Start here: choose a pattern#
Start with Minimum sufficient provider set for everyday research. Use Common one-provider patterns for a restricted provider, Exa Connect patterns for specialized datasets, or Complementary-provider patterns for distinct roles.
Flexible routing and provider audits contains advanced patterns. The General-purpose fill-in template lets you specify required, optional, and excluded providers.
Use the section links below to jump to a pattern.
Testing status: Library blocks are reusable templates, not verified end-to-end runs unless a linked run record says otherwise. For an executed product comparison, see the ChatGPT + TinyFish walkthrough. The Exa Ultra case study records a separate advanced run; it does not validate every Ultra prompt here.
Minimum sufficient provider set#
With the current routing skill installed, start with this everyday prompt:
Use the multi-provider-research skill.
Research question: [YOUR QUESTION]The skill handles tool discovery, source checks, avoiding duplicate work, and reporting gaps. Add only the choices specific to your task: providers, who writes the analysis, fallback permission, scope, and budget. The control clauses below cover those choices.
Defaults at a glance
- Use the smallest useful set of available providers; ordinary lookups use source retrieval and analysis by the current assistant.
- Native search and page reading may help when your restrictions allow them. They cannot replace a required provider or bypass an exclusive tool set.
- If a required provider or reader is unavailable, report the gap and complete other permitted work. Do not substitute unless a fallback is already permitted. For unrestricted tasks, another suitable tool may be used with disclosure.
- Reuse sufficient source text. Reading the same page again is unnecessary unless important context is missing or needs checking.
The examples below rely on these skill defaults rather than repeating them. Each copied block retains its task-specific restrictions. Use the current skill package (new tab); the prose around a prompt is not a substitute for loading the skill.
When not to invoke the skill#
Installing the skill does not authorize it for every task. A mention of a provider in a copyediting task or a general browser task does not by itself call for research routing. Use this prompt when only the current application’s built-in search and page-reading tools are permitted:
Use only the web search and page-fetching tools provided natively by this harness.
Do not invoke, load, or apply the multi-provider-research skill. Do not call any separately connected MCP server, plugin, connector, provider skill, command-line tool, or external search or fetch provider.
If a native tool is unavailable or cannot access a required source, report the gap and stop that evidence lane. Do not substitute another interface.
In the final response, list the exact search and fetch tools used.“Native” describes the tool surface presented by the application. It does not prove that the application vendor uses no external infrastructure.
For a hard technical guarantee, disable external tools and automatic skill loading or use a clean profile, then review the tool-call record.
Useful control clauses#
These clauses can be added to any example.
Require, permit, or restrict providers
Reusable template · no end-to-end run record attached
Link to this exampleRequired participation, optional participation, and exclusivity are separate choices:
Use Exa for discovery. You may add another provider only when it supplies a distinct necessary role within this task.Use exactly Exa and Octen as external providers. Both must contribute a distinct role. Do not call another external provider. You may use the harness-native page reader to verify cited sources.You may use TinyFish if a material primary source cannot be read adequately with the normal page reader.You may use exactly one task-appropriate Perplexity mode if it contributes a distinct discovery or synthesis role.You may use Firecrawl Map to discover relevant URLs on a site, then Scrape selected pages or run one conservatively bounded Crawl. Do not use all three when one operation is sufficient.You may use Parallel Search and Fetch for one evidence lane. Use Parallel Deep Research or Task Group only if provider-side synthesis is permitted and materially useful.Exclude providers or interfaces
Reusable template · no end-to-end run record attached
Link to this exampleUse Exa Search and Fetch, but do not use Exa Agent or Connect.Use Perplexity Search, but do not use Perplexity Ask, Reason, or Research.Do not use generic web search, browser fetching, or scraping as a substitute for X-native evidence.Control fallback behavior
Reusable template · no end-to-end run record attached
Link to this exampleStrict:
If a required provider is unavailable, report the gap and stop that evidence lane. Do not substitute another provider.Flexible:
If a preferred provider is unavailable, disclose that first and use the closest available fallback. Label the substitution clearly.Approval-based:
If a required provider is unavailable, do not substitute it without asking me first.Apply the selected fallback policy to the blocked evidence role. Complete independent permitted work rather than stopping the entire task. Continue until the requested deliverable has sufficient inspected evidence or remaining gaps are blocked by access, explicit restrictions, budget, or the selected tool’s lifecycle. Ask for clarification only when an unresolved boundary materially affects scope, cost, or data handling.
Control duplication
Reusable template · no end-to-end run record attached
Link to this exampleDo not send the same discovery query through multiple providers unless I explicitly request cross-validation.Use Octen and Exa for independent cross-checking, but do not fetch the same URL twice unless the evidence conflicts.Do not send the same subquestion through multiple Perplexity modes. Choose Search, Ask, Reason, or Research according to the required output.Do not launch an equivalent duplicate workflow while an asynchronous run remains active. Retain its identifier and follow the selected tool’s lifecycle instructions. If the tool instructs you to return a progress URL and stop, do so unless I have requested monitoring.Require an auditable report
Reusable template · no end-to-end run record attached
Link to this exampleFor ordinary work, report tools, contributions, and important limitations briefly. Use the fuller ledger below when provenance, coverage, reproducibility, or an explicit audit request warrants it; a simple X lookup or two-provider task does not automatically need one.
Finish with a concise provider ledger covering requested providers, providers actually used, exact interfaces, query transformations, operators, date ranges, contributions, failures, substitutions, and evidence gaps.Common one-provider patterns#
Tested beginner example: Compare free Google Meet and Zoom with ChatGPT + TinyFish. Read the prompt, actual answer, and verification limits before adapting it.
Octen Search, Broad Search, and Extract
Reusable template · no end-to-end run record attached
Link to this exampleChoose Highlights first for focused research or Full content when needed for closer reading. Both patterns inspect returned evidence and allow fuller retrieval when necessary. They work through the supported Octen interface available in your application.
Highlights first
Use the multi-provider-research skill.
Use only Octen for search and page reading; exclude native tools. Keep analysis and writing in this assistant, using retrieval and source-content modes only.
Start with query-relevant highlights: enable highlights and disable full content using the available controls. Start with at most 5 results per search (including each Broad Search sub-query) and at most 3 Broad Search sub-queries. Choose a suitable highlight token cap. Expand these starting budgets only to address a material evidence gap.
Use Octen Extract when highlights lack sufficient context or freshness. Briefly report whether highlights were sufficient, any fuller-text retrieval, and unread sources. Do not claim billing savings without billing evidence.
Research question: [YOUR QUESTION]Full content when needed
Octen’s pricing page, checked on 13 September 2026, includes full content for the first ten results per search without an additional content charge. Search charges and model context costs remain. Explicitly request full content when additional context is needed rather than relying on an integration’s defaults. The following prompt uses conservative starting limits while allowing evidence-driven follow-up work. Pricing (new tab)
Use the multi-provider-research skill.
Use only Octen for search and page reading; exclude native tools. Keep analysis and writing in this assistant, using retrieval and source-content modes only.
Enable full content when additional source context is needed. Start with at most 10 results per search (including each Broad Search sub-query) and at most 5 Broad Search sub-queries. Choose a suitable per-result token cap; 2048 is a possible starting point. These limits do not guarantee complete text or coverage. Expand them only to address a material evidence gap and explain why.
Briefly report the operations and important count/sub-query/token settings, whether the text was sufficient, and extraction fallbacks or unread sources. Do not claim billing savings without billing evidence.
Research question: What are the strongest arguments for and against open-weight AI regulation?Exa Search and Exa Fetch
Reusable template · no end-to-end run record attached
Link to this exampleUse the multi-provider-research skill.
Use only regular Exa Search and Exa Fetch as external research tools. Do not use Exa Agent, Exa Connect, or another external provider.
Find authoritative and recent sources. Inspect returned source content before citing it. Reuse sufficient excerpts or full text; retrieve more context when evidence is ambiguous, incomplete, conflicting, stale, or missing material qualifications.
Use only the providers and operations named above for discovery and page reading; do not use a native page reader or add another provider.
Retrieve source content or faithful extracts only; no provider-generated summaries, answers, reports, or agent analysis. Keep all analysis and final writing in this assistant.
Research question: Compare the documented export formats and offline access of Notion and Obsidian for a personal research notebook.Exa Advanced Search
Reusable template · no end-to-end run record attached
Link to this exampleUse the multi-provider-research skill.
Use Exa Advanced Search as the only external discovery provider because precise date and domain controls are required. Search publications from January 1, 2026 onward and prioritize official regulatory domains.
Use Exa Fetch for selected result URLs if page content is required. Do not use Exa Agent, Connect, or another external provider.
Use only the providers and operations named above for discovery and page reading; do not use a native page reader or add another provider.
Retrieve source content or faithful extracts only; no provider-generated summaries, answers, reports, or agent analysis. Keep all analysis and final writing in this assistant.
Research question: Find newly published official guidance concerning AI model evaluations.Exa specialized search verticals
Reusable template · no end-to-end run record attached
Link to this exampleExa’s people, company, code, and news verticals are search indexes, not Connect providers.
Use the multi-provider-research skill.
Use Exa Search as the only external discovery provider and select the specialized vertical matching the evidence required:
- professional profiles: `category:people` in regular Search, or the `people` category in Advanced Search when exposed;
- companies: `category:company` in regular Search, or the `company` category in Advanced Search when exposed;
- working code examples: a code-specific semantic query and the Advanced Search `github` category when exposed;
- current news: a news-specific query with the Advanced Search `news` category and date controls when exposed.
Inspect returned source content before citing it. Reuse sufficient excerpts or full text; retrieve more context when evidence is ambiguous, incomplete, conflicting, stale, or missing material qualifications. Keep profile, company, code, and news evidence separate from inference.
Use only the providers and operations named above for discovery and page reading; do not use a native page reader or add another provider.
Use Exa Fetch when selected result excerpts lack necessary context.
Retrieve source content or faithful extracts only; no provider-generated summaries, answers, reports, or agent analysis. Keep all analysis and final writing in this assistant.
Research question: [YOUR PEOPLE, COMPANY, CODE, OR NEWS QUESTION]Exa Agent Ultra for comprehensive discovery
Reusable template · no end-to-end run record attached
Link to this exampleUse this when broad coverage and difficult qualification criteria justify provider-generated research. This is an illustrative, unexecuted prompt. The $10 and 30-minute settings are example limits: adjust them before use. The duration control is a soft ceiling. See when Ultra is worth it.
Use the multi-provider-research skill.
Use Exa Agent Ultra for comprehensive discovery within this scope:
- Entity type and topic: [WHAT TO FIND]
- Geography and date window: [SCOPE]
- Required qualification criteria: [CRITERIA]
Provider-side Agent research is permitted. Use Exa's web research only; do not enable Connect datasets, another provider, or a native page reader.
Before starting, inspect the callable schema for Ultra and budget controls. Set effort to ultra, maxCostDollars to 10, and maxDurationSeconds to 1800 through the supported budget fields. If this interface cannot express those limits, explain the gap and do not start the run. Do not substitute a spending instruction in the query for a budget control.
Return structured records with entity name, canonical URL, evidence URL and supporting detail for each criterion, relevant source dates, and unresolved fields. Allow unknown values and separate verified matches from uncertain candidates. Match output limits to the requested scope; do not silently cap the list at an arbitrary small sample.
Retain and resume the same run ID until a terminal state. Report actual returned cost, effort, stop reason, verified-match count, and coverage gaps. A completed run or satisfied schema is not proof of exhaustive coverage. Do not start additional runs beyond this one-run budget.Exa Agent Ultra to expand an existing list
Reusable template · no end-to-end run record attached
Link to this exampleUse this for a new, budgeted follow-up to a completed run, not for resuming an active one. input.exclusion avoids previously found entities; input.data supplies rows for processing or enrichment. This example has not been executed.
Use the multi-provider-research skill.
Expand this completed Exa Agent list with additional qualifying entities:
- Completed run ID: [RUN ID]
- Existing entities, including canonical identifiers: [ROWS]
- Original scope and qualification criteria: [SCOPE AND CRITERIA]
Provider-side Agent research is permitted. Use only Exa Agent's web research, without Connect datasets, another provider, or a native page reader.
This authorizes one new Ultra run with maxCostDollars 10 and the soft maxDurationSeconds limit 1800. Inspect the live interface and set both supported budget controls before starting. If either control is unavailable, explain the gap and do not start or silently switch interfaces.
Use previousRunId for the completed run's context and input.exclusion for the existing entities. Do not use previousRunId to poll an active run. If the earlier run is still active, report that state instead of starting this follow-up. Use input.data only if rows also need processing or enrichment.
Preserve the original criteria, require evidence for each new match, allow unknown fields, and deduplicate against the supplied list and within the new results. Report new verified matches separately from uncertain candidates. Do not fill gaps with invented records or silently return only an arbitrary small sample.
Resume the new run by its own run ID. Report its actual returned cost, effort, stop reason, remaining gaps, and the combined unique count supported by the supplied list and new results. Do not claim exhaustive coverage or start another run automatically.Parallel Search and Fetch
Reusable template · no end-to-end run record attached
Link to this exampleUse the multi-provider-research skill.
Use only Parallel Search and Fetch as external research tools. Give Search a concise objective plus short retrieval queries. Reuse the returned session identifier for any related Fetch calls, and request full content only for selected URLs whose excerpts are insufficient.
Do not launch Parallel Deep Research, Task Group, or another external provider.
Use only the providers and operations named above for discovery and page reading; do not use a native page reader or add another provider.
Retrieve source content or faithful extracts only; no provider-generated summaries, answers, reports, or agent analysis. Keep all analysis and final writing in this assistant.
Research question: What are the official installation requirements for the latest stable Python release on Windows and macOS?Parallel Deep Research
Reusable template · no end-to-end run record attached
Link to this exampleUse the multi-provider-research skill.
Use only Parallel Deep Research because this task requires one substantial provider-generated report. Preserve its task URL or identifier and stop after creation unless I explicitly ask you to monitor it.
Do not launch a duplicate Search workflow while the task is active. When the result is later retrieved, distinguish provider-generated synthesis from underlying sources and report the exposed processor or model, if any.
Do not use another provider or a native page reader. Report missing source evidence as a gap when completed results can be inspected.
Research question: Produce a comprehensive report on frontier-model incident-reporting regimes in the United States, European Union, and United Kingdom.Parallel Task Group
Reusable template · no end-to-end run record attached
Link to this exampleUse the multi-provider-research skill.
Use only Parallel Task Group because the same structured output contract must be applied independently to several inputs. Begin with three to five inputs, use one shared output schema, and preserve the task-group URL or identifier.
Do not poll or duplicate the batch unless I explicitly ask you to monitor it.
Do not use another provider or a native page reader. Report missing source evidence as a gap when completed results can be inspected.
Inputs: [COMPANY OR PRODUCT 1], [2], [3]
Required fields: official name, primary domain, current product category, latest dated launch evidence, and source URLs.Firecrawl Map, Scrape, Crawl, or Parse
Reusable template · no end-to-end run record attached
Link to this exampleUse the multi-provider-research skill.
Use only Firecrawl as the external provider. Inspect its live schema and choose the smallest operation that fits:
- Map to identify relevant URLs on one site without reading their bodies;
- Scrape for one known webpage or schema-constrained fields;
- Crawl for a bounded collection of related pages, with explicit path, depth, and page limits;
- Parse for a supported document.
Do not use Agent, Interact, or Monitor. Preserve source URLs and report excluded, failed, or inaccessible pages.
Use only the providers and operations named above for discovery and page reading; do not use a native page reader or add another provider.
Retrieve source content or faithful extracts only; no provider-generated summaries, answers, reports, or agent analysis. Keep all analysis and final writing in this assistant.
Research task: Map [DOCUMENTATION SITE], then extract only the pages describing authentication, rate limits, and pricing.Firecrawl with optional Alexandria data
Reusable template · no end-to-end run record attached
Link to this exampleUse this when the task needs comparable fields or structured records. Alexandria is optional: sufficient search or page evidence can complete the task. Catalog discovery is free, while data execution uses the selected capability's listed price. Ordinary web search and page extraction have their own charges. See Alexandria coverage and costs.
Use the multi-provider-research skill and Firecrawl.
Research question: [QUESTION]
Entities or scope: [ENTITIES, LOCATION, AND PERIOD]
Required fields: [FIELDS AND SOURCE OR DATE REQUIREMENTS]
Total Firecrawl credit budget for this task: [MAXIMUM CREDITS, OR ASK ME BEFORE PAID CALLS]
Use sufficient search results or page content when they answer the question. Do not add paid enrichment merely because a provider capability is available.
If structured provider data would fill a remaining evidence gap or avoid repeated page collection, discover relevant Alexandria capabilities. Inspect the selected capability's inputs, response fields, coverage, published price, and billing unit, including any stated rounding or special-mode rates. Some capabilities may list zero execution credits. Execute only a discovered fit within my budget; include additional result pages in the cost estimate. Ask before paid execution if the price or budget is unresolved. Do not accept provider terms without my explicit approval.
Use source records or faithful extraction and have the model selected in this application perform the analysis. Preserve sources and observation dates. Report the underlying provider, material gaps, and credits per call, distinguishing published prices from confirmed usage. If no capability fits, use sufficient permitted web evidence or explain the gap.TinyFish discovery, page extraction, or browser automation
Reusable template · no end-to-end run record attached
Link to this exampleTinyFish-only requests work best when you know whether the task needs discovery, rendered page content, or browser interaction. Exact tool names vary by application.
Use the multi-provider-research skill.
Use only TinyFish as the external provider. Inspect its live tools, then choose its discovery capability for web, news, or research-paper discovery; page extraction for browser-rendered content; or browser automation when navigation, form filling, or another stateful interaction is necessary.
Inspect the following JavaScript-heavy websites and extract their current product features, pricing, and stated limitations:
- [URL 1]
- [URL 2]
Report any pages, interactions, or fields that could not be accessed.
Use only the providers and operations named above for discovery and page reading; do not use a native page reader or add another provider.
Retrieve source content or faithful extracts only; no provider-generated summaries, answers, reports, or agent analysis. Keep all analysis and final writing in this assistant.
Perplexity Search
Reusable template · no end-to-end run record attached
Link to this exampleStart here to find sources with Perplexity and have your own assistant turn the evidence into an answer. Copy the prompt and add your question; the assistant chooses the search type for you, using Fast Search for routine lookups and standard search for harder questions or important evidence gaps.
For the details, see Fast or standard Search?. Ask, Reason, and Research below are optional alternatives when you want Perplexity itself to generate an answer or analysis.
Use the multi-provider-research skill.
Use Perplexity Search as the only external discovery provider. Inspect its live schema. Prefer Fast Search (search_type: "fast") for routine questions; choose standard search (search_type: "web") directly for rare, difficult, or ambiguous questions. Apply supported domain, date, or country controls when relevant.
Reuse sufficient source excerpts. You may use the harness-native page reader to inspect selected sources. If a relevant source already found could resolve a gap and its excerpts are insufficient, read it before making another search request.
If a successful Fast Search leaves a material evidence gap, use targeted standard-search follow-ups for the missing or conflicting evidence. Do not repeat parts already answered or automatically run both search types. Stop when the evidence is sufficient and say when standard search was used.
Keep synthesis in this assistant. Do not use Perplexity Ask, Reason, or Research, or another external provider.
Use only supported controls. If Fast Search is unavailable but standard Perplexity Search is supported, disclose this and use standard search. Report failed requests separately from evidence gaps; do not silently switch providers or answer-generation modes.
Cite the sources supporting the answer and report any unresolved evidence gaps.
Research question: [YOUR QUESTION]Perplexity Ask
Reusable template · no end-to-end run record attached
Link to this exampleUse the multi-provider-research skill.
Use Perplexity Ask as the only external research mode. You may use the harness-native page reader to verify cited sources; do not use another external provider. Use it for a focused factual answer or concise web-grounded explanation with citations. Make the input specific and apply authoritative-domain, recency, and context controls when the live schema exposes them.
Return no more than six bullets. If searches return only adjacent or outdated material, state the mismatch and do not guess.
Treat the generated answer as synthesis. Preserve its cited URLs and use the harness-native page reader to open the underlying sources supporting material claims before final citation.
Research question: How does Zotero help a student collect sources and create a bibliography? Use its official documentation.Perplexity Reason
Reusable template · no end-to-end run record attached
Link to this exampleUse the multi-provider-research skill.
Use Perplexity Reason as the only external research mode. You may use the harness-native page reader to verify cited sources; do not use another external provider. Use it because the task requires evaluating alternatives against explicit criteria using current web evidence. Inspect the live schema and apply only the domain, recency, or context controls it supports.
Separate sourced facts from comparative reasoning. Preserve cited URLs and open the underlying sources supporting material claims before final citation.
Research question: Compare the compliance trade-offs of releasing a frontier model through an API versus publishing its weights under the current EU regulatory framework.Perplexity Research
Reusable template · no end-to-end run record attached
Link to this exampleUse Research only when comprehensive synthesis justifies its additional latency.
Use the multi-provider-research skill.
Use Perplexity Research as the only external research mode. You may use the harness-native page reader to verify cited sources; do not use another external provider. Use it for a deep multi-source investigation. Inspect and follow the live Research schema rather than assuming that filters available to other Perplexity tools are supported.
Treat the generated report as synthesis. Retain its cited URLs, open the underlying sources for material claims, and report citations that cannot be verified.
Research question: Produce a literature-style review of empirical methods used to evaluate frontier-model cyber capabilities since 2024.xurl
Reusable template · no end-to-end run record attached
Link to this exampleUse the multi-provider-research skill.
Use only xurl for X-native evidence. Search read-only X data from [START DATE] through [END DATE]. Preserve query operators, post dates, authors, engagement fields, and result limits.
Do not use another X API integration, web search, browser fetching, or scraping. If xurl is unavailable or unauthenticated, report the access gap and stop.
Research question: What are AI researchers on X saying about mechanistic interpretability?Authenticated X API integration
Reusable template · no end-to-end run record attached
Link to this exampleUse the multi-provider-research skill.
Use only the authenticated X API integration available in the current environment. Inspect its live capabilities and choose the structured operation that supports full-archive post search.
Do not use xurl, web search, browser fetching, or scraping. Report query operators, date range, archive coverage, pagination, and access limitations.
Research question: Find posts by major AI laboratory leaders discussing model regulation during 2025.Exa Connect patterns#
On 3 September 2026, the live Exa dataSources schema inspected while preparing this guide exposed eight self-serve Connect providers: Fiber, Financial Datasets, Similarweb, Baselayer, Affiliate.com, Particle, Jinko, and Polymarket. Treat this dated list as a routing hint rather than a permanent specification.
Exa also documents additional partners (new tab) requiring account enablement.
Inspect the available Connect data sources and their schema before first use; reuse that knowledge unless the surface changes or validation fails. The Agent interface may expose these through a dataSources field or another supported control. Do not assume access or commercial terms from documentation.
Use this common structure:
Use the multi-provider-research skill.
Use Exa Agent with only the [CONNECT PROVIDER] data source. Confirm that it appears in the live `dataSources` schema before starting.
Request provider-backed fields for [FIELDS] in the `outputSchema`. Include sources, observation dates, and field-level provenance where available.
Retain the run identifier and follow the selected interface’s lifecycle, resuming the same run when it remains active. Inspect completed results for actual provider contribution, field provenance, errors, and partial coverage. Requested data sources alone do not prove contribution. Do not replace unavailable provider-backed fields with generic web inference.
Use only the named Connect data sources through Exa Agent. Do not add generic web search, another provider, or a native page reader. Provider-side Agent processing is permitted; distinguish returned provider-backed fields from generated conclusions.
Research question: [YOUR QUESTION]Fiber
Reusable template · no end-to-end run record attached
Link to this exampleUse the multi-provider-research skill.
Use Exa Agent with only the Fiber Connect provider. Confirm that Fiber is present in the live dataSources schema.
Request provider-backed fields for company identity, domain, industry, employee count, funding stage, relevant professional role, employer, job-change signal, observation date, and field-level provenance.
Inspect the completed run to confirm that Fiber contributed. Do not replace unavailable Fiber fields with generic web inference.
Use only the named Connect data sources through Exa Agent. Do not add generic web search, another provider, or a native page reader. Provider-side Agent processing is permitted; distinguish returned provider-backed fields from generated conclusions.
Retain the run identifier and follow the selected interface’s lifecycle. Resume the same active run rather than starting a duplicate; report pending work separately from inspected completed results.
Research question: Identify U.S. AI infrastructure companies with 50–500 employees and recent senior engineering hires.Similarweb
Reusable template · no end-to-end run record attached
Link to this exampleUse the multi-provider-research skill.
Use Exa Agent with only the Similarweb Connect provider. Confirm that Similarweb is present in the live dataSources schema.
Request provider-backed fields for domain, estimated visits, reporting period, rank, engagement measures, competitor domains, observation date, and field-level provenance.
Inspect the completed run to confirm that Similarweb contributed. Do not replace unavailable Similarweb fields with generic web estimates.
Use only the named Connect data sources through Exa Agent. Do not add generic web search, another provider, or a native page reader. Provider-side Agent processing is permitted; distinguish returned provider-backed fields from generated conclusions.
Retain the run identifier and follow the selected interface’s lifecycle. Resume the same active run rather than starting a duplicate; report pending work separately from inspected completed results.
Research question: Compare the recent digital reach of ten AI developer-tool companies.Financial Datasets
Reusable template · no end-to-end run record attached
Link to this exampleUse the multi-provider-research skill.
Use Exa Agent with only the Financial Datasets Connect provider. Confirm that `financial_datasets` is present in the live dataSources schema.
Request provider-backed fields for ticker, reporting period, price timestamp, revenue, operating income, free cash flow, ownership or filing data as applicable, SEC form, filing date, and direct filing link. Preserve units and observation dates.
Inspect the completed run to confirm that Financial Datasets contributed. Do not replace unavailable provider fields with generic web figures.
Use only the named Connect data sources through Exa Agent. Do not add generic web search, another provider, or a native page reader. Provider-side Agent processing is permitted; distinguish returned provider-backed fields from generated conclusions.
Retain the run identifier and follow the selected interface’s lifecycle. Resume the same active run rather than starting a duplicate; report pending work separately from inspected completed results.
Research question: Compare the latest reported revenue growth, operating margin, and free cash flow of five publicly traded U.S. data-center infrastructure companies.Baselayer
Reusable template · no end-to-end run record attached
Link to this exampleUse the multi-provider-research skill.
Use Exa Agent with only the Baselayer Connect provider. Confirm that Baselayer is present in the live dataSources schema.
Request provider-backed fields for legal business name, jurisdiction, registration status, entity structure, officers, verification status, risk indicators, source evidence, and observation date. Do not collect private contact details.
Inspect the completed run to confirm that Baselayer contributed. Do not replace unavailable verification fields with company-website claims.
Use only the named Connect data sources through Exa Agent. Do not add generic web search, another provider, or a native page reader. Provider-side Agent processing is permitted; distinguish returned provider-backed fields from generated conclusions.
Retain the run identifier and follow the selected interface’s lifecycle. Resume the same active run rather than starting a duplicate; report pending work separately from inspected completed results.
Research question: Verify the legal registration and current business status of the following U.S. companies: [COMPANY LIST].Affiliate.com
Reusable template · no end-to-end run record attached
Link to this exampleUse the multi-provider-research skill.
Use Exa Agent with only the Affiliate.com Connect provider. Confirm that `affiliate` is present in the live dataSources schema.
Request provider-backed fields for product, brand, model, price, currency, merchant, availability, offer URL, and observation time. Keep variants separate.
Inspect the completed run to confirm that Affiliate.com contributed. Do not replace unavailable offers with generic shopping-search results.
Use only the named Connect data sources through Exa Agent. Do not add generic web search, another provider, or a native page reader. Provider-side Agent processing is permitted; distinguish returned provider-backed fields from generated conclusions.
Retain the run identifier and follow the selected interface’s lifecycle. Resume the same active run rather than starting a duplicate; report pending work separately from inspected completed results.
Research question: Find current offers for laptops with at least 32 GB of RAM, a discrete GPU, and a price below [BUDGET] in [COUNTRY].Particle
Reusable template · no end-to-end run record attached
Link to this exampleUse the multi-provider-research skill.
Use Exa Agent with only the Particle Connect provider. Confirm that Particle is present in the live dataSources schema.
Request provider-backed fields for show, episode, publication date, host, guest, speaker role, timestamp, transcript evidence, and position expressed.
Inspect the completed run to confirm that Particle contributed. Do not replace unavailable transcript evidence with generic podcast search.
Use only the named Connect data sources through Exa Agent. Do not add generic web search, another provider, or a native page reader. Provider-side Agent processing is permitted; distinguish returned provider-backed fields from generated conclusions.
Retain the run identifier and follow the selected interface’s lifecycle. Resume the same active run rather than starting a duplicate; report pending work separately from inspected completed results.
Research question: What have prominent podcast hosts and guests said about AI regulation during 2025?Jinko
Reusable template · no end-to-end run record attached
Link to this exampleUse the multi-provider-research skill.
Use Exa Agent with only the Jinko Connect provider. Confirm that Jinko is present in the live dataSources schema.
Request provider-backed fields for origin, destination, departure and return dates, cabin, fare, currency, itinerary, and observation time. Treat fares as time-sensitive discovery rather than a booking guarantee.
Inspect the completed run to confirm that Jinko contributed. Do not replace unavailable fares with generic travel-search results.
Use only the named Connect data sources through Exa Agent. Do not add generic web search, another provider, or a native page reader. Provider-side Agent processing is permitted; distinguish returned provider-backed fields from generated conclusions.
Retain the run identifier and follow the selected interface’s lifecycle. Resume the same active run rather than starting a duplicate; report pending work separately from inspected completed results.
Research question: Identify five destinations reachable from Hong Kong for [DATES] with return economy fares below [BUDGET].Polymarket
Reusable template · no end-to-end run record attached
Link to this exampleUse the multi-provider-research skill.
Use Exa Agent with only the Polymarket Connect provider. Confirm that Polymarket is present in the live dataSources schema.
Request provider-backed fields for market or event, outcome, implied probability, observation time, trading volume, liquidity, price history, market status, and direct market URL. Keep market data separate from commentary about the event.
Inspect the completed run to confirm that Polymarket contributed. Do not replace unavailable market fields with generic web commentary or inferred probabilities.
Use only the named Connect data sources through Exa Agent. Do not add generic web search, another provider, or a native page reader. Provider-side Agent processing is permitted; distinguish returned provider-backed fields from generated conclusions.
Retain the run identifier and follow the selected interface’s lifecycle. Resume the same active run rather than starting a duplicate; report pending work separately from inspected completed results.
Research question: Compare current prediction-market probabilities for [EVENT] and explain how they changed during [DATE RANGE].Complementary-provider patterns#
Octen and Exa
Reusable template · no end-to-end run record attached
Link to this exampleUse the multi-provider-research skill.
Use exactly Octen and Exa as external providers. Both must contribute a distinct role. Use their source-reading tools; do not add another external provider or a native page reader.
Use retrieval and source-content modes. Have the model selected in this application perform the analysis and final answer; do not use provider-generated summaries, answers, or agent research.
- Use Octen Broad Search for multi-angle discovery. Give it one self-contained question that preserves my intent and constraints, resolves conversational references, and fits the current input limit. Let Octen generate the subqueries.
- When source context is needed, explicitly request Octen full content using the current interface’s controls, then inspect the returned text.
- Use regular Exa Search for an independent semantic cross-check with a distinct verification purpose.
- Reuse sufficient excerpts or full text returned by either provider. Use Octen Extract or Exa Fetch for supplied URLs or missing, incomplete, unusable, or stale content. Prefer the corresponding provider’s reader for its unique results.
Avoid repeating successful queries or reading the same URL twice without a material verification reason. If a required provider is unavailable, report the gap, do not substitute another provider, and complete independent permitted work.
Briefly report each provider’s contribution and material evidence limitations.
Research question: Compare the major arguments surrounding mandatory AI incident reporting.Exa with optional TinyFish escalation
Reusable template · no end-to-end run record attached
Link to this exampleUse the multi-provider-research skill.
Use Exa as the required external provider. You may use TinyFish only for the distinct roles described below. Do not use another external provider.
- Use Exa Search for discovery and Exa Fetch for accessible sources.
- Use TinyFish only for page extraction when Exa cannot adequately retrieve an important source, or browser automation when that source requires interaction that simpler readers cannot perform.
Report which pages or evidence required TinyFish and why.
Use only the providers and operations named above for discovery and page reading; do not use a native page reader or add another provider.
Retrieve source content or faithful extracts only; no provider-generated summaries, answers, reports, or agent analysis. Keep all analysis and final writing in this assistant.
Research question: Compare the current enterprise plans and security claims of [Company A] and [Company B].Octen and Perplexity Search
Reusable template · no end-to-end run record attached
Link to this exampleUse the multi-provider-research skill.
Use exactly Octen and Perplexity Search as external providers. Both must contribute a distinct role. You may also use the harness-native page reader for citation verification; do not add another external provider.
Use retrieval and source-content modes. Have the model selected in this application perform the analysis and final answer. Do not use Perplexity Ask, Reason, or Research, or request generated summaries or answers through retrieval options.
- Use Octen Broad Search for multi-angle discovery. Give it one self-contained question that preserves my intent and constraints, resolves conversational references, and fits the current input limit. Let Octen generate the subqueries.
- When source context is needed, explicitly request Octen full content using the current interface’s controls, then inspect what returned.
- Use Perplexity Search for a distinct targeted query, applying supported domain, country, or recency controls when useful.
- Reuse sufficient excerpts or full text. Use Octen Extract for supplied URLs or Octen results with missing, incomplete, unusable, or stale content. Use the native page reader for Perplexity sources that need additional context.
Do not retrieve an adequately inspected page again without a material reason. If a required provider or reader is unavailable, disclose the gap, preserve the restrictions, and complete independent permitted work.
Briefly report provider contributions, source dates, and material access or freshness limitations.
Research question: What new government initiatives are shaping sovereign AI infrastructure in Asia?Exa and Perplexity Reason
Reusable template · no end-to-end run record attached
Link to this exampleUse the multi-provider-research skill.
Use regular Exa Search and Fetch plus Perplexity Reason.
- Use Exa for authoritative and independent source discovery and page reading.
- Use Perplexity Reason once for a grounded comparison of the documented choices and trade-offs. Do not use it as duplicate general discovery.
Treat Perplexity's response as synthesis. Preserve cited URLs and open any cited sources supporting material claims that Exa has not already retrieved.
Use only Exa Search and Fetch plus Perplexity Reason; do not add another provider or a native page reader. Use Exa Fetch for cited sources that need more context.
Research question: Compare mandatory frontier-model licensing with post-deployment incident reporting.Perplexity Research with optional TinyFish verification
Reusable template · no end-to-end run record attached
Link to this exampleUse the multi-provider-research skill.
Use Perplexity Research as the required research mode. You may use TinyFish only for the verification role described below. Do not use another external provider.
- Use Perplexity Research for one comprehensive investigation.
- Use TinyFish only to verify a cited primary source supporting a material claim when that source is dynamic, interactive, or inaccessible through ordinary page reading.
Do not send every citation to TinyFish. Report citations that remain inaccessible.
You may use the harness-native page reader to verify citations before escalating to TinyFish. Treat Perplexity output as provider-generated synthesis and inspect supporting source content for material claims.
Research question: Produce an evidence-backed review of public frontier-model transparency reporting practices.Firecrawl extraction with Parallel source verification
Reusable template · no end-to-end run record attached
Link to this exampleUse the multi-provider-research skill.
Use exactly Firecrawl and Parallel Search.
- Use Firecrawl Map to identify the relevant documentation pages, then Scrape only the selected pages needed for direct evidence.
- Use Parallel Search for a distinct objective: find independent official or primary sources that verify the documented claims.
- Do not use Firecrawl Crawl unless the map shows that a bounded multi-page collection is necessary. Do not use Parallel Task tools.
Use only the providers and operations named above for discovery and page reading; do not use a native page reader or add another provider.
Retrieve source content or faithful extracts only; no provider-generated summaries, answers, reports, or agent analysis. Keep all analysis and final writing in this assistant.
Research question: Compare the documented enterprise security controls of [PRODUCT A] and [PRODUCT B].Parallel Task Group after focused discovery
Reusable template · no end-to-end run record attached
Link to this exampleUse the multi-provider-research skill.
Use one discovery provider to identify a small qualified input set, then use Parallel Task Group once to apply the same structured enrichment contract to every input.
Preserve the selection evidence and task-group URL or identifier. Follow the selected tool’s lifecycle: if it directs you to share the URL and stop, report that the batch is pending and do not poll unless I requested monitoring. Do not run duplicate searches for each item while the batch is active. Report completed fields and gaps only after results have been retrieved and inspected.
Research question: Identify five qualifying [COMPANY TYPE] companies, then produce the same sourced product and market-position fields for each.Fiber and Baselayer through Exa Connect
Reusable template · no end-to-end run record attached
Link to this exampleUse the multi-provider-research skill.
Use Exa Agent with exactly the Fiber and Baselayer Connect providers.
- Use Fiber for company identity, domain, headcount, funding stage, and relevant professional roles.
- Use Baselayer for legal name, jurisdiction, registration status, officers, verification status, and risk indicators.
Require field-level provenance and observation dates. Keep firmographics separate from verification results.
Inspect the completed run to confirm which provider contributed to each field. Do not fill missing provider fields with generic web inference.
Use only the named Connect data sources through Exa Agent. Do not add generic web search, another provider, or a native page reader. Provider-side Agent processing is permitted; distinguish returned provider-backed fields from generated conclusions.
Retain the run identifier and follow the selected interface’s lifecycle. Resume the same active run rather than starting a duplicate; report pending work separately from inspected completed results.
Research question: Identify U.S. AI infrastructure vendors with 50–500 employees, then verify their legal registration and current business status.Fiber and Similarweb through Exa Connect
Reusable template · no end-to-end run record attached
Link to this exampleUse the multi-provider-research skill.
Use Exa Agent with exactly the Fiber and Similarweb Connect providers.
- Use Fiber for company identity, headcount, funding stage, and relevant personnel.
- Use Similarweb for estimated traffic, engagement period, rankings, and competitor domains.
Require field-level provenance and observation dates. Do not replace missing provider fields with web estimates.
Inspect the completed run to confirm which provider contributed to each field.
Use only the named Connect data sources through Exa Agent. Do not add generic web search, another provider, or a native page reader. Provider-side Agent processing is permitted; distinguish returned provider-backed fields from generated conclusions.
Retain the run identifier and follow the selected interface’s lifecycle. Resume the same active run rather than starting a duplicate; report pending work separately from inspected completed results.
Research question: Identify ten AI developer-tool companies with 50–500 employees and meaningful recent website traffic.Web, finance, and traffic evidence
Reusable template · no end-to-end run record attached
Link to this exampleUse the multi-provider-research skill.
Use three distinct evidence lanes:
1. Regular Exa Search and Fetch for official announcements and contextual web evidence.
2. Exa Agent with Financial Datasets for filed financials, price timestamps, earnings, ownership, and SEC filing links.
3. Exa Agent with Similarweb for estimated visits, reporting periods, rankings, engagement, and competitor domains.
Keep filed financial data, estimated traffic, company claims, and analyst inference separate. Require reporting periods, observation dates, provenance, and direct filing links where available.
Research question: Compare the financial performance and digital reach of five publicly traded AI infrastructure companies.Web and X-native evidence
Reusable template · no end-to-end run record attached
Link to this exampleUse the multi-provider-research skill.
Use exactly Octen, Exa, and authenticated X evidence.
- Use Octen for broad web discovery.
- Use Exa for semantic cross-checking and authoritative page reading.
- Use xurl for X-native discussion, or one authenticated X API integration instead when a dedicated structured operation fits better.
Keep web evidence and X-native evidence in separate sections. Do not send the same X query through both interfaces.
Use only the providers and operations named above for discovery and page reading; do not use a native page reader or add another provider.
Retrieve source content or faithful extracts only; no provider-generated summaries, answers, reports, or agent analysis. Keep all analysis and final writing in this assistant.
Research question: How did researchers and developers react to the announcement of [MODEL OR PRODUCT]?Advanced multi-lane investigation
Reusable template · no end-to-end run record attached
Link to this exampleThis pattern is for a comprehensive task that needs several distinct evidence families, not the normal default.
Use the multi-provider-research skill.
You may use any supported provider family only when it has a distinct role: Octen, Exa, exactly one Perplexity mode, Parallel Search or Task, Firecrawl, TinyFish, and authenticated X data.
- Octen: broad, multi-angle discovery.
- Exa: semantic cross-checking, controlled search, page reading, or relevant Connect data.
- Perplexity: choose one mode according to the required output.
- Parallel: objective-shaped Search and Fetch, or one provider-side Task workflow—not both for duplicate work.
- Firecrawl: site mapping, selected-page scraping, bounded crawling, document parsing, or an explicitly justified Agent.
- TinyFish: distinct discovery, browser-rendered page extraction, or necessary browser automation.
- X: one authenticated X-native interface; use a dedicated structured interface only when its operation fits better.
Do not automatically duplicate queries or page reads. Use a provider only when it contributes a distinct evidence lane. Report all providers attempted, their contributions, failures, date ranges, and material gaps.
Research question: Produce a comprehensive evidence-backed analysis of reactions to the proposed U.S. AI liability framework.Flexible routing and provider audits#
Broad permission without mandatory use
Reusable template · no end-to-end run record attached
Link to this exampleUse the multi-provider-research skill.
You may use Octen, Exa Search and Fetch, relevant Exa Connect providers, one task-appropriate Perplexity mode, Parallel, Firecrawl, TinyFish, and authenticated X interfaces. Use only those that make a material contribution.
Do not repeat successful searches or page reads merely for provider coverage. Report tools used, distinct contributions, and material limitations. Expand query, date-range, and provenance details when they affect interpretation or reproducibility.
Research question: [YOUR QUESTION]Field-driven provider selection
Reusable template · no end-to-end run record attached
Link to this exampleUse the multi-provider-research skill.
Inspect the relevant available research tools and specialized datasets. Choose the minimum sufficient combination that supplies the required fields within my provider and synthesis restrictions. A supplied authoritative document may already satisfy a field; do not add a dataset or another provider when it would only duplicate sufficient evidence.
For each selected provider, identify the provider-backed fields it is expected to supply. If a matched provider is unavailable, preserve the field as an evidence gap instead of silently replacing it with generic web inference.
Research question: Build a sourced profile of [COMPANY], covering corporate identity, people, funding, website traffic, public-company financials if applicable, podcast commentary, product offers, relevant prediction-market data, and X-native discussion.Strict required-provider audit
Reusable template · no end-to-end run record attached
Link to this exampleUse this pattern when the goal includes testing whether specific providers are callable and return evidence.
Use the multi-provider-research skill.
This is a provider-access audit as well as a research task.
Required providers:
- Octen
- Exa Search
- Exa Agent with Particle
- Perplexity Search
- Parallel Search
- Firecrawl Map and selected Scrapes
- TinyFish
- authenticated X data through xurl or one X API integration
Attempt each required provider only for its distinct assigned role. Do not silently replace an unavailable provider.
For every provider, report the callable interface discovered, authentication and availability status, query or operation attempted, successful contribution, errors or limitations, and whether the final answer contains provider-backed evidence.
Research question: What are prominent podcast hosts, guests, journalists, and X users saying about AI regulation?General-purpose fill-in template#
Use this when the short everyday prompt needs more control. Fill in the relevant fields and delete the rest; the skill supplies the shared research rules.
Use the multi-provider-research skill.
Research question: [YOUR QUESTION]
Scope and desired output: [DATES, GEOGRAPHY, COMPARISON CRITERIA, FORMAT]
Providers:
- Required: [PROVIDERS AND THEIR ROLES]
- Optional: [OTHER PROVIDERS ALLOWED WHEN USEFUL]
- Excluded: [PROVIDERS OR OPERATIONS NOT TO USE]
- Is this an exclusive list? [YES / NO]
Native tools: [ALLOW SEARCH AND PAGE READING / PAGE READING ONLY / NEITHER]
Analysis: [SOURCE RETRIEVAL ONLY; THIS ASSISTANT DOES ALL ANALYSIS / PROVIDER-GENERATED ANALYSIS ALLOWED WITH DISCLOSURE]
If a required tool is unavailable: [REPORT THE GAP WITHOUT SUBSTITUTION / ASK BEFORE SUBSTITUTING / USE THIS PERMITTED FALLBACK: ...]
Budget: [SPENDING LIMIT AND ANY CALL, PAGE, OR TIME LIMITS]Native-tool permission does not waive a required provider's role. If you make the provider list exclusive, the native-tools field states any exception. Spending limits need supported controls; text in a prompt is not an enforced provider-side cap. Ask for a fuller provider ledger only when you need an audit.
Portability note#
These prompts are portable at the language level, but provider access is not. Each application may expose integrations through built-in tools, MCP servers, command-line utilities, plugins, extensions, or none at all.
The prompt requires the agent to discover and disclose the tools available in the current session rather than pretend that a named provider is accessible.