用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/rlaope/bestwork-agent --skill plan命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
Evidence-driven causal tracing — competing hypotheses, falsification, rebuttal round, recommended next probes
Feature validation gate — research, define purpose, stress-test from real user perspective before building
Maximum parallelism burst — all tasks execute simultaneously with no coordination overhead
基于 SOC 职业分类
正在显示 SKILL.md
| name | plan |
| description | Analyze task scope and recommend a bestwork team composition before executing |
When this skill is invoked, IMMEDIATELY print:
[BW] analyzing scope...
The user's request follows after the skill invocation. Analyze it thoroughly before doing anything.
Gather concrete data about what the task requires. Do ALL of the following:
Identify affected files — based on the user's request, determine which files will need to be created, modified, or deleted. Use git ls-files, find, and your knowledge of the codebase to list them. Count them.
Identify domains — classify each file into a domain:
Estimate complexity — score 1-5 based on:
Based on the analysis, recommend ONE team mode:
Conditions: 1-2 files, single domain, complexity 1-2.
Pick the best-fit specialist from:
Print:
[BW] plan: SOLO
scope: {N} files, {domain}
complexity: {score}/5
agent: {specialist}
tasks:
1. {what will be done}
2. {what will be done}
confirm? (y/n)
Conditions: 3-5 files, 2 domains, complexity 2-3.
Pick 2 complementary specialists. Common combos:
Print:
[BW] plan: PAIR
scope: {N} files across {domains}
complexity: {score}/5
agents: {specialist1} + {specialist2}
tasks:
1. [{agent}] {what will be done}
2. [{agent}] {what will be done}
confirm? (y/n)
Conditions: 5-10 files, 3+ domains, complexity 3-4, or any task benefiting from quality gates.
Pick the best-fit trio from all 49 specialists:
Print:
[BW] plan: TRIO
scope: {N} files across {domains}
complexity: {score}/5
agents:
tech: {tech-specialist} — implements
pm: {pm-specialist} — verifies requirements
critic: {critic-specialist} — reviews quality
execution:
1. [{tech}] {implementation task}
2. [{pm}] verify: {what PM checks}
3. [{critic}] review: {what critic checks}
(max 3 feedback rounds if rejected)
confirm? (y/n)
Conditions: 10+ files, 4+ domains, complexity 4-5, feature work.
Pick a squad preset or custom composition:
Print:
[BW] plan: SQUAD
scope: {N} files across {domains}
complexity: {score}/5
squad: {preset or "custom"}
- {specialist1} ({role})
- {specialist2} ({role})
- {specialist3} ({role})
- {specialist4} ({role})
execution (parallel):
1. [{agent1}] {task}
2. [{agent2}] {task}
3. [{agent3}] {task}
4. [{agent4}] {task}
resolution: majority vote on disagreements
confirm? (y/n)
Conditions: 10+ files, architecture changes, complexity 5, refactoring, or security-critical.
Pick a hierarchy preset:
Print:
[BW] plan: HIERARCHY
scope: {N} files, architecture-level
complexity: {score}/5
chain: {preset}
4. {C-level} — final approval
3. {Lead} — architecture review
2. {Senior} — implementation review
1. {Junior} — implementation
execution (bottom-up, review top-down):
1. [{junior}] implement: {task}
2. [{senior}] review + improve
3. [{lead}] architecture check
4. [{c-level}] final decision
confirm? (y/n)
Wait for user input. Accept:
[BW] plan cancelled.Before execution, ALWAYS save the plan to disk:
mkdir -p .bestwork/plans
Write the plan as JSON to .bestwork/plans/<timestamp>-<slug>.json:
{
"created": "<ISO timestamp>",
"task": "<user's original request>",
"mode": "solo|pair|trio|squad|hierarchy",
"scope": { "files": ["..."], "domains": ["..."], "complexity": N },
"agents": [{ "name": "tech-backend", "role": "tech", "tasks": ["..."] }],
"status": "pending|executing|complete|cancelled"
}
Also write a human-readable summary to .bestwork/plans/<timestamp>-<slug>.md.
This enables:
./plan --last — re-load and resume the most recent plan./plan --list — list all saved plansWhen the user says ./plan --last:
.json file from .bestwork/plans/Once confirmed, update the saved plan's status to "executing". Then route to the appropriate team mode. Print:
[BW] executing plan...
Then execute based on the confirmed mode:
SOLO: Load the specialist's prompt from prompts/tech/{name}.md (or the matching category). Execute the task directly using that agent's expertise and perspective.
PAIR: Execute sequentially — first specialist handles their domain files, second specialist handles theirs. Cross-review at the end.
TRIO: Execute as ./trio — spawn Tech agent (implement), PM agent (verify), Critic agent (review). Feedback loop up to 3 rounds. Record to meeting log at .bestwork/state/meeting.jsonl.
SQUAD: Spawn all agents in parallel using run_in_background. Each agent works on their assigned files. Disagreements resolved by majority vote. Record to meeting log.
HIERARCHY: Execute bottom-up (Junior implements first), review top-down (C-level approves last). Each level can send work back down with feedback. Record to meeting log.
For trio/squad/hierarchy, write the meeting log:
mkdir -p ~/.bestwork/state && echo '{"type":"header","teamName":"plan","mode":"<MODE>","task":"<TASK>","classification":"PLAN","developerCount":<N>,"routingReason":"plan skill recommendation"}' > .bestwork/state/meeting.jsonl
After execution completes, update the saved plan's status to "complete" and print:
[BW] plan complete. mode={MODE}, agents={N}, files={N}.
saved: .bestwork/plans/<filename>.md