laststance/search

/search — Iterative Multi-Tool Research

Codex Compatibility

When running this skill in Codex, translate Claude Code-only primitives before acting: AskUserQuestion -> chat/request_user_input, TodoWrite -> update_plan, Task/TaskCreate/TeamCreate/SendMessage -> spawn_agent/send_input/wait_agent when available and allowed, and EnterPlanMode/ExitPlanMode -> a concise chat plan plus explicit approval. Resolve Read/Write/Edit/Bash/WebSearch/WebFetch to Codex file/shell/web tools, and map ~/.claude/... paths to ~/.agents/... or ~/.codex/... unless the task explicitly targets Claude Code.

Cursor Compatibility

When running this skill in Cursor Agent, translate Claude Code-only primitives before acting: AskUserQuestion -> AskQuestion; TodoWrite -> Cursor TodoWrite or an equivalent checklist; Task/TaskCreate/TeamCreate/SendMessage/multi-agent flows -> Cursor Task (subagents), parallel Tasks, or run_in_background when allowed (TeamCreate/SendMessage may have no exact match); EnterPlanMode/ExitPlanMode -> Plan mode (SwitchMode / CreatePlan) plus explicit user approval. Resolve Read/Write/Edit/StrReplace/Bash/web/search/MCP via Cursor Composer or Agent equivalents. MCP names written as mcp__server__tool typically map to call_mcp_tool with configured server identifiers. Map ~/.claude/... to ~/.cursor/skills/, .cursor/skills/, and .cursor/rules/ unless the task explicitly targets Claude Code.

Research the user's question by orchestrating every available search tool — WebSearch, WebFetch, and installed MCPs (Exa, Perplexity, Tavily, Context7, DeepWiki) — iterating with different tool families until the answer is satisfactory or three passes are exhausted.

Workflow

Step 1 — Classify the question

Pick the best first tool based on question type. Use the table below; if the chosen MCP is not installed, fall back to the next row.

| Question type | First tool | Why | |---|---|---| | Named library / framework / SDK API | mcp__context7__resolve-library-id → mcp__context7__query-docs | Authoritative, version-correct | | Specific GitHub repo internals | mcp__deepwiki__ask_question | Indexed wiki + code | | Known URL to read | WebFetch (or mcp__exa__web_fetch_exa) | Direct content | | Comparison / "which is better" / synthesis | mcp__perplexity__reason | Built for reasoning across sources | | Deep multi-source research | mcp__perplexity__deep_research or mcp__tavily__tavily_research | Long-form, multi-citation | | Recent events / news | WebSearch + mcp__tavily__tavily_search | Fresh index | | General factual lookup | WebSearch | Fast default |

If no MCPs are installed, use WebSearch + WebFetch only.

Step 2 — Search loop (max 3 passes)

Pass 1: Run the best-fit tool with a focused query.
        → Apply Step 3 satisfaction criteria.
Pass 2: If gaps remain → switch to a DIFFERENT tool family
        (different web index or different reasoning style).
Pass 3: If still unresolved → escalate to deep_research
        / tavily_research, OR fetch the most promising URL
        directly with WebFetch.

Stop early as soon as the answer is satisfactory. After Pass 3, surface what is known with a confidence label and ask the user to narrow scope rather than burning more passes.

Step 3 — Satisfaction criteria

A result is satisfactory only when all of the following hold:

• The core question is directly answered (not just adjacent topics).
• At least one citation / source URL backs each non-trivial claim.
• No major contradiction between sources. If contradicted, reconcile or flag.
• Recency matches the question (current-events questions → sources < 6 months old).

If any criterion fails, run another pass with a different tool family or a materially refined query (new keywords, narrower scope, different framing).

Step 4 — Tool diversity rule

Each pass MUST use a different tool family. Re-running the same tool with reworded text rarely yields new information. Prefer crossing these boundaries:

• WebSearch ↔ Tavily ↔ Perplexity (different web indices)
• Context7 ↔ DeepWiki (different doc sources)
• Search ↔ Fetch (broad → specific)

Step 5 — Answer format

Respond in the user's language. Default template:

**Answer**: <one or two sentences, direct>

**Why** (key evidence):
- <claim> — <source url>
- <claim> — <source url>

**Confidence**: high / medium / low
**Tools used**: <list>
**Open questions** (if any): <…>

Keep raw tool dumps out of the answer — synthesize, then cite.

Tool quick-reference

Always available

• WebSearch — generic web search
• WebFetch — fetch a known URL

MCPs (use whichever are installed)

• mcp__exa__web_search_exa, mcp__exa__web_fetch_exa — Exa search & fetch
• mcp__perplexity__search — Perplexity quick search
• mcp__perplexity__reason — synthesis / comparison
• mcp__perplexity__deep_research — long-form research
• mcp__tavily__tavily_search — Tavily search
• mcp__tavily__tavily_extract — extract content from a URL
• mcp__tavily__tavily_research — Tavily deep research
• mcp__tavily__tavily_crawl, mcp__tavily__tavily_map — site exploration
• mcp__context7__resolve-library-id → mcp__context7__query-docs — library docs
• mcp__deepwiki__ask_question, read_wiki_contents, read_wiki_structure — GitHub wiki

Anti-patterns

• Re-running the same tool with a slightly reworded query — switch tool families instead.
• Skipping classification (Step 1) and defaulting to WebSearch for everything.
• Stopping at Pass 1 when the answer is partial or uncited.
• Going past Pass 3 — diminishing returns; surface what is known and ask the user to narrow scope.
• Pasting raw tool output as the answer — always synthesize and cite.
• Using deep_research / tavily_research on Pass 1 — too slow; reserve for escalation.

Examples

Example 1 — "How do I use Suspense with use() in React 19?"

• Pass 1: context7__resolve-library-id → query-docs for React 19.
• Satisfactory? Yes (official docs, code example, version-correct) → answer with cited snippet. Done in one pass.

Example 2 — "Latest funding round for Anthropic"

• Pass 1: WebSearch → news headlines.
• Pass 2: tavily_search to confirm date and amount across multiple outlets.
• Cross-checked → answer with two source URLs and confidence: high.

Example 3 — "Why is my Next.js v16 build slow with Turbopack?"

• Pass 1: context7__query-docs for Next.js v16 perf docs.
• Pass 2: deepwiki__ask_question on vercel/next.js for known issues.
• Pass 3: WebSearch for recent GitHub discussions / blog posts.
• Synthesize root causes with citations; flag any unresolved hypotheses as open questions.