メニュー
Build or refine harness context - skills, runbooks, or specs. Interview for shape, identify the right layer, study existing files for voice, draft, iterate.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
Sign the Firefox Browser Bridge extension via Mozilla AMO
Record progress - update by default, complete a topic with "done", record everything with "full".
Orient a new session — load memories, announce on the roster, follow memory pointers, report what you see
Name and describe unnamed Claude Code sessions, one at a time with confirmation.
Draft a commit message for the current changes.
Run the lint pipeline and fix issues with judgment. Don't sweep errors under the rug.
| name | build-context |
| description | Build or refine harness context - skills, runbooks, or specs. Interview for shape, identify the right layer, study existing files for voice, draft, iterate. |
| argument-hint | [name or description] |
Three layers hold what the next Claude needs to know. Each has a different job. Placing something in the wrong layer makes it invisible or noisy.
Skills - the choreography. Loaded into context when invoked, stays for the session. Two kinds:
Two homes: ${CLAUDE_PLUGIN_ROOT}/skills/<name>/SKILL.md
for general-purpose skills shared through the plugin
(invoked as /soil:<name>), and .claude/skills/<name>/SKILL.md
in the repository at hand for skills specific to it or its
infrastructure.
doc/ai/runbook/<name>.md - the operations manual.
Discrete procedures and lookups: which tools, which
arguments, taxonomies, conventions, known issues. Read
on demand, put down after. Multiple independent
operations in one doc, no single narrative. Don't
duplicate self-documenting surfaces - point at --help
or the spec for inventory and annotate only judgment,
when to reach for what. A copied list teaches its own
completeness, then drifts. Runbooks
backing plugin skills live at
${CLAUDE_PLUGIN_ROOT}/doc/ai/runbook/; repository
runbooks live in the repository itself.
doc/ai/spec/<name>.md - the rules. What's correct.
Coding conventions, naming patterns, deployment shapes,
file structure constraints. Prescriptive - defines right
and wrong. Doesn't change per session. Shared specs live
at ${CLAUDE_PLUGIN_ROOT}/doc/ai/spec/ (naming, packages,
error handling, lifecycle, and others); repository-specific
specs live in the repository's own doc/ai/spec/.
For each: general-purpose content goes to the plugin (a public repository - no internal hosts, names, or credentials); content specific to a repository or its infrastructure stays in that repository. Personal context - voice, preferences, working relationships - belongs in memories, not in files; always-tagged memories carry it into every session.
A skill points at its runbook for mechanics. A runbook points at specs for rules. None duplicate each other.
Read existing files - ls .claude/skills/ and
ls ${CLAUDE_PLUGIN_ROOT}/skills/ for skills,
ls doc/ai/runbook/ doc/ai/spec/ in the repository
and under ${CLAUDE_PLUGIN_ROOT} for documents.
Read 2-3 from each layer to feel the voice and density.
Know the skill format - frontmatter is optional
YAML between --- markers. Key fields: name,
description, disable-model-invocation,
allowed-tools, argument-hint, context.
Full docs: https://code.claude.com/docs/en/skills
Before drafting, understand the intent:
Don't assume. Two or three questions, then draft.
Match the voice of the existing files in that layer. Skills are direct, second person - instructions to the next Claude. Runbooks are operational - tool names, steps, tables. Specs are prescriptive - rules, patterns, rationale.
If the content spans layers, draft each side together on its own side of the line.
Present the draft. Corrections are the design. Iterate until it fits, then write.
When improving existing context:
