Run Research

Technical Research

One technical question. Web + Reddit evidence. Source-backed synthesis. Optional multi-agent fan-out when the question spans 3+ subdomains.

When to use

Trigger on phrasings like:

• "research X for me", "do some research on…", "find sources on…"
• "why does library Y do Z", "is bug B fixed in version V", "what changed between V1 and V2"
• "compare A vs B for production", "should we migrate from X to Y", "any gotchas with…"
• "what do practitioners say about…", "any horror stories with…", "reddit experience with…"
• "verify that vendor V still does W", "current pricing/quota/CVE for…"
• "give me the current state of "

Do NOT use when:

• The user asks to research a market, vendor category, or 5+ entities with per-entity packs, comparison templates, or a reusable corpus. Use run-deep-research.
• The user asks to find, shortlist, or compare GitHub repositories for a concrete need. Use run-github-scout.
• The question is answerable from the local codebase alone, or a docs page already in context is sufficient — just answer it.
• The user explicitly asked not to search the web.

The fence between this skill and the corpus skills: run-research answers one question and returns one synthesis (Markdown by default). Corpus skills answer N questions across N entities and return a multi-file evidence corpus. If the deliverable is a folder, you are in the wrong skill.

Research tool surface

Use the Research Powerpack MCP tools first. If unavailable or denied, fall back to built-in web tools; if those fail, use curl and parse manually.

| Capability | First choice | Fallback 1 | Fallback 2 | |---|---|---|---| | Research prelude | mcp__research-powerpack__get-research-consultancy | manual query plan | - | | Targeted search | mcp__research-powerpack__web-search | WebSearch | curl | | Page extraction | mcp__research-powerpack__scrape-link | WebFetch | curl + parse |

Short aliases used throughout this skill:

• get-research-consultancy -> mcp__research-powerpack__get-research-consultancy
• web-search -> mcp__research-powerpack__web-search
• scrape-link -> mcp__research-powerpack__scrape-link

There are only three tools: a planner, a keywords-only search, and an extraction-always scrape. web-search never classifies, tiers, or synthesizes — it always returns the same ranked URL pool, so there is no "prefer smart vs raw" decision to make. Scope is chosen by how you write keywords, not by a parameter: write plain probes for web evidence, and write explicit site:reddit.com/r/.../comments probes for Reddit permalink discovery. Mixing both intents in one keyword set wastes ranking budget — split into separate calls instead.

scrape-link always requires extract and always runs LLM extraction, including on Reddit permalinks — the Reddit API still fetches the full threaded post + comments first, then extraction runs on top. For sentiment or dissent work, write an extract that explicitly asks for verbatim quotes with author/score attribution (e.g. verbatim quotes with author + score | agreement reasons | dissent reasons | migration drivers) so the threading detail survives extraction.

get-research-consultancy is the planner for substantive sessions. It returns gaps_to_watch and stop_criteria — treat both as binding contracts.

For tool-by-tool API and operational thresholds, read references/tools.md. For prompting each tool well, read references/prompting.md.

The research loop

Five steps. One pass minimum. Iterate until every gap is closed.

1. Plan. Call get-research-consultancy with a goal that names the

topic, the user's use case, known unknowns to skip, what NOT to research, freshness window, and quote discipline. The goal is the highest-leverage prompting decision in the entire loop — a weak goal produces a generic brief, which produces wandering keywords, which produces shallow synthesis. See references/prompting.md.

1. Reconnoiter. Fan out 15-50 keywords with web-search. Write

keywords as Google retrieval probes — name the source class, anchor on discriminating terms, use one operator. For Reddit permalink discovery, write explicit site:reddit.com/r/.../comments probes; there is no separate scope parameter. Adjective-rotation on the same noun phrase is wasted budget.

Fire search calls in parallel when intents differ. Two web-search calls in one turn — one with plain web probes (vendor docs, GitHub, blogs, changelog), one with site:reddit.com/r/.../comments probes (sentiment, migration, dissent) — is the canonical reconnaissance pattern. The round runs in roughly the time of one call.

1. Triage. Read the ranked URL list. Aim for 5-15 candidate URLs to

scrape. Sort by CONSENSUS score and source authority.

1. Capture. Use scrape-link with a defined extract

(≤5 URLs per call, ≤7 facets per call). For Reddit threads, write an extract that preserves attribution and dissent (e.g. verbatim quotes with author + score | agreement reasons | dissent reasons | migration drivers) — the Reddit API path still returns the full threaded post and comments before extraction runs on top. Read every ## Not found section returned; it tells you which gaps to chase next round.

1. Synthesize. Every numeric, versioned, priced, or error-string claim

traces to a verbatim scraped quote. Snippet citations are forbidden — snippets lie; the page is canonical. Mark inference vs evidence explicitly. Surface contradictions; do not paper over disagreement.

Two to four search rounds per substantive session is normal. After each capture, harvest ## Follow-up signals and ## Not found from scrape-link, and re-run web-search with terms gathered from those sections. Stop only when gaps_to_watch and stop_criteria are closed, or when remaining gaps are explicitly unresolvable from available sources.

Multi-agent orchestration (deep single-question path)

Single-agent research is the default. Use the orchestrated path only when one technical question genuinely spans 3+ distinct technical subdomains that benefit from independent reading lenses — e.g. security + performance + maintainer intent + migration experience.

Pattern: dispatch one subagent per subdomain in parallel, each with its own get-research-consultancy goal and its own search intent. Each returns a section synthesis. The orchestrator merges, reconciles contradictions between sections, and produces the unified answer.

The output still defaults to one Markdown synthesis. If the user explicitly asks for files, a small numbered folder is allowed; do not build per-entity packs, product profiles, comparison templates, or reusable source-ledger corpora here — those are corpus shapes and belong in run-deep-research.

For the full orchestration protocol — subagent prompts, scope allocation, merge strategy, and contradiction resolution — read references/orchestrator.md. Below the threshold, stay single-agent for coherence.

Reference routing

| Question | Read | |---|---| | How do I drive a specific tool? Parameters, output formats, thresholds. | references/tools.md | | How do I write a get-research-consultancy goal or a scrape-link extract? | references/prompting.md | | What does an end-to-end research session look like for my scenario? | references/workflows.md | | How do I cite, mark inference, surface contradictions, format output? | references/synthesis.md | | A scrape timed out. A search returned 0 results. The provider cascade failed. Now what? | references/failure-modes.md | | The question spans 3+ technical subdomains and needs parallel evidence gathering. | references/orchestrator.md |

Output and citation contract

Default to in-chat Markdown unless the user asks for a file. Use JSON only when explicitly requested.

| Request shape | Default output | |---|---| | quick fact check | 3-8 bullets with sources | | bug / root cause | likely cause, fix, caveats, fallback | | decision / comparison (≤4 options) | recommendation, confidence, table, flip conditions, counter-arguments | | deep single-question research | 800-2,000 words plus source ledger |

For any non-trivial answer, include compact source notes:

• URL or source identifier
• source type (docs, changelog, issue, advisory, Reddit thread, blog)
• author/date when available
• access date or research date for time-sensitive claims
• claim supported
• confidence or caveat when source quality is weak

Minimum citation rules:

• Cite scraped pages, official docs, issues, posts, advisories, or other

concrete sources. Never cite search snippets or tool-provided synthesis as evidence.

• For APIs, prices, CVEs, versions, model behavior, deprecations, and

fast-moving libraries, verify before synthesizing. Prefer official docs, changelogs, release notes, and advisories for exact facts.

• Use practitioner sources for production behavior, not exact API truth.
• Separate confirmed facts from inference. Mark unresolved gaps instead

of smoothing them into a confident answer.

• "Reddit consensus" is not a citation. Attribute Reddit evidence with

username, subreddit, date, and preferably score/comment context.

Read references/synthesis.md for credibility tiers, contradiction handling, and worked output examples.

Operational guardrails

• Use get-research-consultancy first for substantive sessions. A quick

fact check can skip it when the overhead does not pay back.

• Cap scrape-link at 5 URLs and 7 facets per call; split beyond that.
• Read every ## Not found section and feed unresolved gaps into the

next query.

• Plan for output volume before parallel web-search calls; large URL

pools may need file-backed triage.

• Treat provider cascade failure as blocking or WAF behavior. Route

around to mirrors, archives, postmortems, or quoted discussions — see references/failure-modes.md.

Final checks

• description is single-line, starts with Use if, and

is ≤100 characters

• run-research target-specific validator checks pass
• every reference file remains routed from SKILL.md
• output contract includes source attribution and unresolved gaps
• sibling redirects still name run-deep-research and

run-github-scout

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: yigitkonur
• Source: yigitkonur/skills-by-yigitkonur
• License: MIT

Install and usage instructions live in the source repository linked above.