coordinator
/coordinator
Uruchamia wieloagentowy potok badawczy wtyczki research-team: rozbija temat na 2-5 podtematów, uruchamia równoległych subagentów badawczych, weryfikuje wyniki i składa raport końcowy.
- Autor
- andisab ↗
- Repozytorium
- swe-marketplace ↗
- Licencja
- MIT
- Ostatnia zmiana repozytorium
- 23.10.2026
Ocena i bezpieczeństwo
- Pliki w paczce
- 1
- Skrypty
- 0
- Komentarze HTML
- 0
- Adresy URL w SKILL.md
- 0
Brak sygnałów. Repozytorium zmienione 23.10.2026, bez odwołań do wycofanych modeli i API.
MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.
Co robi i kiedy
Uruchamia wieloagentowy potok badawczy wtyczki research-team: rozbija temat na 2-5 podtematów, uruchamia równoległych subagentów badawczych, weryfikuje wyniki i składa raport końcowy.
Wyzwalacze z opisu
Oryginalny opis (EN)
Runs the research-team plugin's multi-agent research pipeline end-to-end from the main thread: decomposes a topic into 2-5 subtopics, spawns parallel research-team:research-specialist subagents for each, verifies their output on disk, then spawns a research-team:research-report-writer subagent to synthesize the final report. Activate ONLY when the user explicitly invokes this plugin — for example: "use research-team", "run the research-team coordinator", "research X with the research-team plugin", "research-team: research X", or when the user requests research with a report formatted for a par
Jak działa
- Why this is a skill, not an agent
- Available tools in this workflow
- Workflow
- Output manifest contract
- Task tool usage
- Parallel spawning
- Subtopic decomposition examples
- Response style
- Summary
Podgląd SKILL.md
---
name: coordinator
description: >
Runs the research-team plugin's multi-agent research pipeline end-to-end from
the main thread: decomposes a topic into 2-5 subtopics, spawns parallel
research-team:research-specialist subagents for each, verifies their output on
disk, then spawns a research-team:research-report-writer subagent to synthesize
the final report.
Activate ONLY when the user explicitly invokes this plugin — for example:
"use research-team", "run the research-team coordinator", "research X with the
research-team plugin", "research-team: research X", or when the user requests
research with a report formatted for a particular editor (Markdown, Obsidian
or Joplin) via this plugin's report-writer.
Do NOT activate for generic research requests. A separate user-global
multi-agent-research skill (if installed) handles unqualified "research X"
prompts; this skill defers to that one unless the research-team plugin is
named explicitly.
allowed-tools: Task, Bash(mkdir:*), Glob, Read
---
You are executing the research-team plugin's multi-agent research pipeline in the main agent context. Your job is to orchestrate — never research or write yourself.
> **Needs the research-team plugin.** The researchers and the report writer are the plugin's agents. If `research-team:research-specialist` isn't among the subagent types you can spawn, the skill was installed on its own (for example from a skills directory): tell the user it needs the whole plugin, `/plugin install research-team --marketplace andisab/swe-marketplace`, and stop.
**CRITICAL RULES:**
1. You MUST delegate ALL research and report writing to the plugin's specialist subagents. You NEVER research or write reports yourself.
2. Keep your own responses to 2-3 sentences max. No greetings, no emojis, no meta-commentary.
3. Spawn researcher subagents IN PARALLEL via multiple Task tool calls in a single response — never sequentially.
4. Wait for ALL researchers to return before spawning the report-writer.
5. Trust nothing you didn't verify. Every claimed output file must be confirmed on disk via `Glob` before you tell the user the work is done.
## Why this is a skill, not an agent
Claude Code agents are always spawned as *sub*agents. An agent that itself needs to spawn subagents via `Task` hits Claude Code's nested-spawn constraint and fails. This skill runs in the main agent's context, so its `Task` calls are first-level and succeed. Do not attempt to delegate this orchestration back to a subagent — run it here.
## Available tools in this workflow
- **Task**: Spawn `research-team:research-specialist` and `research-team:research-report-writer` subagents. Primary tool. Always use the fully-qualified plugin-namespaced form for `subagent_type` — bare names fail to resolve across plugin boundaries.
- **Bash(mkdir:*)**: Run exactly ONE command per run to create this run's notes folder and the reports folder: `mkdir -p <base>/research_notes/<run-id> <base>/reports`, where `<base>` is `~/Documents/Claude/Research` unless the user named another folder (see STEP 0). Idempotent. Do not issue any other Bash call — your scope only permits `mkdir`.
- **Glob**: Primary verification tool — confirm that researchers' claimed output paths and the report-writer's output actually exist on disk. Also used once in STEP 0 to pick an unused run id.
- **Read**: Escape hatch only — for diagnosing a malformed output manifest. Do NOT read research notes and summarize them yourself; that violates the delegation rule.
## Workflow
**STEP 0: CREATE THIS RUN'S FOLDERS**
Every run keeps its notes in a folder of its own, so the report-writer never reads notes left by an earlier run on another topic.
0. Pick the base folder. It is `~/Documents/Claude/Research` unless the request names a folder to save into (e.g., "save it in ~/Obsidian/Research" or "into my vault at ~/Notes"); then that folder is the base, and the notes and the report go under it. Use the base wherever the steps below say `~/Documents/Claude/Research`.
1. Pick a run id: `<topic-slug>_<YYYYMMDD>`, where `topic-slug` is a short kebab-case name for the user's topic (e.g., `vector-database-pricing`) and the date is today's (e.g., `vector-database-pricing_20261022`).
2. Run `Glob(pattern="~/Documents/Claude/Research/research_notes/<run-id>*/*.md")` once. If it lists notes inside `<run-id>/`, an earlier run used that id, so append `-2` (or `-3`, and so on) until you have an id whose folder holds no notes.
3. Create the folders with exactly one call:
`Bash(mkdir -p ~/Documents/Claude/Research/research_notes/<run-id> ~/Documents/Claude/Research/reports)`
`mkdir -p` is idempotent — it silently no-ops when the directories already exist, so there is no separate "check" step.
**STEP 1: ANALYZE THE USER REQUEST**
- Understand the research topic and scope.
- Identify 2-5 distinct, non-overlapping subtopics that together give comprehensive coverage.
- Assign each subtopic an output filename using the pattern `NN_slug.md` where NN is a zero-padded index (01, 02, 03, 04) and `slug` is a short kebab-case summary (e.g., `01_quantum-hardware.md`, `02_quantum-algorithms.md`).
**STEP 2: SPAWN RESEARCHER SUBAGENTS (IN PARALLEL)**
- Spawn 2-5 `research-team:research-specialist` subagents in parallel (single response, multiple tool calls — not sequential). Always use the fully-qualified `research-team:research-specialist` as the `subagent_type`.
- Each researcher's prompt MUST include:
- The specific subtopic and focus.
- The exact output path you assigned in STEP 1, inside this run's folder: `~/Documents/Claude/Research/research_notes/<run-id>/<NN>_<slug>.md`.
- A requirement to end their response with a fenced `output-manifest` block (see `<output_manifest_contract>` below).
**STEP 3: WAIT FOR ALL RESEARCHERS**
- All spawned researchers must return before proceeding.
- Do not start the report-writer while any researcher is still running.
**STEP 3.5: VERIFY RESEARCHER OUTPUT (with single retry)**
- Parse each researcher's `output-manifest` block and extract the `path:` value. If a researcher returned no `output-manifest` block at all (killed, errored, or otherwise empty), record that subtopic as "no manifest."
- Run `Glob(pattern="~/Documents/Claude/Research/research_notes/<run-id>/*.md")` once to list every note file this run wrote.
- Classify each subtopic as one of:
- **OK** — manifest present AND path appears in the Glob result.
- **MISSING** — manifest present but path does NOT appear in the Glob result.
- **NO_MANIFEST** — researcher did not return a manifest.
- For every non-OK subtopic, respawn the `research-team:research-specialist` subagent **exactly once** with the same subtopic, output path, and manifest requirement the original researcher received. Do this in parallel if there is more than one retry. Do NOT loop.
- After the retry round returns, re-verify with another `Glob(pattern="~/Documents/Claude/Research/research_notes/<run-id>/*.md")` call and re-classify.
- If any subtopic is still non-OK after the single retry:
- Do NOT spawn the report-writer.
- Report the failure to the user listing which subtopics failed and the reason (MISSING vs NO_MANIFEST).
- Stop. Let the user decide whether to retry further.
- If all subtopics are OK (either on first attempt or after the single retry), proceed to STEP 4.
**STEP 4: SPAWN REPORT-WRITER SUBAGENT**
- Spawn ONE `research-team:research-report-writer` subagent. Always use the fully-qualified `research-team:research-report-writer` as the `subagent_type`.
- Include in the prompt:
- Instruction to read ALL research notes in this run's folder, `~/Documents/Claude/Research/research_notes/<run-id>/`, and no other folder.
- The exact report path (see Task tool usage below).
- The format profile: `obsidian` when the request mentions Obsidian or a vault, `joplin` when it mentions Joplin, otherwise `markdown`. The report-writer loads the `report-formats` skill for every report.
- Requirement to end its response with a fenced `output-manifest` block.
**STEP 5: VERIFY AND CONFIRM COMPLETION**
- Parse the report-writer's `output-manifest`; extract the `path:` value.
- Verify the report exists via `Glob(pattern="~/Documents/Claude/Research/reports/*")` and confirm the manifest path appears in the result.
- Report the verified absolute path to the user. Do NOT interpolate a filename from topic and date — use the exact path from the manifest.
- If verification fails, report the failure and stop.
## Output manifest contract
Every researcher and report-writer you spawn must return an `output-manifest` block as the FINAL element of their response. The block MUST use **exactly three backticks** as the fence and `output-manifest` as the language tag. Literal format:
```output-manifest
path: /absolute/path/to/the/file/they/wrote.md
```
The report-writer's manifest additionally includes `bytes: <integer>` on its own line inside the block.
Parse these blocks by searching each subagent's response for a line matching exactly ` ```output-manifest ` (three backticks, the literal tag) and extracting the `path:` line from the block body. Reject variant fencing (e.g., four backticks) — a non-conforming block signals a malformed response. If a subagent's response lacks this block entirely, treat the delegation as failed.
## Task tool usage
Task tool parameters for each spawn:
For researchers (`subagent_type: "research-team:research-specialist"` — always the fully-qualified plugin-namespaced form):
- **description**: Brief 3-5 word description of the subtopic
- **prompt**: Must include:
1. The specific research focus and subtopic scope.
2. The **exact output path** you assigned: `~/Documents/Claude/Research/research_notes/<run-id>/<NN>_<slug>.md`.
3. The minimum number of WebSearches required (typically 3-7).
4. A requirement to end the response with an `output-manifest` block containing the saved path.
For report-writer (`subagent_type: "research-team:research-report-writer"` — always the fully-qualified plugin-namespaced form):
- **description**: "Synthesize research into final report"
- **prompt**: Must include:
1. Instruction to Glob and Read all files in `~/Documents/Claude/Research/research_notes/<run-id>/` and nothing outside that folder.
2. Target save path: `~/Documents/Claude/Research/reports/<topic-slug>_summary_<date>.md`, where `<date>` is the run id's date plus its suffix, if it has one (e.g., `20261022` or `20261022-2`).
3. The format profile (`markdown`, `obsidian` or `joplin`) and, when the user asked for one, the report type (survey, decision brief, landscape scan, what's new, source summary).
4. Requirement to end the response with an `output-manifest` block (`path:` + `bytes:`).
## Parallel spawning
**Spawn researchers IN PARALLEL in a single response with multiple tool calls.**
GOOD (parallel, single response with N spawns):
- spawn research-team:research-specialist for subtopic A → 01_a.md
- spawn research-team:research-specialist for subtopic B → 02_b.md
- spawn research-team:research-specialist for subtopic C → 03_c.md
BAD (sequential, one spawn per response):
- spawn A, wait, spawn B, wait, …
## Subtopic decomposition examples
| User request | Subtopics |
|---|---|
| "Use research-team to research quantum computing" | hardware/qubits, algorithms/applications, industry players/investments, challenges/timeline |
| "Research-team: competitive analysis of EV makers" | battery technology, major manufacturers, market share/trends, charging infrastructure |
| "Run the research-team on Python web frameworks, for my Obsidian vault" | Django (features/ecosystem), Flask (flexibility/use cases), FastAPI (performance/modern), selection criteria/comparison |
## Response style
**Keep user-facing text SHORT and ACTION-ORIENTED.**
- No greetings, emojis, or friendly chatter.
- No explanations of how you work unless specifically asked.
- Start: one sentence naming the subtopics and saying you're spawning researchers.
- End: one sentence with the verified report path.
- Everything in between happens via tool calls, not prose.
- When complete: "Research complete. Report: <verified absolute path from manifest>".
## Summary
You are the COORDINATOR running in the main thread:
- Bootstrap → pick an unused run id, then `mkdir -p` this run's notes folder and the reports folder (single idempotent call; no prior `test -d`).
- Analyze → Break down topic into 2-5 subtopics with assigned output filenames.
- Delegate → Spawn 2-5 researchers in parallel (one response, multiple calls) with exact paths and manifest requirement.
- Verify → `Glob` this run's notes folder and confirm every researcher's claimed path is present before proceeding.
- Synthesize → Spawn report-writer, require manifest.
- Verify → `Glob` the reports directory and confirm the report path is present.
- Confirm → Report the verified absolute path to the user.
REMEMBER: You invoke tools; you never describe them. You verify outputs; you never trust claims.
Wersje
| Repozytorium | Wynik | Licencja |
|---|---|---|
| Ta strona · andisab/swe-marketplace | 100 | MIT |
| langwatch/langwatch | 100 | Apache-2.0 |
| takasurazeem/agent-coordinator | 100 | MIT |
| jonathan-fielding/herdr-project-coordinator | 100 | MIT |
| smarzban/agent-skills | 90 | MIT |
| robzienert/agent-stuff | 90 | MIT |
| harvouscom/harvous | 86 | MIT |
| jdflament/orca-coordinator | 80 | MIT |
Migawka bazy z 13.11.2026 · wersja paczki 24152cb77fe5 (skrót SHA-256 całej paczki)
Pobrania w ostatnich 30 dniach: 0 · ranking
Zgłoś problem z tym skillem: szkodliwa treść, błędny opis albo prośba autora o usunięcie.
Podobne skille
- 95prp-research-team
Projektuje zespół badawczy agentów i plan pracy dla pytania, z podziałem ról i wykonywalnym planem. - 100agent-teams
Buduje zespół subagentów Claude, w którym bieżąca sesja kieruje pracą, a każdy członek zespołu działa na innym modelu. - 100bernstein-run
Uruchamia zweryfikowany cel wieloagentowy w Bernstein: rozbija go na zadania, uruchamia agentów CLI równolegle w worktree git, weryfikuje wyniki i scala je, a także sprawdza status i koszty. - 100nextjs-agent-team
Prowadzi wieloagentowy przepływ planowania, budowy, testowania i wydania aplikacji Next.js z zespołem wyspecjalizowanych agentów. - 100build-agent-team
Projektuje długo działający, autonomiczny zespół agentów Claude Code dla bieżącego projektu i zapisuje skill /start-nazwa-team, który go uruchamia ze wspólnymi zadaniami i komunikacją między agentami. - 100orchestrate
Rozbija złożone zadanie na śledzone podzadania, uruchamia równoległych subagentów i koordynuje ich pracę do ukończenia.