Frontier Operating Manual — AI skill for Claude Code

Frontier Operating Manual

An operating discipline for AI coding agents and the people who direct them: how to read the real request, decompose work into checkable pieces, spend verification effort in proportion to risk, label what is known versus guessed, attack a conclusion before shipping it, and stay anchored across long tasks.

It ships as a Claude Code skill. The skill file is the always-on hot layer, about seventy lines. Four reference files hold the full procedures, worked examples and rubrics, about sixteen thousand words.

Why it exists

Capable models fail on hard tasks in repeatable ways: answering a nearby easier question, treating a request's premise as fact, re-reading a source and calling it verification, sounding more certain than the evidence, drifting from the original ask over a long session. None of these is a knowledge gap. Each is a discipline gap under load. This manual encodes the discipline as standing rules with pass or fail checks, so a session can be held to them and audited against them.

What is in it

File	Holds	Load when
SKILL.md	The operating loop, ten standing rules, risk tiers, calibration vocabulary, the five-question pre-send self-test, hard stops	Always. This is the hot layer.
references/core-competencies.md	S1 to S8: reading the request, decomposition, risk-weighted effort, verification by re-derivation, known versus guessed, self-attack, answer-first communication, competence mimics	Starting a hard task, or suspicious of a draft
references/long-horizon.md	S9: the ledger, re-anchoring, thread discipline	Tasks spanning hours or days, context compaction, drift
references/gates.md	S10 the adversarial QC audit, S11 the self-test rules	Shipping outward-facing or irreversible work
references/appendices.md	A to F: the risk rubric, label rules, symptom library, escalation triggers, an evaluation harness, versioning	Exact anchors and thresholds

The ten