Real output from Claude Code with this plugin, on 2026-10-11:
> Which Node.js release line is the current Active LTS, and is that line installed on this Mac? Grade the sources. Answer in English, at most 5 lines plus the legend.
Active LTS today is Node.js 24 ("Krypton"), and this Mac has it: v24.14.0.
- Active LTS is line 24, and line 26 is not LTS yet [D1] (nodejs/Release `schedule.json` and nodejs.org `dist/index.json`, both read through WebFetch)
- Line 26 becomes Active LTS on 2026-10-28 [D2] (`schedule.json` through WebFetch)
- v24.14.0 is installed and active through fnm [A2] (`node --version`, `which -a node`)
- Your copy is behind the newest 24.x release, v24.21.0 [D2] (`dist/index.json` through WebFetch)
Grades: letter = source (A seen here, B primary, C secondary, D can distort, E wrong before, F unknown). Digit = confirmation (1 two checks, 2 seen once, 3 inferred, 4 no evidence, 5 contradicted, 6 unchecked).
Agents state a guess and an observation in the same confident voice. With this skill, each claim carries a grade, so you verify the weak claims first.
- Grade each claim: a letter for the source and a digit for the proof.
- Separate seen from inferred:
Ais only for what the agent saw in this session. - Catch silent passes: an empty output or a timeout never counts as proof.
- Name the next step: a weak grade comes with the step that verifies the claim.
- Preload into subagents: list the skill in the
skillsfield of a subagent.
Claude Code:
claude plugin marketplace add zyx1121/admiralty-code
claude plugin install admiralty-code@admiralty-codeCodex:
codex plugin marketplace add zyx1121/admiralty-code --ref main
codex plugin add admiralty-code@admiralty-codeAsk "grade the sources", or type /admiralty-code in Claude Code or $admiralty-code in Codex. The next report puts a grade such as B2 after each claim. For Antigravity, Cursor, OpenCode and other skill hosts, see INSTALL.md.
The letter grades the source. A higher letter is weaker.
| Grade | FM 2-22.3 term | Sources in agent work |
|---|---|---|
| A | Completely reliable |
A command, file or response that the agent saw in this session |
| B | Usually reliable |
Official docs for the exact version, pinned source code, a log from the user |
| C | Fairly reliable |
Memory notes, repository docs, issue threads |
| D | Not usually reliable |
LLM page summaries, search snippets, recall from training |
| E | Unreliable |
A source that was wrong on this topic before |
| F | Cannot be judged |
An unknown site or a paste with no origin |
The digit grades the claim. A higher digit is weaker.
| Grade | FM 2-22.3 term | Test in agent work |
|---|---|---|
| 1 | Confirmed |
2 independent sources or tools agree |
| 2 | Probably true |
The agent saw the claim once |
| 3 | Possibly true |
The agent inferred the claim |
| 4 | Doubtful |
No evidence supports the claim |
| 5 | Improbable |
Other evidence contradicts the claim |
| 6 | Cannot be judged |
Nobody verified the claim |
To preload the skill into a Claude Code subagent, add it to the frontmatter of the agent:
skills:
- admiralty-code:admiralty-codeThe 9 rules for the grades are in SKILL.md.
None. The agent loads the skill when a request matches its description. To grade every report, add the snippet from INSTALL.md to your agent instructions.
The skill is 1 Markdown file, and no code runs. The agent reads it when you call it or when a request matches the description. A subagent reads it at start when its skills list names the skill.
Facts for coding agents and LLM tools. The rules for changing the code are in AGENTS.md.
- Summary: one agent skill that puts an Admiralty Code grade (source A to F, credibility 1 to 6) after each claim in a report.
- Type: plugin
- Interfaces: skill
admiralty-code(/admiralty-codein Claude Code,$admiralty-codein Codex,admiralty-code:admiralty-codein a subagentskillslist) - Install:
claude plugin marketplace add zyx1121/admiralty-code && claude plugin install admiralty-code@admiralty-code. Other hosts: INSTALL.md. - Run: ask "grade the sources", or type
/admiralty-code - Verify:
claude plugin listshows the plugin, and the next report has a grade such asB2after each claim - Config: none
- Secrets: none
- Limits: the grades show the evidence that the agent reports. They do not prove that the agent reports honestly.
git clone https://github.com/zyx1121/admiralty-code && cd admiralty-code
claude plugin validate .Edit skills/admiralty-code/SKILL.md. Keep the manifests in plugin.json, .claude-plugin/, .codex-plugin/ and .agents/plugins/ in agreement.
The A to F and 1 to 6 meanings come from US Army FM 2-22.3. The British Admiralty used the scheme for naval intelligence, and NATO uses it today (Wikipedia). The repo layout comes from zyx1121/i-have-dyslexia.