AGENT SKILL

ywc-team-assemble

(ywc) Use when the user explicitly asks Codex to assemble a specialist team, delegate work to subagents, run a swarm, or split a complex task across parallel expert roles. Triggers: "assemble a team", "use agents", "delegate", "parallel agents", "swarm", "specialist team", "팀 구성", "전문가 팀", "エージェントチーム", "専門家チーム". Do not use for simple questions, single-file edits, strictly sequential work, or any task where the user has not explicitly authorized subagents/parallel/delegated execution.

License 许可
MIT
最近更新
2026-11-08
Stars
★ 8
Forks
1

01 前言 Front Matter

name
ywc-team-assemble
description
(ywc) Use when the user explicitly asks Codex to assemble a specialist team, delegate work to subagents, run a swarm, or split a complex task across parallel expert roles. Triggers: "assemble a team", "use agents", "delegate", "parallel agents", "swarm", "specialist team", "팀 구성", "전문가 팀", "エージェントチーム", "専門家チーム". Do not use for simple questions, single-file edits, strictly sequential work, or any task where the user has not explicitly authorized subagents/parallel/delegated execution.

02 资源文件 9

03 SKILL.md 正文

Team Assemble

Announce at start: "I'm using the ywc-team-assemble skill to split explicitly authorized work across specialist roles."

Use this skill to decompose complex work into specialist roles, run independent roles in parallel with Codex subagents, and synthesize the results into one coherent outcome.

Rationalization Defense

When tempted to skip a step, check this table first:

Excuse Reality
"This is complex, so subagents are automatically allowed" Complexity is not authorization. Use subagents only when the user explicitly asked for team, parallel, delegated, or agent work.
"The hardest task can go to a worker while I wait" Keep the immediate critical-path blocker local. Delegate sidecar work that can run in parallel.
"Two workers can edit nearby files if they coordinate" Assign disjoint write scopes. Shared ownership causes merge conflicts and lost work.
"A broad prompt gives specialists flexibility" Broad prompts waste context and produce overlapping work. Give concrete scope, exclusions, ownership, and output format.
"Subagent results are accepted by default" Review every returned output before integration. Subagents are contributors, not source of truth.
"One slow agent should be waited on repeatedly" Wait only when its result blocks the next critical-path decision. Continue non-overlapping local work otherwise.

Violating the letter of these rules is violating the spirit. Team execution is useful only when boundaries are explicit.

Eligibility Gate

Proceed only when all conditions are true:

  1. The user explicitly requests a team, swarm, parallel execution, delegated agents, or specialist roles.
  2. The work has at least two independent or partially independent workstreams.
  3. The subagent outputs can be integrated without shared write conflicts.

If the task is a generated task wave from ywc-task-generator, prefer ywc-parallel-executor. If the task is a normal implementation with one clear critical path, work locally or use ywc-sequential-executor.

Workflow

  1. Analyze the task and identify role boundaries.
  2. Decide what must stay local on the critical path.
  3. Assign only sidecar or independent workstreams to subagents.
  4. Give each subagent a concrete scope, expected output, and write boundary.
  5. Continue useful local work while subagents run.
  6. Review returned outputs, integrate only relevant changes, and report the final result.

Claim/Evidence and role isolation

Every team prompt is assembled from a validated, bounded context packet. The packet contains only Included scope, Excluded scope, Artifacts (paths), and optional Claims. Validate the complete packet recursively before making any role projection; a rejected packet is terminal and must not be redacted, summarized, retried, or forwarded.

Claims follow the canonical Claim/Evidence contract in the shared subagent-status-actions.md reference owned by ywc-sequential-executor:

  • Emit at most three Claims. Each Claim has exactly Statement and Evidence.
  • Statement is factual and at most 1,024 characters.
  • Evidence is a project-relative file:line citation with a positive line number or a project-relative artifact path; absolute paths, .., URIs, and unlabelled prose are invalid.
  • Reject missing, uncited, over-cap, malformed, or private fields as BLOCKED and name only the field and bounded rule in the diagnostic.

Role projections are strict allowlists:

  • An independent reviewer receives only included scope, excluded scope, and artifact paths. It receives no Claims, peer conclusions, peer recommendations, transcript, or raw content.
  • A dependent role receives only Claims and the artifact paths cited by those Claims. It receives no uncited artifacts, peer transcript, peer conclusions, peer recommendations, or raw content. Missing cited context returns NEEDS_CONTEXT with the missing path/citation.

Privacy validation rejects forbidden raw-content keys recursively, including case/separator variants of transcript, chain_of_thought, generated_source, full_diff, raw_tool_output, raw_response, and tool_output. This check also covers nested Claim/Evidence data, artifact metadata, concerns, blockers, and wrapper objects before projection.

Role Design

Use a small team. Two to four roles is usually enough.

Role type Use for Typical mode
Researcher external or repo-aware investigation that can run independently read-only report
Explorer bounded codebase questions with no file edits read-only answer
Worker implementation in a disjoint file or module scope direct file edits
Reviewer focused verification of a completed or near-completed change findings report

Keep role ownership explicit:

  • Files or modules the role may edit
  • Files or modules the role must not edit
  • Commands the role should run, when known
  • Output format expected from the role

Subagent Rules

Use spawn_agent only for tasks that are independent enough to run in parallel with local work.

Do not delegate:

  • The immediate blocker for the next local step
  • Broad ownership of the whole repository
  • Destructive git, deploy, database, or production operations
  • Work that requires secrets or live credentials unless the user has explicitly approved that path

When delegating implementation, tell each worker:

  • It is not alone in the codebase.
  • It must not revert changes made by others.
  • It must adapt to concurrent changes.
  • It must list the files it changed in its final response.

Prompt Template

Use references/prompt-templates.md when preparing prompts for multiple subagents or when the task has dependencies between roles.

Each subagent prompt should include:

  1. Context
  2. Goal
  3. Scope and exclusions
  4. Ownership boundary
  5. Output format
  6. Whether file edits are allowed

For reviewer and dependent packets, the prompt must also name the projection (independent or dependent) and list the exact allowed fields. Never pass a source payload and ask a role to ignore disallowed fields; construct the allowlisted projection first.

Execution Pattern

Start independent subagents in the same turn when possible. After spawning, immediately continue local work that does not overlap with delegated scopes.

Wait for subagents only when their results are needed for the next critical-path decision. Review their outputs before treating them as accepted work.

For dependent roles, pass only the necessary prior result into the next subagent. Avoid forwarding full transcripts or unrelated context.

Output Format

Return a concise synthesis:

  • Team roles used
  • Local work completed
  • Subagent outputs accepted or rejected
  • Files changed
  • Verification run
  • Remaining risks or follow-up decisions
Status: <DONE | DONE_WITH_CONCERNS | BLOCKED | NEEDS_CONTEXT>
Team: <roles used and why>
Accepted outputs: <subagent outputs accepted or rejected>
Local work: <what the main agent completed>
Verification: <commands or checks run>
Next action: <follow-up decision or "none">

Validation

Before finalizing, verify that each subagent had a bounded prompt, ownership boundaries did not overlap unsafely, accepted outputs were reviewed before use, files changed are listed, and any verification gaps or rejected outputs are reported.

04 同仓库更多 Skills

yongwoon
yongwoon/ywc-agent-toolkit

missing-frontmatter

暂无简介

★ 8
yongwoon
yongwoon/ywc-agent-toolkit

malformed-frontmatter

暂无简介

★ 8
yongwoon
yongwoon/ywc-agent-toolkit

fixture-501-fail

(ywc) Use when aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa Do not use for x.

★ 8
yongwoon
yongwoon/ywc-agent-toolkit

fixture-500-pass

(ywc) Use when aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa Do not use for x.

★ 8
yongwoon
yongwoon/ywc-agent-toolkit

ywc-worktrees

(ywc) Use when creating, auditing, pruning, or resolving git worktrees for isolated task workspaces, whether delegated from an executor or used standalone. Triggers: "worktree 생성", "worktree audit", "worktree 정리", "worktree prune", "ywc-worktrees", "create worktree". Do not use for branch creation/deletion alone, PR creation (use ywc-create-pr), CI or merge delivery (use ywc-finish-branch), task orchestration (use ywc-sequential-executor or ywc-parallel-executor), or non-git isolation.

★ 8
yongwoon
yongwoon/ywc-agent-toolkit

ywc-wayfinder

(ywc) Use when a large or uncertain change needs multi-session discovery with a deterministic local map and exactly one active ticket before planning or implementation. Triggers: "wayfinder", "길찾기", "discovery map", "multi-session discovery", "탐색 ticket", "経路整理", "探索マップ", "ywc-wayfinder". Do not use for ordinary planning (use ywc-plan), idea clarification (use ywc-brainstorm), direct implementation, or tracker/database writes outside the local Markdown map.

★ 8