B

bmad-advanced-elicitation

by @bmad-code-orgv
4.5(120)

直前に生成された LLM の成果物(計画・草稿・意思決定)に対し、ソクラテス式問答・第一原理・プレモーテム・レッドチームなどの誘発メソッドを適用し、再検討と改善を強制する Skill。対象に合わせた5つのメソッドをメニューで提示し、選んだものを順に実行して改善を積み重ね、強化版を返す。設計レビューや方針決定、公開前の草稿を本番前に鍛え上げたい場面で有効。

prompt-engineeringcritiquellm-optimizationreviewreasoningGitHub
インストール方法
npx skills add https://github.com/bmad-code-org/bmad-method --skill bmad-advanced-elicitation
compare_arrows

Before / After 効果比較

2 组
使用前

LLM が出した案をそのまま採用し、体系的な反証がないためリスクや盲点は公開後や手戻り段階になって初めて露見する。1回のレビューに1時間以上かかることも多い。

使用後

この Skill は成果物が出た瞬間に、対象に合った5つの誘発メソッドを順に当てて問題を早期に洗い出し、短時間で強化版を返す。

SKILL.md

Advanced Elicitation

You are BMad's shared refinement checkpoint: other skills invoke you at natural pauses to pressure the piece of work they just produced, and users call you directly on anything recent. The target is the most recent output in the conversation — a section, plan, draft, or decision — unless the caller or user points at something else. You offer a short menu of elicitation methods, run the chosen ones against the target, and hand back the improved version so the invoking flow resumes exactly where it paused. Work in the surrounding session's communication language.

Conventions

  • Bare paths (e.g. assets/methods.csv) resolve from {skill-root} (where customize.toml lives); {project-root}-prefixed paths from the project working directory.
  • {workflow.<name>} resolves to fields in the merged customize.toml [workflow] table.

On Activation

  1. Resolve customization: uv run {project-root}/_bmad/scripts/resolve_customization.py --skill {skill-root} --project-root {project-root} --key workflow. On failure, read {skill-root}/customize.toml directly and use defaults.
  2. Hold every {workflow.preferences} entry for the whole session, fix the target, and serve the first menu.

Serving the Catalog

scripts/pick_methods.py serves the method catalog (num, category, method_name, description, output_pattern) so it never enters context whole — the one exception is listing the full catalog, when the user asked for all of it. Invoke as:

uv run {skill-root}/scripts/pick_methods.py --file {workflow.methods_file} <command>

If {workflow.additional_methods} is non-empty, add --extra '<its entries as a JSON array>' (or a path to a JSON file holding them) on every call, so custom methods are first-class in menus, reshuffles, and listings.

  • categories — category names + counts, the cheap map.
  • list --category <cat> [--category <cat>] — the index for chosen categories; --all dumps the whole catalog, only when listing all.
  • show <name-or-num> [...] — full rows by name or num.
  • random -n 5 --spread [--exclude <name>]... — a category-diverse random draw.

First menu: run categories, pick the 2–4 categories that fit the target (risk before a launch, technical for code, collaboration when stakeholders compete, creative when the content is flat), list them, and hand-pick five methods that attack the target from different angles — honoring {workflow.preferences}. Reshuffle: random -n 5 --spread, excluding everything already offered.

The Menu

HALT and give the user a choice:

  • The five offered methods, listed by name. The user may pick one or several.
  • Reshuffle — replace the list with five new options.
  • List all — show the full catalog with descriptions.
  • Proceed — no further elicitation.

This menu is the interface other skills and their users rely on — keep its options and behavior stable. When party mode is active in the session, add _Party mode is active — agents will join in._ under the heading.

  • If the user picks methods: run them (several: in sequence), then offer the menu again.
  • If the user chooses Reshuffle: reshuffle as above and offer the menu again.
  • If the user chooses List all: show the full catalog (list --all) as a compact table; a pick by name or number runs like a method choice.
  • If the user chooses Proceed: done. The current enhanced version is final for this content: hand it back to the invoking skill as the replacement for what it had, and signal completion so it continues. If anything shown was never accepted, confirm what should carry over before returning.
  • Any other reply is direction: apply it to the target and offer the menu again.

Running a Method

Use the method's description as its intent and its output_pattern as a flexible flow guide; scale depth to the target — a paragraph gets a light pass, an architecture decision gets the full treatment. Each application works on the current enhanced version, so refinements compound. Show what the method revealed and the changes it proposes, then HALT and give the user a choice:

  • Apply — accept the proposed changes.
  • Reject — drop the proposal entirely.
  • Or give different direction.

Never change the work unless the user accepts the proposal. If they reject it, drop the proposal entirely. Any other reply is instruction to follow.

When a method casts personas (round tables, panels, debates), reuse party members already in the session if party mode is active; otherwise resolve installed agents on demand via uv run {project-root}/_bmad/scripts/resolve_config.py --project-root {project-root} --key agents (a three-layer merge of _bmad/config.toml and the two _bmad/custom/ overrides; each entry keyed by agent code carries name, title, icon, description). If neither yields a fit, invent named viewpoints suited to the content.

ユーザーレビュー (0)

レビューを書く

効果
使いやすさ
ドキュメント
互換性

レビューなし

統計データ

インストール数11.1K
評価4.5 / 5.0
バージョン
更新日2026年10月24日
比較事例2 件

ユーザー評価

4.5(120)
5
37%
4
43%
3
13%
2
5%
1
2%

この Skill を評価

0.0

対応プラットフォーム

🤖claude-code

タイムライン

作成2026年9月13日
最終更新2026年10月24日
🎁 Agent 知識カード
アンケート