Building a simple research workflow
Combine a search skill, a reasoning model, and a summary skill into a working research pipeline.
A research workflow takes a question and returns a structured, reliable answer — not just a web search, but a process that finds information, evaluates it, and synthesises it into something usable. Here's how to build one from scratch using harness-compatible components.
What we're building
A three-stage pipeline:
1. Find — A search skill pulls relevant sources from the web or a specific data source
2. Filter — A reasoning model reads the sources and identifies what's actually relevant to the question
3. Summarise — A summary skill condenses the relevant material into a clear, structured output
This is a pattern. Once you understand it, you can adapt it to almost any research task.
Stage 1: The search skill
Search skills vary significantly in quality. For a research workflow, you want one that:
• Returns sources, not just snippets
• Includes enough context to evaluate relevance
• Has a trust score of 80+ (search skills with lower scores often return unreliable or outdated results)
• Supports filtering by date or source type if your use case requires fresh or domain-specific information
Connect the skill so its output (a list of sources with snippets) flows directly into Stage 2.
Stage 2: The reasoning model
The reasoning model's job is to read what the search returned and decide what's actually relevant to your question. Configure it with a clear system prompt:
*'You will receive a list of web sources and excerpts. Your job is to identify which ones directly address [your question]. Return only the relevant ones, with a one-sentence explanation of why each is relevant.'*
This step prevents noise from propagating to the summary. A search returns many things; you only want the right things to continue.
Stage 3: The summary skill
Pass the filtered, relevant sources to a summary skill. Configure the output format based on how you'll use the result:
• Bullet points — for quick scanning
• Structured sections — for documents or reports
• Single paragraph — for inserting into a larger piece
Always include a source list in the output. A summary without sources is an opinion. A summary with sources is a research output.
Testing and tuning
Run the workflow on three test questions before using it for real work:
1. A question you already know the answer to (validates accuracy)
2. A question about something very recent (validates freshness)
3. A question with no good answer (validates how it handles uncertainty)
If it fails on any of these, adjust the prompt in Stage 2 first. The filtering stage is where most quality problems originate.
Search → filter → summarise. The filtering step is the one most people skip, and it's the one that separates useful research outputs from noise wrapped in a summary.