メニュー
Use when writing Instagram caption variants and channel copy from media tags or analysis output. Reads brand from instance repo config. Do NOT use for selecting the final caption (edit-content) or vision analysis (analyse-media).
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
Web development practice setup interview — detects or creates .agency/target.json binding, reads instance profile from `/agency-hub:setup`, interviews tech stack, deployment platform, connectors, and persona preference (six engineering personas vs merged), writes practice profile. Use on first install after `/agency-hub:setup`, when the user says "set up web development" or "bind target repo", or to redo engineering defaults only.
Use when the user wants to run a Ralph loop: an autonomous, self-referential delivery loop that drives a whole epic through implement, code review, fix, UX review, validate, commit, and merge request, one task per iteration, until a completion promise is emitted (ralph setup, ralph start), or to inspect or stop a running loop (ralph status, ralph cancel). Also supports simple ad-hoc loops that repeat a single prompt until done. Do NOT use to implement a single task once (implement), review a diff (code-review), or sign off an epic (validate) — the loop orchestrates those skills.
Use when the user wants a UX/design review of implemented UI — frontend components, pages, or flows — against the design source of truth (Figma via MCP, mockups, tokens, or style guide, whatever the repo provides), covering design fidelity, accessibility (WCAG 2.2 AA), interaction states, responsiveness, and design-system conformity; or to address UX review findings (ux-design-review fix, optionally scoped to an action tier). Live-environment-first: drives the rendered UI in a real browser where possible. Works with any framework or design tooling. Do NOT use for code correctness/security review (code-review), to review a design.md document (design), or to implement features (implement).
Use when the user is the reviewer of a merge request or pull request and wants it reviewed and the verdict published to the provider — inline comments on the right lines, threads resolved or replied to, and an approve / request-changes / comment decision. Handles first reviews and re-review rounds after the author pushes. Works with any codebase and any git provider — GitHub, GitLab, or Bitbucket — using MCP tools where available, provider CLIs otherwise. Do NOT use to review your own working diff before opening an MR (code-review), to open or babysit an MR (merge-request), or to implement changes (implement).
Use when the user wants to open a merge request or pull request for the current branch (merge-request create — the default — with a generated title, description, labels, and reviewer suggestions), or to babysit an open MR/PR to a merge-ready state (merge-request babysit) by watching CI, review comments, and merge conflicts. Works with any codebase and any git provider — GitHub, GitLab, or Bitbucket — using MCP tools where available, provider CLIs otherwise, and plain git as a last resort. Do NOT use to review code or an MR as its reviewer (code-review, merge-request-review) or to implement changes (implement).
Use when the user wants a code review of a branch, PR, or diff against its acceptance criteria and declared scope (whatever form they take in this repo), or to address review feedback without changing behaviour (code-review fix, optionally scoped to a blocking or warning action tier). Works with any language, delivery process, or issue tracker. Do NOT use to implement code, sign off completion of a larger body of work, or review rendered UI/UX (design fidelity, accessibility — use ux-design-review).
| name | write-captions |
| description | Use when writing Instagram caption variants and channel copy from media tags or analysis output. Reads brand from instance repo config. Do NOT use for selecting the final caption (edit-content) or vision analysis (analyse-media). |
| license | MIT |
| allowed-tools | ["Read","Write","Glob","Grep"] |
| argument-hint | <tags json or media context> |
| metadata | {"version":"0.1.0","owner":"content-marketing","review_cadence":"quarterly","work_shape":"generate-draft","output_class":"draft-for-review"} |
Generate three distinct Instagram caption variants plus channel copy from media tags or analyse-media output.
edit-content)analyse-media)| Failure mode | Mitigation |
|---|---|
| Brand safety | First-person plural; cliché avoid list; no hashtags in body |
| Direct apply vs draft | Variants for editorial review — edit-content selects |
| DoD bypass | Three differentiated angles per prompt-refinement quality checks |
Follow prompts/run.prompt.md. Pass tags JSON or media context after the skill name.
JSON with caption_a (storytelling), caption_b (short/visual, max 150 chars),
caption_c (educational), plus google_post and email_excerpt channel copy.
Quality checks from ../../references/prompt-refinement.md: distinct opening words across variants; no embedded hashtags.