Hacktoberfest 2026：メンテナが10月に向けて印を付けた、オープンで初心者向けの issue。 Hacktoberfest の issue を見る

Agent Persona Exploration - 2026-10-27

クローズ 初心者向け
  • 何か問題がありますか？
#72,480 コメント 0 件 リアクション 0 件 担当者 0 名 GitHub で見る

メンテナーはふだん 1 日以内に返信

まだ誰も着手していません。

評価

難易度
2/5
見積もり時間
1〜3時間
初心者へのやさしさ
72/100
issue の種類
ドキュメント
明瞭さ
明確に書かれている
活発さ
活発
技術スタック
github-actions, go
領域
devtools

調査の方向性

推奨事項で挙げられているファイル .github/aw/create-agentic-workflow.md から始め、トリガーに関する指針と create-issue の重複排除に関する指針を確認します。イベント固有のラベルおよびパスフィルターに関するチェックリスト項目を追加し、古いissueを置き換える場合と履歴を残すために保持する場合を文書化し、プロンプトでは具体的なデフォルトの前提を明示すべきことを明確にします。ガイドですべての推奨事項が網羅されれば完了です。

索引モデルが issue の本文から書いたものです。

説明

agent-research
Persona Overview
  • Agent: agentic-workflows (persona-evaluator sub-agent)
  • Personas This Run: Program Manager, Designer, Information Worker
  • Scenarios Tested: 4 (of 6 generated this run)
  • Average Quality Score: 4.65/5.0
Key Findings
  • All 4 scenario invocations succeeded via the persona-evaluator sub-agent — no invocation failures this run.
  • The agent consistently front-loads ambiguity disclosure (undefined pairing rules, wiki scope, trigger path uncertainty) before committing to a design, rather than guessing silently.
  • Dedup design is a strong recurring pattern: sentinel HTML-comment markers for add-comment, file-path/dedup-key search before create-issue, and a deliberate choice between close-older-issues (replace-previous snapshot) vs. one-issue-per-event (preserve history) based on task semantics.
  • Security defaults were consistently solid: read-only agent-job permissions, explicit network: allowed: [] deny-by-default, and one scenario proactively added mentions: false / allowed-github-references: [] to suppress accidental pings from untrusted echoed content.
  • The weakest spot across all 4 responses is the same: draft prompts leave a core design placeholder unresolved (directory paths, pairing convention, wiki scope) rather than forcing a concrete default — reasonable caution, but it means none of the 4 outputs is directly shippable without a follow-up round-trip.
Top Patterns
  1. Triggers: schedule (weekly) + workflow_dispatch for digest/audit tasks; issues/push with paths:/labels: scoping for reactive single-item checks.
  2. Tools: github (gh-proxy, default toolset) for GitHub reads/dedup search; bash for deterministic local analysis (git log diffing, link grep) in preference to asking the model to infer facts it can compute cheaply.
  3. Security practices: minimal read-only permissions on the agent job in every scenario; explicit network deny-all by default rather than relying on prompt wording; untrusted-content handling called out explicitly for user-authored text (meeting notes, wiki pages).
View High Quality Responses (Top 2-3)
  • IW-6 (broken internal links): Opened by explicitly flagging that "wiki" is ambiguous (GitHub wiki vs. in-repo docs vs. external system) before picking a design, and added mentions: false / allowed-github-references: [] specifically because wiki content could contain accidental @name/#123 text — a non-obvious, well-targeted safeguard.
  • PM-3 (dependency risk triage): Clean sentinel-marker dedup for add-comment so repeated labeled events don't re-nag, paired with conservative "don't flag if phrasing is ambiguous" heuristics.
  • DS-6 (component doc drift): Correctly rejected a PR-reactive trigger (misses time-based drift with no further code changes) in favor of a scheduled digest, and separately reasoned about trivial-commit noise (formatting/bot commits inflating false positives).
View Areas for Improvement (Top 2-3)
  • All 4 draft prompts ship with at least one unresolved placeholder (exact directory paths, source-to-doc pairing rule, meeting-notes authoring path, wiki scope) instead of committing to a stated default assumption the user could quickly confirm or override.
  • PM-3's trigger relies on an in-prompt early-exit for the issues: opened path because GitHub's labels: trigger filter doesn't apply to opened — every new issue in the repo invokes the agent regardless of label, which is a real (if cheap) cost worth calling out more prominently in authoring guidance.
  • None of the 4 scenarios considered a "resolution" follow-up (e.g., acknowledging/closing a prior flag once the gap is fixed) — each only prevents re-flagging, not closing the loop.
Recommendations
  1. Add a short checklist item to .github/aw/create-agentic-workflow.md reminding authors that an issues:/pull_request: trigger with a labels:/paths: filter only applies to the label/path-changing event type (e.g. labeled), not to opened/co-occurring events — so the prompt must still perform a cheap early-exit check, and this should be flagged as a known trigger-design gotcha rather than left implicit.
  2. Document the "replace-previous-snapshot" vs "preserve-history" create-issue dedup decision explicitly in .github/aw/create-agentic-workflow.md (when to use close-older-issues: true vs. a stable per-item key) — this run's agent re-derived the right answer each time from first principles, which suggests the guidance isn't yet centralized.
  3. Encourage the agent to propose one concrete default assumption (clearly labeled as an assumption to confirm) instead of leaving a bare placeholder in the draft prompt, so a single round-trip with the user is enough to make the design shippable.

Generated by 🎭 Agent Persona Explorer · claude · agent · ~146 AIC · ⌖ ~1 AIC · ⊞ 4.5K · ◷

主要言語
Go
スター
5.5k
フォーク
606
平均マージ
4時間 18分
マージ済み PR（30日）
796

環境構築

Codespaces で開く

このプロジェクトの開発コンテナを、あなたの GitHub アカウントでブラウザ上に起動します。

はじめの一歩

  1. issue を最後まで読み、次にプロジェクトのコントリビューションガイドを読みます。
  2. 着手することを issue にコメントします — 二人が同じ作業をするのを防げます。
  3. リポジトリをフォークし、ブランチを切って変更します。
  4. issue 番号を参照したプルリクエストを送ります。

github/gh-aw のほかの issue

github/gh-aw の issue をすべて見る

似ている issue

Go の issue をもっと見る

新しい issue をメールで受け取る

初心者向けの GitHub issue を短くまとめたダイジェスト。