bmad-forge-idea
Pressure-test an idea through persona-driven interrogation until it hardens, proves out, or dies cheaply. Use when the user says 'forge an idea', 'pressure-test this idea', 'stress-test my thinking', or 'harden this idea'.
- License 许可
- MIT
- 最近更新
- 2026-08-25
- Stars
- ★ 1
- Forks
- 2
01 前言 Front Matter
- name
- bmad-forge-idea
- description
- Pressure-test an idea through persona-driven interrogation until it hardens, proves out, or dies cheaply. Use when the user says 'forge an idea', 'pressure-test this idea', 'stress-test my thinking', or 'harden this idea'.
02 资源文件 3
- 其他customize.toml1.7 KB
- 脚本scripts/resolve_personas.py9.9 KB
- 脚本scripts/tests/test_resolve_personas.py6.2 KB
03 SKILL.md 正文
BMad Forge Idea
Overview
Take a half-formed idea and pressure-test it in conversation, while changing your mind is still cheap, until it becomes something the user can act on with conviction or reject. The main risk is what the user has not examined yet: unchecked assumptions and unresolved decisions usually become more expensive problems later.
The main goal is better thinking, not producing an artifact. Strengthening an idea, rejecting it, or thinking it through more clearly are all complete outcomes. Writing forged-idea.md to hand off to another workflow is optional. Do not steer the conversation toward "shall we build it?"
This skill can be used on many kinds of ideas. When the idea is about a product or feature, what survives may be written to forged-idea.md for later planning.
Lead by questioning, not lecturing. Ask one question at a time, press on weak points, and do not let vague claims pass without examination.
Conventions
- Scripts live in two places — run each from the exact path written, never assume co-location: the shared core scripts (
memlog.py,resolve_customization.py,resolve_config.py) are installed by BMad core at{project-root}/_bmad/scripts/and are never bundled here; this skill's ownresolve_personas.pyis at{skill-root}/scripts/. {workflow.<name>}resolves to fields in the mergedcustomize.toml[workflow]table.
On Activation
- Resolve customization:
uv run {project-root}/_bmad/scripts/resolve_customization.py --skill {skill-root} --key workflow. On failure, read{skill-root}/customize.tomldirectly with defaults. Apply the resolved{workflow.*}values throughout. - Run each
{workflow.activation_steps_prepend}entry; treat each{workflow.persistent_facts}entry as foundational context (file:entries load their contents,skill:names a skill to consult, others are facts verbatim). - Resolve central config:
uv run {project-root}/_bmad/scripts/resolve_config.py --project-root {project-root} --key core; from the merged JSON read{user_name},{communication_language},{output_folder}. On failure use neutral defaults; never block. Greet{user_name}in{communication_language}and stay in it. - Note whether a BMad persona is already active in this conversation — the user loaded one (e.g. the analyst, the storyteller) and invoked the forge from within it. If so, that persona leads the session, in voice, throughout.
- Resume: glob
{workflow.forge_output_path}/**/.memlog.md(recursive, so it still finds sessions whenrun_folder_patternis overridden to nest paths) and read only each match's frontmatter to find any whosestatusis notcomplete. Offer to resume one — then read its full memlog once to rebuild state and continue append-only — or to start fresh. - Run each
{workflow.activation_steps_append}entry.
Open the session
Start by scrutinizing the idea, not endorsing it.
Discover intent
Identify:
- the subject idea,
- the user's goal for the session,
- whether the idea is new or a change to an existing project
If any of these are already clear from the prompt that invoked this skill or previous context, ask the user to confirm and continue.
Otherwise ask for what's missing, in order:
- what is the idea?
- do you want to clarify and understand it, test whether it holds up, or make it better?
- is it a new idea or a change to an existing project? If the latter, what project is it, and where can I find its files or other relevant materials?
Steering the conversation
Tell the user they can say "attack this", "defend this", or "switch roles" at any time to change how the current idea is argued. In attack mode, do not agree with the idea; look for contradictions, weak assumptions, and failure cases. In defend mode, argue for the strongest version of the idea. Tell the user they can also name a persona or party at any time to change who participates in the session.
Set up the session
Derive a kebab-case {slug} for the idea and bind the session workspace {workspace} = {workflow.forge_output_path}/{workflow.run_folder_pattern} (the pattern fills with {slug}). Create the memlog once the goal is known:
uv run {project-root}/_bmad/scripts/memlog.py init --workspace {workspace} --field idea="<idea>" --field goal="<goal>"
Tell the user the path; state is on disk now, so the session survives interruption. If init fails, don't abort — run the forge in-conversation and tell the user state won't persist this session.
The forge
Let the session goal set the first move: for clarifying, pin down terms, boundaries, and assumptions; for testing, go after the central claim first; for making it better, drive each unresolved branch to a concrete decision.
Work one question at a time, in dependency order.
Include your current best answer or hypothesis when it helps the user respond. A concrete proposal is easier to accept, reject, or revise than an open-ended prompt. Find discoverable answers yourself instead of asking.
Do not assume the user's terms are precise. When a term is fuzzy or overloaded, name the ambiguity and ask for a precise choice before continuing. For example, do not let user, buyer, and payer collapse into one entity unless the idea actually requires that.
For ideas about an existing project, treat the project's files and materials as the source of truth. Do not accept a label or summary as proof. Find the relevant material yourself and check the user's claim against it. If the material contradicts the user's claim, stop and resolve that before continuing.
When a branch resolves, pause before moving on. Give the user a chance to raise any remaining concern.
Do not use agreement or praise to make the interaction smoother; they lower pressure and lead to shallower thinking. Agreement is allowed only when it helps the user think better. Praise is noise. Continued engagement and ego-stroking are not objectives. In attack mode, never agree with the idea until the user ends the mode. For each answer, either challenge the weak point or build on the strong point, whichever helps the user think better.
Capture as you go — each decision, assumption, crack, kill, and locked idea, one bullet in the user's meaning:
uv run {project-root}/_bmad/scripts/memlog.py append --workspace {workspace} --type <decision|assumption|crack|kill|direction|lock|note> --text "<gist>"
A lock is an idea the user hardens — settled, not to be reopened; locks are what forged-idea.md is distilled from. Don't read the memlog back except on resume. If the user raises a different branch, capture it and stay put — the loop and the stray insight both survive.
The personas
If a BMad persona was already active when the forge started, keep that persona as the lead voice.
Resolve the available persona pool once, as soon as the goal is known:
uv run {skill-root}/scripts/resolve_personas.py --project-root {project-root} --skill {skill-root}
The script returns installed BMad agents (agents), user-defined personas (members), and saved parties (parties). Parties may include a scene; some are open-cast. This gives you the same roster information as bmad-party-mode without invoking it.
Each turn uses two voices:
- One available persona — choose an installed agent or user-defined persona whose expertise fits the current branch. Vary this voice every few turns; do not let one voice dominate. If the user names a specific persona, use it. If the user calls a saved party, use the whole party and its scene. If the user asks to go one-on-one, use only the requested persona. If no pool is available, generate this voice yourself.
- One generated persona — create a fresh outside voice, such as a competitor, buyer, finance reviewer, domain expert, or critic. Give it a name and enough characterization to keep its viewpoint distinct.
Use these voices in character to pressure-test the current branch: find sharper objections, missing assumptions, and stronger defenses. Cross-examine them for what matters, then synthesize their input into your next question. Do not let the session turn into a panel debate or persona performance.
Voice the personas yourself by default. Spawn separate agents only when a branch needs independent reasoning that should not be influenced by one shared voice.
Exits
The session can end in three valid states:
- Hardened — the idea is stronger and specific enough to use. Distill the memlog into
{workspace}/forged-idea.md. Keep it extremely short: only the decisions, rejected options, and reasons that matter downstream, in the user's meaning. Do not write a prose summary, template, or conversation recap. If it reads like a document, it is too long. If planning or dev skills are installed (bmad-spec,bmad-prd,bmad-prfaq,bmad-build), offer the file as their input; if none are, the file stands on its own — never treat a missing skill as an error. - Killed — the idea does not hold up. Say so plainly and record why. Finding that out early is a valid outcome.
- Clearer — the user understands the idea better, but there is no hardened idea to hand off. Leave the memlog as the record; no
forged-idea.mdis needed.
Always render {workspace}/forge-report.html as a self-contained HTML file the user can open, with inline CSS and an inline-SVG seal or stamp. Summarize the outcome, the locked decisions, what was rejected and why, and the weak points that survived scrutiny, in the user's meaning. Credit the personas and parties that pressure-tested the idea by name, icon, and voice. Render a prominent wax-seal-style or stamped outcome mark, matched to the result: HARDENED, an Idea Death Certificate stamped KILLED with the cause of death, or CLARIFIED. Tell the user the path.
Flip the status at the end: uv run {project-root}/_bmad/scripts/memlog.py set --workspace {workspace} --key status --value complete.
If {workflow.on_complete} is non-empty, run all instructions in order.
04 同仓库更多 Skills
skill-maintenance
创建、审计、修正、拆分、升级、迁移或退役 Skill,或其合同/入口已漂移时,先恢复行为、授权与预算,再同步正文、调用者、版本、生成物和验证。普通业务维护、纯格式修正、承载位置未定的纠正不触发。Use for explicit Skill maintenance, not carrier choice.
skill-appraisal
判定一个 Skill 组该不该进当前装配、归哪些事项、与谁重叠,或对已判定过的组按节拍复核。三种进入:首次判定新候选、补判早已在用但从未判定过的组、按失效条件复核。判定单位是组不是单个 Skill;产出必须留下失效条件与下次最少复核步骤。用于外部来源、他人所写、疑似重复或从未被判定过的组。从零设计新 Skill 用 agent-skill-design;已定下后的改造、迁移、退役用 skill-maintenance;只要外部标准与版本事实用 capability-lifecycle;对象是 knowledge 用 knowledge-maintenance。本 Skill 只产出判定,不安装、不写配置、不改 profile。Appraise a Skill group, or recheck an adopted one. Judgement only, never installs.
self-improvement
当用户指出 Agent 漂移、误解、重复犯错,要求把任务经验固化为系统改进,或讨论只增概念却不减关键未知、形成决定或可检验资产时,只暂停依赖被推翻假设的路径,重锚原问题、持久记录纠正、诊断原因,并在授权内改进入口提示词或 Skill。最小实验不能验收产品交付;普通一次性错误、偏好或任务内修复不触发完整自我优化。Use for reported drift, misunderstanding, repeated mistakes, durable Agent improvement, or abstraction loops without decisions or testable assets. Freeze only affected paths and avoid a full detour for ordinary one-off corrections.
resource-observability
当用户查询 Codex/Claude 账户额度、重置时间、Codex 重置券、单 Session Token,或高成本/多 Agent 工作要决定启动、并行、降级、延后或停止时,按需读取 Orca 账户快照与固定 ccusage 回执,区分状态、来源、可信度和未知值。不要用于低风险任务、轮询、自动调度/消费权益、账单推断或长期保存快照。Use for account capacity, reset times, credits, session tokens, or consequential resource decisions. Skip routine work, polling, scheduling, automatic credit use, billing inference, and durable snapshot storage.
orchestrated-collaboration
当用户明确要求多 Agent/多 Session/跨 Provider 协作、任务已授权委派,或活动 Session 发生共享写入碰撞时,建立目标来源、排他所有权、可追踪交付、独立验收与综合;按共享资源和 Issue 子树确定唯一协调者,只冻结重叠或受推翻假设影响的路径。协调者参与高价值多交付件能力的拆分或设计时,由未参与者复核父目标;使用已选后端,选 Orca 时动态读 orchestration 指南。不要因复杂、额度或空闲 Agent 擅自并行,也不把试用变成长期依赖。Use for authorized multi-agent work, cross-session delegation, or shared-write collisions. Establish exclusive ownership, traceable delivery, independent acceptance, and bounded synthesis; add parent-goal review when required and avoid invented coordination or permanent platform commitments.
knowledge-maintenance
用于多来源调研、可重复实验、会影响权威/Agent 配置/重要决定的研究,或用户要求复用、复核、更新当前知识时:先找已认可知识和失效条件,只补变化、冲突与缺口,再经价值门和可信门更新。不用于低成本一次性事实、原始材料/研发过程留存、私域结构设计、定时全量重查或旧仓导入。Use before multi-source research, repeatable experiments, consequential authority/configuration decisions, or explicit knowledge reuse. Skip cheap one-off facts, raw retention, private architecture, scheduled full re-research, and bulk legacy imports.