원클릭으로
write-docs
expert: "First-draft technical documentation following inverted-pyramid
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
메뉴
expert: "First-draft technical documentation following inverted-pyramid
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
Use at the start of every agent session, even if the task seems obvious and you think you already know what to do. Activates when the agent has just opened a repository, does not know what to do, needs to orient itself, or is beginning any new session. Also activates for "where do I start?", "what should I work on?", "what is the current state?". Skipping orientation leads to wasted effort and missed context.
expert: "Structured task execution procedure for individual implementation
expert: "Multi-agent development orchestration with conflict-aware parallel
expert: "Review orchestration with adaptive specialist dispatch,
Use at the start of every agent session, even if the task seems obvious and you think you already know what to do. Activates when the agent has just opened a repository, does not know what to do, needs to orient itself, or is beginning any new session. Also activates for "where do I start?", "what should I work on?", "what is the current state?". Skipping orientation leads to wasted effort and missed context.
expert: "Structured task execution procedure for individual implementation
| name | write-docs |
| description | {"expert":"First-draft technical documentation following inverted-pyramid structure with audience-aware progressive disclosure and implementation-verified accuracy","natural":"Write a first draft of a document — plan it, structure it, write it following the inverted pyramid"} |
| triggers | ["write documentation","draft a document","create a README","write a getting-started guide","write a manual","author technical documentation"] |
| roles | ["documenter"] |
| stage | documenting |
| constraint_level | medium |
refs/documentation-structure-guide.md §7Write a single sentence that captures what the document must accomplish. This sentence constrains every decision that follows — what to include, what to cut, how deep to go. If you cannot write this sentence, you are not ready to write.
Choose the structural template that matches the purpose: README, getting-started
guide, manual, reference, or design document. Read the structure for your type in
refs/documentation-structure-guide.md §7. The template gives you a section
order; do not invent one from scratch.
Who is the reader? What do they already know? What are they trying to accomplish? Write these down explicitly. All our documentation assumes Git fluency and command-line comfort. State any assumptions beyond these for this specific document.
These are the things the reader must understand after reading. They become the spine of your structure. Everything in the document supports one of these messages. Everything that does not will be cut.
Write each section's content as a complete sentence, not a keyword. Sentences force verbs and therefore meaning. Compare: "Configuration" tells you nothing; "The user creates a configuration file that controls which features are enabled" tells you exactly what the section says. The outline also reveals gaps — places where you do not yet know what you want to tell the reader.
Code examples, diagrams, tables, and transcripts are often the most valuable part of a document. Draft them before writing prose. Prose exists to connect and explain the examples, not the other way around.
At every level — document, section, paragraph — place the most important information first. Follow the tone gradient: accessible and conversational at the top, precise and formal at depth. Each section opens with its key point; each paragraph opens with its topic sentence.
Check every factual claim against code, configuration, or runtime behaviour. Design documents provide concepts and intentions but must not be trusted as the sole source of truth. Record what you verified and flag anything you could not check.
Summarise what you actually wrote, not what you planned to write. The opening must answer three questions: what is this, who is it for, what will you learn. Write it only after the body is complete.
The Write stage edits the document file directly — it does not just produce a report. Write the draft (or revise the existing document) in place, then produce the metadata and verification notes below as a record of what was produced. When the input is an existing document, edit the file to improve it rather than returning an assessment of what could be improved.
## Document Draft
**Purpose:** {one-sentence purpose statement}
**Type:** {README | getting-started | manual | reference | design}
**Audience:** {who this is for and what they already know}
**Key messages:**
1. {message}
2. {message}
3. {message}
---
{document body following the chosen structural template}
---
**Verification notes:**
- {list of facts verified against implementation}
- {list of facts that could not be verified — flagged for check stage}
## Configuration
The configuration system was originally designed in Q3 2024 as part of
the platform modernisation initiative. It replaced the legacy INI-based
system that had been in use since version 1.2. The team evaluated several
alternatives including TOML, JSON, and YAML before settling on YAML due
to its readability advantages. Configuration files are stored in the
`.config/` directory.
WHY BAD: The reader scrolls through four sentences of history before learning where configuration files actually live. The key fact — the location and format — is buried at the end. A scanning reader stops before reaching it. This violates the inverted pyramid at the section level.
## Configuration
Configuration files live in `.config/` and use YAML format. Each file
controls one aspect of the system — see the sections below for details.
The system migrated from INI to YAML in v2.0 for readability. If you are
upgrading from v1.x, see the migration guide.
WHY GOOD: The key fact is first. The reader who stops after one sentence still knows where configuration lives and what format it uses. History appears only where the reader who needs it will find it — below the key point, clearly marked as upgrade context.
These criteria are for evaluating the document draft, not for self-evaluation during writing. They are phrased as gradable questions to support automated LLM-as-judge evaluation.