Przejdź do treści
Biblioteka skilli
Nauka i badania · Research

deep-research

/deep-research

Prowadzi otwarte badania z wielu źródeł, uzgadnia z użytkownikiem zakres i plan, równolegle zleca wątki subagentom i składa raport z cytatami.

Repozytorium
tejit-plugin ↗
Licencja
MIT
Ostatnia zmiana repozytorium
10.11.2026

Ocena i bezpieczeństwo

100/ 100Klasa: dobry
Format25 / 25
Frontmatter z nazwą i opisem, poprawna nazwa, opis do 1024 znaków, nazwa zgodna z folderem.
Odkrywalność25 / 25
Opis mówi, kiedy użyć skilla, ma sensowną długość i jest w 3. osobie.
Treść35 / 35
Co najmniej 150 słów i do 500 linii, nagłówki, listy kroków, przykłady albo kod.
Higiena15 / 15
Bez zaślepek typu TODO, bez ścieżek Windows, z licencją.
Ryzykobrak sygnałów w skanie
Pliki w paczce
4
Skrypty
0
Komentarze HTML
0
Adresy URL w SKILL.md
0
Przestarzałość

Brak sygnałów. Repozytorium zmienione 10.11.2026, bez odwołań do wycofanych modeli i API.

Licencja

MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.

Co robi i kiedy

Prowadzi otwarte badania z wielu źródeł, uzgadnia z użytkownikiem zakres i plan, równolegle zleca wątki subagentom i składa raport z cytatami.

Oryginalny opis (EN)

Use for open-ended, multi-source research on any topic — web research, idea/market validation, or deep code/repo investigation. Clarifies scope, proposes a research plan for user confirmation, then fans out parallel sub-agents (max 2 concurrent, depth ≤2) to investigate each direction and returns one synthesized, cited report. Not for a single quick lookup — use web search, page fetch or a read-only exploration sub-agent directly for that.

Jak działa

  1. Overview
  2. Phase 0-1 — Scope & Plan (inline, interactive)
  3. Phase 2 — Execute (delegated)
  4. Phase 3 — Synthesize (inline, by you)
  5. Common Mistakes
Sekcje z nagłówków SKILL.md, w oryginale · 83 linii, 710 słów
Podgląd SKILL.md
---
name: deep-research
description: Use for open-ended, multi-source research on any topic — web research, idea/market validation, or deep code/repo investigation. Clarifies scope, proposes a research plan for user confirmation, then fans out parallel sub-agents (max 2 concurrent, depth ≤2) to investigate each direction and returns one synthesized, cited report. Not for a single quick lookup — use web search, page fetch or a read-only exploration sub-agent directly for that.
metadata:
  version: "0.1.0"
---

# deep-research — Multi-Agent Research Orchestrator

Before anything else, follow `../_shared/update-check.md`.

## Overview

Two-phase flow: (1) scope the question and get the user to confirm a plan, (2) execute
via parallel sub-agents and return one synthesized report. Domain-agnostic — the
same flow covers internet research, idea/market validation, and deep-dives into a
codebase or repo. Nothing here is specific to one domain; don't hardcode examples into
prompts when using this skill.

Core principles:

- **Confirm before you fan out.** Phase 0/1 happen inline, in this conversation, because
  they need the user. Phase 2 onward is fully delegated.
- **No narration.** Once the plan is confirmed, do not post progress narration or partial
  findings. If the user asks "status", answer in at most five lines, minimal words: which
  directions are done, running, or queued. Otherwise the next thing you say is the
  finished report.
- **Depth ≤ 2.** You (this conversation) are depth 0. You may launch sub-agents (depth 1).
  Each depth-1 agent may launch at most one further sub-agent (depth 2) if its assigned
  direction genuinely benefits from splitting. Depth-2 agents may not spawn further
  agents. This is a hard limit, not a suggestion — state it explicitly in every sub-agent
  prompt so it survives the handoff.
- **A mid-tier model (medium reasoning effort) for workers, your strongest model for the plan.** Spawn every research
  sub-agent on a mid-tier model, medium reasoning effort. Your own
  scoping/decomposition/synthesis is the highest-leverage reasoning in this workflow — if
  you're not already running your strongest available model with high effort, mention once at the start of Phase
  0 that switching models will improve plan quality, then proceed regardless of the
  answer.
- **Cite everything non-trivial.** Every claim in the final report traces to a source
  (URL, `file:line`, PR#, commit, etc.). A source's silence on a point is reported as
  silence, not filled in from assumption.

## Phase 0-1 — Scope & Plan (inline, interactive)

Read `clarify-and-plan.md` and follow it. Produces:

1. Zero or one round of clarifying questions (skip entirely if the request is already
   unambiguous — don't ask just to be thorough).
2. A research plan: a short list of independent research directions, each of which will
   become one depth-1 sub-agent.
3. Explicit user confirmation of the plan before any agent is launched.

Do not proceed past this phase without an explicit go-ahead.

## Phase 2 — Execute (delegated)

1. Launch one depth-1 sub-agent per confirmed research direction, using
   `research-agent-prompt.md`, filled in per direction. Run in waves of at most 2
   concurrent (in the background); when one completes, launch the next queued direction,
   keeping ≤2 in flight at all times.
2. Each depth-1 agent may spawn exactly one depth-2 helper for a genuinely parallel
   sub-split of its own direction — the prompt template tells it this is the ceiling.
   Most directions won't need one; don't force it.
3. Wait for every direction to complete. Do not message the user in between except to
   answer "status", and do not preview, summarize, or speculate about results while waiting.
4. If a direction fails or returns only partial results, note it as a gap in the final
   report rather than retrying more than once.

## Phase 3 — Synthesize (inline, by you)

1. Read every depth-1 agent's returned findings in full.
2. Merge into one report using `report-template.md`: preserve contradictions between
   directions instead of smoothing them over, name gaps explicitly, keep every citation
   intact through the merge.
3. Send the report as your first and only message to the user since plan confirmation.

## Common Mistakes

- Skipping Phase 0/1 confirmation and diving straight into agents — the plan is the
  contract; don't fan out on a guess at scope.
- Narrating progress unprompted ("agent 1 of 3 done...") between plan confirmation and the final
  report — defeats the entire point of delegating.
- Launching more than 2 depth-1 agents concurrently, or letting a depth-1 agent spawn
  more than one depth-2 helper.
- Standing up this whole flow for a single quick fact — if one web search or page fetch, or
  one read-only exploration sub-agent, would answer it, just do that instead.
- Losing citations during synthesis — a claim that loses its source in the merge is worse
  than a claim that was never included.
Pliki w paczce (4)
  • SKILL.md 5 KB
  • clarify-and-plan.md 3 KB
  • report-template.md 2 KB
  • research-agent-prompt.md 3 KB

Wersje

RepozytoriumWynikLicencja
Ta strona · tejitpabari99/tejit-plugin100MIT
shepherdjerred/monorepo100GPL-3.0
3x-haust/ondra100MIT
glebis/claude-skills100MIT
drvoss/everything-copilot-cli100MIT
5dive-ai/skills100MIT
huayu-pro/skills100MIT
lovstudio/skills100MIT
bytedance/deer-flow100MIT

Migawka bazy z 13.11.2026 · wersja paczki a6a4509ed26b (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

  • 100research-quick
    Prowadzi otwarty research jednego pytania: dzieli je na podpytania, sprawdza równolegle i zapisuje raport z cytowaniami.
  • 100research
    Bada temat z wielu źródeł, weryfikuje twierdzenia w niezależnych publikacjach i przedstawia wyniki z datowanymi cytatami, w trybie szybkim lub pogłębionym.
  • 100deep-web-research
    Prowadzi wielożródłowy research w sieci z porównaniem, kontrolą jakości źródeł i raportem z cytatami, przy przeglądach rynku, decyzjach technicznych i due diligence.
  • 100firecrawl-deep-research
    Prowadzi głęboki research z wieloma źródłami i zwraca odpowiedź, raport lub tabelę z cytatami, np. do analiz rynku.
  • 100research
    Uruchamia zapytanie deep research w agencie Google Gemini i zwraca raport z cytowaniami, z trybem max o większym nakładzie pracy.
  • 100deep-research-discipline
    Pilnuje, żeby badanie opierało się na dowodach: cytowane źródła, trzy poziomy pewności, weryfikację krzyżową i niezależną recenzję.

Uwagi i zgłoszenia

Wiadomość trafi do Bartosza Sokolińskiego. Adres e-mail jest potrzebny tylko do odpowiedzi. Szczegóły w polityce prywatności.