operator-skills
A Claude Code plugin of 10 skills, 117 short rules and 2 hooks. Each rule names one common failure in agent-assisted coding and prescribes a specific correction.
decision
A coding-agent operator decides whether to accept or override a correction during a build. Each rule names a recurring failure and when the agent should stay quiet.
baseline
A single long instruction file is simpler to maintain. Separate situation-specific files reduce irrelevant context, but no live comparison proves better routing or productivity.
review
The operator reads the suggested correction and can override ordinary guidance. These are prompt instructions, not an authorization system. Structure and phrase checks cannot guarantee that an agent follows them.
data
The portfolio shows invented situations without session transcripts. The linked package includes local activity logging hooks. A deployment must inspect those hooks and define access, retention and deletion before using real session data.
platform
The reusable pieces are rule files, skill descriptions and their structure validator. Coding agents consume the prompts. The 28 invented phrase checks test description consistency rather than the host agent routing or enforcing a rule.
limits
No live activation, adherence, speed or user-outcome evaluation. Prompted behavior is not an implemented security boundary.
Evaluation
Run 2026-10-30. npm test runs a structure validator (manifests, frontmatter, rule links, hook targets) and a routing test over 28 invented phrases, each paired with the skill and rule that should handle it. Two hooks ship with the plugin: one warns when a session ends in a worktree with unmerged commits, and one appends a line per subagent run to a local activity log.
Evidence: Structure validator | 10 skills, 117 rules, valid manifests and hooks
Evidence: Routing tests | 28 of 28 invented phrases route to the expected skill and rule (keyword overlap)
Evidence: claude plugin validate | Passes
Skills: claude-quality-watch | Agent drift: sycophancy, invented APIs and citations, unchecked claims about current state, being led by momentum | 8
Skills: engineering-floor | Secrets in chat or repos, in-place data changes, auth only in the UI, missing error and empty states, vague commit messages | 20
Skills: build-discipline | Output mistaken for outcome, missing brief, scope creep, polish before the feature works | 20
Skills: decision-discipline | Options with no pick, unstated trade-offs, no kill criteria, reversible and irreversible calls treated alike | 8
Skills: solo-run | Unattended runs without a written contract, stop conditions, log and debrief | 5
Skills: honesty-and-calibration | "Done" or "stable" claimed without outside use, rejecting an option without its strongest case | 4
Skills: claude-effective-usage | Wrong surface or tool tier, serial tool calls that could be parallel, polling, stale context | 10
Skills: claude-usage-discipline | Context and cost budgeting, handoffs, parking stray ideas, repeated re-prompts | 11
Skills: communication-craft | Hedged asks, weak closings, over-explaining, drafts that never pick a position | 23
Skills: strategy-and-business | Moats, unit economics, cost versus investment, recurring spend | 8
- It does not prove that a skill fires in a live session. The routing test is a consistency check on the written descriptions, not a measure of how Claude Code routes.
- Rules are written as prompts. They shape what the agent says, they do not enforce anything.
- No measurement of speed, accuracy or user outcomes has been made.