An AI Research Agent for Technical Writers

Overview

Technical writers often begin a feature by piecing together information from Jira, Confluence, Google Docs, Slack, and conversations with Engineering. I designed and built an AI research agent that searches approved sources, cites its findings, surfaces conflicting information, and marks missing information as GAP instead of trying to infer an answer. The goal was not to automate writing. It was to give writers a faster, more reliable starting point for research. The original version was built in Gemini Enterprise. Its core research workflow later became a Claude skill that writers now use regularly as part of feature planning. Writers on the documentation team now use it.

My role

I designed the workflow, built the agent and skill, defined its trust model, and created the guidance writers use with it.

• Designed. Mapped the research workflow and created operating modes for the main stages of feature research.
• Built. Implemented the original Gemini Enterprise agent and later the Claude skill version.
• Defined. Established the trust model: approved sources, verification gates, citations, conflict detection, and an explicit GAP rule for missing information.
• Tested. Designed tests for ambiguous feature names, incomplete tickets, unsupported UI text, missing sources, and broken links.
• Enabled. Wrote prompt templates, a quick-reference guide, and usage guidance for the writing team.

Worked with

• Technical writers. The primary users, responsible for verifying findings and producing the final documentation.
• Engineering and Product. Resolve gaps and conflicting source information that the agent cannot answer reliably.

The challenge

Research for a new feature was fragmented across several systems.

A writer might need to search Jira for requirements, Confluence or Google Docs for design context, Slack for decisions, and then speak with Engineering to resolve what was still unclear.

In conversations, writers estimated that this research could take roughly three hours for a feature before drafting even began.

The bigger problem was not only time.

Source material could be incomplete, outdated, or contradictory. An unconstrained AI system can turn missing context into plausible-sounding answers, which is particularly risky in technical documentation.

I wanted the workflow to help writers find information faster without hiding uncertainty.

Key decisions

1. Treat GAP as a valid result. If the available sources do not support an answer, the agent returns GAP rather than attempting to complete the missing information. The gap then becomes a precise question for Engineering or Product. Trade-off: The output may be incomplete, but it is explicit about what is unknown.
2. Verify the research context before synthesizing. The workflow first anchors itself to the relevant feature or ticket and searches only approved sources. This reduces the chance that unrelated information is pulled into the result. Trade-off: A narrower search can miss useful information that exists outside the approved source set.
3. Cite every factual claim. Findings link back to their source so writers can verify them before using the information. If a source cannot be identified, the workflow does not invent one. Trade-off: Writers still need to inspect important sources rather than accepting the output at face value.
4. Separate research from authorship. The agent researches and organizes information. The writer remains responsible for interpretation, engineering conversations, documentation decisions, and final content. Trade-off: It is less automated than a system that drafts everything, but responsibility stays with the writer.

My approach

I started with the way writers already researched features rather than designing an AI workflow first.

1. Observe. Map where writers look for information and where research slows down.
2. Define. Establish rules for sources, uncertainty, citations, and human review.
3. Build. Translate the research process into an AI-assisted workflow.
4. Test. Deliberately introduce ambiguous and incomplete cases.
5. Teach. Give writers simple guidance for interpreting and validating the results.
6. Iterate. Move the most useful research behaviour into a reusable Claude skill.

The solution

The initial agent supported several stages of the documentation research cycle, including planning, concepts, workflows, reference data, alignment checks, drafting support, and review.

Its research workflow connects information from Jira, Confluence, Google Docs, and Slack and produces a structured Technical Brief aligned to the team’s feature-planning questionnaire.

The output separates three things: supported findings, which are backed by a cited source; conflicts, where sources disagree and need resolution; and gaps, which are questions that the available sources cannot answer.

That distinction turned out to be more important than simply producing a summary. Instead of discovering missing information late in documentation review, writers can use the gaps list as an agenda for early conversations with Engineering and Product.

The research core was later moved into a Claude skill, making it easier for writers to use within their normal workflow.

Tools: Gemini Enterprise · Claude · Jira · Confluence · Google Docs · Slack

Impact

• Adoption: Writers on the documentation team use the workflow, and the Claude skill is now part of regular feature-planning work.
• Research speed: Writers previously estimated that researching a feature could take roughly three hours. The agent can produce an initial cited technical brief in under a minute, after which the writer validates the findings and resolves gaps.
• Earlier gap detection: Missing information that might previously have surfaced during drafting or review is now visible at the beginning of the documentation process.
• Better engineering conversations: Instead of asking broad questions such as “Can you explain this feature?”, writers can bring Engineering a specific list of unresolved gaps and conflicting information.
• Trust and traceability: Claims remain connected to their sources, and unsupported information is explicitly marked rather than presented as fact.

Nice work on building the AI tools for the team to increase our productivity.

Colleague

What I learned

A gap is useful information. One of the most important design decisions was allowing the system to say that it does not know. A GAP does more than prevent hallucination. It identifies where the product’s internal knowledge is incomplete and gives the writer a concrete question to resolve.

Research and writing should remain separate capabilities. The original agent gradually expanded from research into drafting and review. That made the workflow harder to understand and maintain. I would now design those as separate skills with clear responsibilities, one for research, another for drafting, and another for review.

The best AI workflow doesn’t remove the writer. The biggest value came from reducing low-value searching while giving writers better information for judgment, questioning, and collaboration. The agent makes research faster. The writer still decides what the documentation should say.

What’s next. Continue separating the workflow into smaller, purpose-built skills and expand the set of approved sources where doing so improves research quality without weakening traceability.