Skip to content
Get packs

Guide · Agent skill

How to use the Data Research Pipeline agent skill

Updated · written with AI from the project's own documentation and checked against our catalogue data

The data-research skill teaches an AI agent to run a repeatable pipeline that turns scattered information, mostly email such as investor updates, receipts and company metrics, into a markdown table. It is built for the GBrain knowledge base, so it fits people who want a source-linked tracker they can keep updating. After this guide you will know what the skill does, how to install it, how to trigger it with plain requests, and what to check before trusting its output.

Data Research Pipelinegarrytan/gbrain · MIT

Quick start

  1. 1Download and unzip. Download the ZIP from this page and unzip it into ~/.claude/skills/ or a project's .claude/skills/ folder. Alternatively, upload the ZIP under Settings, Capabilities, Skills in the Claude apps.
  2. 2Set up GBrain and recipes. Make sure your GBrain knowledge base exposes the required tools and that recipe YAML files are in ~/.gbrain/recipes/.
  3. 3Connect email access. Configure email access through your credential gateway. The skill file does not specify credentials, so this is set up outside the skill.
  4. 4Make a natural request. Ask for what you want tracked, using a trigger word such as track, research or investor updates. Name the data you care about, such as MRR or receipts.
  5. 5Review the tracker and sources. Open the tracker page and check a sample of rows against their linked raw sources before relying on any totals.

What is an agent skill, and what does data-research make the agent do?

An agent skill is a folder with a SKILL.md file that gives an AI agent a set of instructions for a specific task. When a request matches the skill's triggers, the agent follows its steps instead of improvising.

The data-research skill runs a seven-phase pipeline. The agent defines a research recipe, searches your brain first, then email, web and APIs. It saves each raw source before extracting fields, classifies items, removes duplicates, and adds each new row to a canonical tracker page with a link back to the source.

  • Searches the GBrain knowledge base before going to outside sources
  • Saves raw email bodies, API responses and attachments before any extraction
  • Uses regex patterns first and an LLM fallback for field extraction, logging every fallback
  • Flags possible duplicates and notes possible corrections instead of silently overwriting
  • Backlinks people and companies to their entity pages

Install the data-research skill

Download the ZIP from this page. The skill ships as a single file, SKILL.md, inside the ZIP. Install it in one of these ways:

  • Personal install: unzip the folder into ~/.claude/skills/ so the skill is available across your projects.
  • Project install: unzip the folder into the .claude/skills/ directory of a single project.
  • Claude apps: open Settings, go to Capabilities, then Skills, and upload the ZIP.

Set up the prerequisites: GBrain, recipes and email access

The skill requires a GBrain knowledge base with tools such as put_page, query, put_raw_data and file_upload. Email is read through a credential gateway; the skill file does not specify the credentials, so set up that access in your own environment. Web search and API access are optional extra sources.

Recipes are YAML files that define search queries, classification rules, the extraction schema and the tracker page path. They live in ~/.gbrain/recipes/. Three built-in recipes are named investor-updates, expense-tracker and company-updates.

  • GBrain knowledge base with the listed tools available to the agent
  • Email access through a credential gateway
  • Optional web search and API access
  • YAML recipe files at ~/.gbrain/recipes/

How do you trigger the skill with a natural request?

You do not need a special command. Ask in plain language using words the skill listens for, such as research, track, extract from email, investor updates, donations, build a tracker or data dig. The agent infers the target from the conversation and recent brain activity, and asks for clarification only when the context is not enough.

If you want recurring collection, say so in the request. The skill notes that it supports a cron recipe for scheduled runs.

  • "Track investor updates from my email this quarter and build a tracker for MRR, runway and headcount."
  • "Extract amounts and platforms from my receipt emails for the last few months and build an expense tracker."
  • "Research our portfolio company update emails and add revenue and user numbers to the tracker."

What happens during a run?

The pipeline is strict about raw sources. Every tracker entry must link back to a saved source, and summaries of batch results must be re-read from saved files rather than from the agent's memory.

  • Phase 1 defines the recipe. Phase 2 searches brain, then email in windows (quarterly, or monthly if results are truncated), then web and APIs.
  • Phase 3 classifies items and skips marketing, newsletters and noise.
  • Phase 4 saves the raw source first, then extracts fields.
  • Phase 5 archives raw sources with put_raw_data or file_upload.
  • Phase 6 deduplicates: exact matches are skipped, fuzzy matches on entity, date and similar amount are flagged, and a different amount for the same entity and date is added with a note.
  • Phase 7 appends rows to the tracker, computes running totals and backlinks entities.

Safety notes: scripts, network and writes

The skill does not run scripts, but it uses the network. It is marked as mutating, meaning it writes data. It writes to your brain and reads email through a credential gateway, so review what the agent is allowed to read before you run it on private or sensitive email.

Check extracted numbers against the saved raw sources before using them for financial or legal decisions. The skill itself warns that LLM-only extraction can be wrong, and its integrity rule exists because batch-processed amounts were wrong when taken from working memory.

  • Runs scripts: no
  • Uses network: yes
  • Writes to the GBrain knowledge base: yes
  • Reads email through a credential gateway: yes

What people use it for

  • Tracking MRR, ARR, burn, runway and headcount from investor update emails over time
  • Building a personal expense or subscription ledger from receipt emails with amounts and platforms
  • Maintaining a source-linked tracker of portfolio company metrics inside GBrain
  • Collecting donation records from email into a structured table for year-end review

Frequently asked questions

How do I install the data-research agent skill?

Download the ZIP from this page and unzip it into ~/.claude/skills/ for personal use or a project's .claude/skills/ folder for one project. In the Claude apps you can instead upload the ZIP under Settings, Capabilities, Skills.

Is the data-research skill free to use?

The skill is listed under the MIT licence, which is stated on this page. Check the licence file in the source repository for the full terms. You still need your own GBrain knowledge base and email access to run it.

What do I need to run the data-research skill?

You need a GBrain knowledge base with tools such as put_page, query, put_raw_data and file_upload, plus email access through a credential gateway. Web search, API access and YAML recipe files in ~/.gbrain/recipes/ are used as needed.

Does the data-research skill run scripts or access the internet?

The skill does not run scripts, but it does use the network. It reads email through a credential gateway and writes to your GBrain knowledge base, so review what it can access first.

Why does the skill save raw sources before extracting data?

Saving the raw source first means every tracker row can link back to the original email or document. The skill's integrity rule requires re-reading saved files when summarising batches, because LLM working memory produced wrong amounts in a known bug.

How does the skill avoid double-counting entries?

Before adding a row, it checks for exact matches, which are skipped, and fuzzy matches on entity, date and similar amount, which are flagged for review. A different amount for the same entity and date is added with a note in case it is a correction.

New agent skills, repos and AI tools — every week

The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.

Our verdict

Use data-research if you are a founder, operator or team that wants a repeatable, source-linked tracker built from email inside GBrain, and you will check the saved sources. Pick something lighter if you only need a one-off question answered or a quick web lookup, since the seven-phase pipeline is heavy and writes data.

Download the skill