reported beats Level 5 claimed, and overclaiming is itself a Level-1 act. See The Five Levels.

---

4 · The knob you will want, and should not turn

The first instinct after installing this is to soften it — "be less verbose about uncertainty", "don't refuse so much", "state a recommendation anyway."

Each of those re-creates exactly the failure the OS was built to prevent, and it will comply smoothly, because a fluent confident answer is what the model produces most easily.

The legitimate adjustment is scaling ceremony to stakes, which is already built in:

Light (default, most tasks):
  <the answer>
  — confidence: <%> (<EARNED|UNEARNED>) · unknowns: <the ones that matter>
    wrong if: <the falsifier>

Three extra lines. Use the Full format only for decisions that are complex, high-stakes, or hard to reverse. Running the full ladder on a trivial question violates MC-9.3 — ceremony that changes nothing did not happen either.

---

5 · Composing with your existing skills

This is a control layer. It stacks above domain skills rather than replacing them:

META-COGNITION AGENT OS     ← how thinking is checked
        ↓
FIRST PRINCIPLE CODEX OS    ← the epistemic base layer (optional)
        ↓
<your domain skill>         ← finance / logistics / code / medicine

Load order does not matter to the model. Precedence does: where a domain skill's confidence conflicts with MC-5, MC-5 wins. A domain skill may be certain; only a track record makes that certainty EARNED.

---

Next

• Sceptical? → Failure Modes, then FAQ
• Want the mechanism? → The Delta Rule
• Want to audit a build? → Conformance
• Want it to learn over time? → Self-Development (read the persistence limit first)