Skip to content
cccskills
FreePublished on GitHub

parallel-investigate

Fans out multiple agents in parallel via the Workflow tool to investigate a codebase question — bug root cause tracing, cross-file or cross-repository behavior tracing, "why does X happen for some cases but not others" questions — then synthesizes a ranked, file:line-cited answer in chat. The point of this skill is explicitness: it announces the investigation dimensions and agent count before launching, narrates progress, and reports a structured synthesis, rather than leaving multi-agent orchestration as an implicit background behavior the user can't see into. Supports investigations that span more than one local repository on disk (e.g. a backend service and a client app that talks to it). Investigation only — does not edit code or fix anything; pair with another skill or a follow-up request for that. Use when the user wants to investigate a bug, trace behavior across files or repositories, understand why something happens under some conditions but not others, or explicitly asks for a parallel/multi-agent investigation. Trigger phrases: "parallel-investigate", "/parallel-investigate", "並列で調査して", "並列調査", "複数エージェントで調べて", "複数エージェントで並列調査して", "investigate this in parallel", "fan out and investigate", "investigate across repos", "trace this across repositories".

The English summary is being prepared. Showing the original description.

hideki5123/agent-skill-set

Run score Not tested

See how to install

Files included (5)

  • SKILL.md7.9 KB
  • references/decomposition-patterns.md4.0 KB
  • references/scenarios.feature3.5 KB
  • references/synthesis-format.md2.1 KB
  • references/workflow-template.md2.7 KB

SKILL.md (original)

Review the instructions the agent will receive before you install.

Parallel Investigate

Investigate a codebase question by fanning out multiple Workflow agents across independent dimensions, then synthesizing their findings into one ranked, cited answer. Always run this via the Workflow tool directly — never simulate the fan-out with sequential Agent calls, and never hide the orchestration from the user. Explicitness is the entire value of this skill over ad hoc investigation.

Feedback Check

If feedback/log.md exists next to this SKILL.md and has 5 or more entries, read the last 10. If a pattern is apparent (the same issue in 3+ entries, or average rating below 3), tell the user (in Japanese): 「過去のフィードバックで類似パターン を検出: [簡潔に]。/skill-improve --skill parallel-investigate で改善案を分析でき ます。」Continue with normal execution either way. If feedback/log.md does not exist, skip silently.

Step 1: Scope the Question

  1. Guard against overkill first. If the question can be answered by reading one file or a single targeted grep, say so briefly and just answer it directly — do not launch a Workflow for it. This applies even when the skill was triggered by natural language, not just the slash command.
  2. Guard against vagueness. If the request has no concrete subject or codebase area ("調べて" with nothing else, no symptom, no file/module named), ask 1-2 clarifying questions before proceeding. Do not guess dimensions from a request with nothing to decompose.
  3. Decompose into investigation dimensions. For genuinely multi-faceted questions, read references/decomposition-patterns.md for the by-layer, by-repository, and by-hypothesis decomposition strategies (with a worked example) before picking dimensions. Aim for 3-5 dimensions.
  4. Check for cross-repository scope. If the codebase references an external component whose source is not present in the current repo, do not conclude it's simply unavailable — follow the repo-discovery steps in decomposition-patterns.md (check common sibling-repo roots, then ask the user) before ruling it out.

Step 2: Announce the Plan

Before launching anything, tell the user in one short paragraph: how many dimensions/agents, what each will investigate, and that this runs in the background. Never launch silently — the whole point of this skill is that the user can see the investigation happening, not just receive a final answer.

Step 3: Launch the Workflow

Read references/workflow-template.md for the exact script pattern (Explore phase with one parallel agent per dimension, then a Synthesize phase). Read references/synthesis-format.md for the exact instructions to give the Synthesize agent (ranking, CONFIRMED/PLAUSIBLE confidence tags, contradiction resolution, convergence notes, a concrete next step, ~900-word cap).

Rules that apply regardless of the specific question:

  • Every Explore-phase agent prompt must be fully self-contained — agents share no context with you or each other. Restate the symptom/question, name the exact files/dirs (and repo path, if not the current one), and specify exactly what to report back with file:line citations.
  • Label every agent (explore:<dimension>, synthesis) so progress in /workflows is legible.
  • Use parallel() as the barrier between Explore and Synthesize — synthesis genuinely needs every finding at once to cross-reference and rank them. Do not use pipeline() here even though it's the general default for multi-stage Workflow scripts.
  • Before reading any repository you didn't already have open (see Step 1's cross-repo case), check git status first. If it's on a non-default branch with uncommitted changes, investigate the working tree read-only; do not switch branches, stash, or pull without asking.

Step 4: Report Back

Relay the Synthesize agent's output to the user in chat, organized and cited. This skill's default output is chat-only — do not write a report file unless the user separately asks for one to be saved.

Behavior Scenarios

BDD spec lives in references/scenarios.feature. Read only when auditing or amending this skill; not needed for normal execution.

Retrospective

After Step 4 completes, reflect on the investigation:

  1. Consider: was the initial decomposition wrong and had to be redone? Did a dimension turn out irrelevant or a needed one get missed? Any Workflow errors, or an agent that returned nothing useful?

  2. Ask the user (in Japanese): 「今回の調査のフィードバック (1-5の評価、気になった 点、または何もなければEnter)」 If the user provides a rating < 5, ALWAYS follow up with: 「なぜその評価ですか？ (改善のために具体的に教えてください)」 Record the response verbatim as Rating reason.

  3. If the user provides feedback OR corrections/issues actually occurred: a. Create feedback/ next to this SKILL.md if it does not exist. b. Read feedback/log.md (create with # Feedback Log header followed by a blank line and the comment <!-- Append new entries at the top. Do not edit previous entries. --> if it does not exist). c. Prepend a new entry directly after the header:

    ## <ISO-8601 timestamp>
- **Skill Version**: 1.0.0
- **Task**: <the question investigated, 1 line>
- **Outcome**: success | partial-success | failure | error
- **Rating**: <N>/5 (or "—" if not provided)
- **Rating reason**: <verbatim WHY answer, or "—" if rating was 5 or not provided>
- **Corrections**: <decomposition redone, dimension added/removed, or "none">
- **Issues**: <specific problems, or "none">
- **User Note**: <user's verbatim feedback, or "—">
---

    d. Confirm in one short Japanese sentence.

  4. If the user skips AND no corrections or issues occurred, end without recording.

References

  • references/decomposition-patterns.md — WHEN TO READ: during Step 1, when deciding how to split the question into investigation dimensions (by layer, by repository, by hypothesis) and how to locate a referenced-but-missing repo.
  • references/workflow-template.md — WHEN TO READ: during Step 3, as the starting script pattern to adapt for the real investigation.
  • references/synthesis-format.md — WHEN TO READ: during Step 3, when writing the Synthesize-phase agent's prompt.
  • references/scenarios.feature — WHEN TO READ: only when auditing or amending this skill.

Reviews

No reviews yet. Tell others what you thought.

Comments

No login needed. Share how it worked for you, ask a question, or note where it did not run.

No comments yet. Be the first to write one.

Comments are published right away. Do not include personal information such as email addresses or phone numbers. Up to 2 URLs.

0 / 1,000 characters

This form is protected by reCAPTCHA, and the Google Privacy Policy and Terms of Service apply.

More skills from this repository

adversarial-panel

Free

Run a multi-model adversarial mutual review with facilitator synthesis: 2+ panelists (different model families where possible — use codex-server / codex-cli for the external-family seat) independently answer, adversarially critique each other across rounds, and the main session acts as facilitator, synthesizing agreements, live disagreements, and calibrated confidence into one answer. Use ONLY on high-stakes, hard-to-reverse claims, designs or decisions, and only when the user EXPLICITLY asks to debate or adversarially review with multiple models — "discuss with Opus and Codex", "red-team this", "devil's advocate", 敵対的レビュー, 相互レビュー, 多モデル討論, 議論させて. A full run costs ~6 subagent/CLI dispatches: state the panel size and round count and get a go-ahead before launching. Where the Workflow tool is available, run the panel as a Workflow (judge-panel / adversarial-verify) instead of ad-hoc fan-out. Do NOT self-trigger on a bare "are you sure?" — that is a cue to OFFER the panel, not to run it.

The English summary is being prepared. Showing the original description.

hideki5123/agent-skill-set2Updated Oct 21, 2026

agent-swarm

Free

Orchestrate multiple coding-agent sessions (Claude Code, Codex, Gemini) over the mcp_agent_mail MCP server, observe their progress on a recurring auto-loop, and prevent inconsistencies (file/edit conflicts and contradictory decisions) between concurrent tasks. Auto-manages the agent-mail server, derives the shared room key from the git repo, and runs either as the orchestrator/observer or as a participating worker. Handles the operational details (server startup, room key, reservation conventions) so the user does not have to remember them. Use when coordinating several agent sessions on different tasks at once, watching a swarm of sessions for conflicts, or wiring already-running sessions into a shared coordination room. Trigger phrases: "agent-swarm", "/agent-swarm", "orchestrate agents", "orchestrate multiple claude sessions", "coordinate sessions", "watch for conflicts between agents", "observe the swarm", "prevent inconsistency between sessions", "multi-agent observe", "複数セッションを協調", "エージェント協調", "セッション間の不整合監視", "スウォーム監視", "複数のClaudeを束ねて".

The English summary is being prepared. Showing the original description.

hideki5123/agent-skill-set2Updated Oct 21, 2026

agents-init

Free

Generate AGENTS.md (cross-tool agent context — Codex CLI, Aider, Cursor, etc.) and wire CLAUDE.md as a symlink so Claude Code reads the same file. Delegates content generation to Claude's built-in /init when needed, then renames + symlinks. Idempotent. Triggers — /agents-init, "generate agents.md", "create agents.md", "share claude and codex context", "set up agents file", "init agents", "agents-init", "cross-tool agent context", "wire claude and agents".

The English summary is being prepared. Showing the original description.

hideki5123/agent-skill-set2Updated Oct 21, 2026

chrome-use

Free

Drive an already-running, LOGGED-IN Chromium browser on macOS via AppleScript — navigate, click, fill forms, scroll, and extract page content from the user's REAL session, with the browser left open and NO profile copy and NO restart. Use when the task needs the existing logged-in session (cookies, auth, current tabs) reused as-is. Trigger phrases: "chrome-use", "ブラウザから取得", "ログイン済みの Chromeで", "ブラウザを操作して", "このページ読んで", "Xのスレッド読んで", "ログインしたまま取得", "自分のセッションでスクレイプ", "read this page in my browser", "scrape with my session", "drive my chrome", "execute JS in my (logged-in) browser", "automate my logged-in browser". Chromium family (Chrome default; Brave/Edge/Arc/Vivaldi via --app). NOT for isolated/throwaway automation where login reuse is unneeded — use playwright-cli for that. macOS only.

The English summary is being prepared. Showing the original description.

hideki5123/agent-skill-set2Updated Oct 21, 2026

codex-cli

Free

Run OpenAI Codex CLI from the terminal for batch / one-shot `codex exec` invocations with output captured to an `-o` file. Use for fire-and-forget runs, scripted CI, code review batches (`codex review`), and resuming a specific session by id (`codex exec resume`). Always use this skill instead of `mcp__codex__codex` and `mcp__codex__codex-reply` MCP tools — CLI is faster and more capable. For streaming UX, multi-turn dialogues that need live state, structured (JSON schema) output, attaching images, or programmatic event handling → **use codex-server instead**. codex-server uses the user's ChatGPT subscription via Codex App Server and decoupled async invocation that bypasses the Bash 2-minute timeout entirely. Trigger patterns (batch-only — chat/streaming triggers belong to codex-server): codex-cli / codex CLI / codex exec / codex review / codex resume / codex fork / batch + {codex, gpt} / one-shot + {codex, gpt} / non-interactive + codex / codex {to a file, -o, output file, captured output} / CI + {codex, gpt} / scripted + {codex, gpt} / fork codex session / resume codex session by id

The English summary is being prepared. Showing the original description.

hideki5123/agent-skill-set2Updated Oct 21, 2026

codex-server

Free

Run OpenAI Codex via the Codex App Server (JSON-RPC-over-stdio) for streaming multi-turn chat sessions with persistent threads, structured output, image input, and rich event streams. Default entry point for ChatGPT/GPT/Codex-like conversations from the terminal. Uses the user's ChatGPT subscription (Plus / Pro / Team) exclusively via `codex login` — never consumes OPENAI_API_KEY billing, by design. Implementation: TypeScript on deno with npm:@openai/codex-sdk; spawns the existing system codex binary. Decoupled async invocation: every turn returns turn-id in <1 s; the actual SDK work runs in a detached worker that streams to per-turn files under ~/.codex-server/turns/<turn-id>/ — Bash's 2-min timeout never applies. Prefer this over codex-cli for streaming UX, multi-turn dialogues that need live state, structured (JSON schema) output, attaching images, or programmatic event handling. Use codex-cli only for one-shot batch invocations that pre-capture to an -o file. For OpenAI Platform API calls specifically (embeddings, image generation, audio, moderation, batches, files) — not open-ended chat — use openai-cli instead; that path bills OPENAI_API_KEY, this one uses the ChatGPT subscription. Trigger patterns (match any variation): codex / codex-server / codex app server / codex app-server / openai / open ai / OpenAI / chatgpt / chat gpt / ChatGPT / GPT / ask gpt / chat with gpt / gpt session / codex session / streaming chat / multi-turn chat / multi turn chat / second opinion / cross-check / cross check / ai pair programming / another AI / external AI / ask another model / get GPT's take / what does openai think / show me what gpt thinks / structured output from gpt / json schema gpt / resume codex thread / continue codex conversation / continue codex session / /codex-server

The English summary is being prepared. Showing the original description.

hideki5123/agent-skill-set2Updated Oct 21, 2026

See all skills by hideki5123

Report a problem with this skill