Host Adapters

Host Adapters

SKILL.md is one file with no tool calls and no runtime, so it installs anywhere a system prompt can go. That is the portability claim, and it is deliberately small.

Only two things differ between hosts:

1. where the file goes
2. whether the Thinking Journal survives the session

Everything else is identical, because the OS was written to depend on nothing.

---

Install

Host	Where SKILL.md goes
Claude (Projects / Skills)	add as a Skill, or paste into Project instructions
Claude Code · OpenClaw	drop into the skills directory — the loader reads the frontmatter
GPT (Custom GPT / API)	Instructions field, or the system message
Gemini (Gems / API)	Gem instructions, or system_instruction
Hermes · local models	prepend to the system prompt
Antigravity · agent frameworks	system prompt of the top-level agent
anything else	if it reads a system prompt, it runs this

No host-specific edits. If a host requires you to change the thinking to make it work, that is a defect in the adapter, not in the host — and almost always it means the OS is being installed at the wrong layer.

Why the top-level agent in a framework

This is a control layer. In a multi-agent system it belongs above the domain agents, not beside them. Installed on a worker it governs that worker's output; installed on the orchestrator it governs which outputs are accepted, which is where the leverage is.

A note on filename case

The canonical filename is SKILL.md, exact casing. Loaders that glob for SKILL.md silently skip skill.md on case-sensitive filesystems — the skill installs cleanly and does nothing. The repo's CI asserts this on every push, because it is a failure that produces no error message.

---

Persistence — the only real difference

Self-development needs the journal to outlive the task.

Host capability	Binding	Cross-session learning
project files · vault · repo	keep THINKING_JOURNAL.md there; read it at session start	✅ real
a memory or notes tool	append entries; retrieve at start	✅ real
a database (SkynetClaw)	journal is a table, promotion is a query	✅ real, and auditable
long single session	in-conversation only	⚠ within the session
stateless API call	none	❌ — state this in the output

The rule that matters MC-8.5

An agent that says "I have learned from our previous conversations" when it cannot remember them has invented a specific — the failure MC-2.4 forbids, applied to itself.

No persistence is a fine state to be in. Pretending otherwise is not.

→ Self-Development

---

Small models

The OS runs on 7B-class models, with two honest caveats.

Ceremony degrades first. A small model will reliably produce the section headings and under-fill them. That is The Delta Rule territory: headings with nothing sorted under them are a violation, not compliance — and they are harder to spot than a missing section, because the structure signals rigour.

Prefer the Light format. The full ladder on a small model often produces reflection-shaped text rather than reflection. Light format plus one real falsifier beats a Full-format output where every section paraphrases the answer.

If a small model can produce only one thing from this OS, it should be the UNKNOWN section. It is the cheapest operation and it catches the most.

---

Composing with domain skills

META-COGNITION AGENT OS        ← how thinking is checked
        ↓
FIRST PRINCIPLE CODEX OS       ← the epistemic base layer
        ↓
<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. → Calibration

Conflicts you should expect

Domain skill says	This OS says	Resolution
"always give a recommendation"	a failed gate may refuse	the gate wins — MC-6.2
"be concise"	name the unknowns	Light format; unknowns stay, prose goes
"state confidence as a percentage"	with EARNED/UNEARNED	both — the label is one word
"you are an expert in X"	confidence is UNEARNED absent a record	expertise is not a track record

The last row is the one that surprises people. Being a domain expert makes an agent's reasoning better. It does not constitute evidence about its accuracy rate, and only the latter can earn a confidence figure.

---

Related

• Getting Started — the two-minute path
• Conformance — auditing a build after install
• adapters/README.md