無料GitHub で公開

my-code-write

Delegate writing predictable code to the `code-write` command, which has a cheaper writer model write the target file directly so the generated code never enters your context. Use for tests, scaffolding, fixtures, boilerplate, or a new file that copies an existing pattern, and when the user says "code write", "生成を委譲", or "雛形を書かせて". Do NOT use for design decisions, changes to public contracts or shared modules, edits inside an existing file that needs judgment, or logic with subtle correctness requirements.

Ynakatsuka/dotfiles

インストール方法を見る

含まれるファイル（1）

SKILL.md3.6 KB

SKILL.md（原文）

インストールする前に、エージェントに与えられる指示の中身を確認できます。

Code Write Delegation

When to delegate

The result follows a pattern that an existing file already shows: a test for another module, a handler or model shaped like a sibling, a fixture, or a scaffold.
The spec can be stated completely up front, including the file path, the behavior, and the reference file to imitate.
The file is new or is owned entirely by this task, so a full rewrite is acceptable.

Write the code yourself when the change needs a design decision, touches a public contract, or must be woven into an existing file with judgment.

Command

code-write --spec "<complete spec>" --target <file> [--reference <file>]...

State the spec completely: what to cover, names to use, framework, and edge cases. The writer cannot ask questions.
Pass one or two reference files that show the exact pattern to copy. Each is sent verbatim.
The target must be inside the current working directory. An existing target is sent to the writer and rewritten as a whole.
Stdout carries the written path with its line count and a short summary; stderr carries progress. A non-zero exit means nothing usable was written. Report the error instead of retrying with the same input.

Collecting completion

This command is a shell process, not a spawn_agent task. Do not use wait_agent, clock.sleep, or shell sleep to wait for its result, and do not rely on a completion notification.
In Codex, when exec_command returns a session_id, collect output with write_stdin using that ID and empty chars. Use yield_time_ms: 1000 while waiting for this result; repeat while a session_id remains, until an exit_code and the final output arrive. The tool may enforce a longer minimum wait, but returns early when the process finishes.
functions.exec saying Script completed only ends the JavaScript wrapper. An inner session_id still needs collection. If the wrapper instead yields a running cell_id, resume it with functions.wait and yield_time_ms: 1000, then inspect the inner result.
Do independent work while the writer runs, but check its result at the next tool boundary. When its result is the only remaining dependency, keep collecting directly without an extra delay. Once the exit code and summary arrive, proceed to verification and stop waiting.
In Claude Code, collect a background Bash task with TaskOutput using its returned task ID; a finished task's output should be read immediately rather than waiting for another notification.
A target file appearing or changing is not proof of completion. Collect the command's exit status before reading, editing, or testing the target.

After the write

Run the narrowest relevant check on the target, such as the new test file or the linter, instead of reading the file to review it.
Confirm with git status that only the target changed.
When a check fails, read the failing region with a bounded range and fix it yourself, or re-run once with a more precise spec. Do not loop on the writer.

レビュー

まだレビューはありません。使ってみた感想をお寄せください。

ログインしてレビューを書く

同じリポジトリのスキル

概要と使いどころ

my-agent

無料日本語概要

Delegate tasks to an external CLI agent (OpenAI Codex, Google Gemini, or Claude Code) for second opinions or parallel execution. Use when the user explicitly mentions "codex", "gemini", or "claude" or asks to delegate work to one of them (e.g., "codexで実行", "codexに聞いて", "geminiで実行", "Geminiに聞いて", "claudeで実行", "Claudeに聞いて", "run with codex", "ask gemini", "ask claude"). Also use when the user mentions Claude's "fable" model alias for delegation or consultation (e.g., "fableで実行", "fableに相談", "Fableに聞いて", "ask fable"). Do NOT use for general coding tasks that don't mention codex, gemini, claude, or fable.