essay-cluster-from-transcript

これはレビュー時に保存された抜粋です。完全で最新の内容は外部ソースを確認してください。

---
name: essay-cluster-from-transcript
description: "Take a long stream-of-thought transcript (voice memo, video reflection, text dump) and turn it into a cluster of 5-20 publishable blog essays. Clusters ideas by theme, drafts each essay via parallel sub-agents, generates Gemini hero images, batch-commits, and (optionally) sends a grouped link list via the user's preferred messaging surface. Use when the user shares a transcript and says 'turn this into essays', 'write a series', 'split into posts', 'draft all of them', or pastes a long stream of ideas they want shipped."
---

# Essay Cluster From Transcript

## When to use

Trigger this skill when:

- The user shares a long stream of thought (voice memo transcript, free-form text, video reflection) and asks to publish from it.
- They say "turn this into essays", "write a series", "split into separate posts", "draft all of these", "ship the cluster".
- They paste many ideas in one message and ask for blog posts.

Do NOT use this for single-essay requests, polishing existing drafts, or extracting one quick takeaway. This is for the case where one transcript becomes 5-20 essays.

## Inputs the skill needs

Before drafting, confirm or read from the project's CLAUDE.md:

| Input | Where it usually lives |
|---|---|
| Blog repo path | Project Path Registry in user's CLAUDE.md |
| Blog post path convention | Repo's CLAUDE.md (e.g. `content/blog/YYYY-MM-DD-slug-en.md`) |
| Hero image path convention | Repo's CLAUDE.md (e.g. `public/images/blog/<slug>/hero.png`) |
| Frontmatter shape | An existing recent post in the same repo |
| Brand voice + style | Repo's CLAUDE.md (tone, words to use, words to avoid) |
| Hero image style guide | Repo's CLAUDE.md (palette, typography, motifs) |
| Default state for new posts | `hidden: true` if the repo has a draft mechanism, else publish directly |
| Messaging surface for the link list | User's CLAUDE.md (Telegram bot, Slack channel, etc.) |

If any are missing or unclear, ask one consolidated question — don't draft 17 essays in the wrong voice.

## Workflow

### 1. Cluster the transcript into a numbered list of titles

Read the full transcript end-to-end. Identify distinct ideas. Group them — typical structure:

- **Core philosophy / framework** (the load-bearing concepts)
- **Mechanics / how it works** (operational claims)
- **Examples / cases** (concrete material)
- **Personal / autobiographical** (vulnerable, first-person)

Aim for 5-20 essays. Fewer if the transcript is thin; more if it's dense. Each title should be a single concrete claim — not a category. Bad: "Thoughts on Productivity". Good: "I Stopped Treating Myself Like a One-Job Person".

Present the numbered list to the user grouped by theme. Suggest a 3-5 essay reading order for the impatient. Confirm before drafting.

### 2. Dispatch one sub-agent per essay (parallel)

For each essay, dispatch one sub-agent (in Claude Code: the `Agent` tool with `subagent_type: general-purpose` and `run_in_background: true`). The prompt for each agent must include:

- The exact markdown file path to write
- The exact hero image path to write
- The full frontmatter template (filled with title and slug; agent fills `description`)
- The voice/tone rules (cite from repo's CLAUDE.md, don't paraphrase)
- The relevant slice of the transcript as **source material** (verbatim — anchor the essay in the user's actual words, not your paraphrase)
- The specific essay angle for this post (one paragraph)
- The Gemini image generation command, with a prompt tuned to the brand style and this essay's metaphor
- Word count target (typical: 700-1100 words)
- Explicit instruction: **NO git commits**

Send all dispatch calls in a single message (parallel). The skill orchestrator does the git work after, not the sub-agents — otherwise they race on `git status`.

### 3. Verify and batch-commit

When all agents report back:

- Confirm each markdown file and hero image exists (`ls`, file size > 0).
- Spot-check one essay's