Pipeline Setup Guide

G: Gather and qualify

You have already identified the signals and public sources you want to explore in Step 1. This guide picks up there: use Claude to gather the data, see what is useful, qualify the results and turn the best sources into a recurring workflow.

Start in the same Claude conversation, or paste in the signals and sources you already chose. The prompts below translate the workshop approach into instructions you can adapt. They are suggested wording, rather than a transcript of the original prompts.

The aim is to learn quickly. Do one broad run, see what comes back, cut the losers and double down on the sources that give you useful data.

1. Do a broad first run

Give Claude the sources you identified and ask it to start collecting. Let it write and run the collection code where needed. You want to see what data is actually available before deciding which filters are worth applying.

Keep this first run focused on the signals you chose, with room to discover companies you would otherwise have filtered out. Save your usual company-size, job-title and other profile filters for after you have seen the list.

Let's gather companies from the public sources we identified.

The problem our offer solves: [problem]
Signals we chose: [signals]
Sources to explore: [links or source names]

Do a broad first run across these sources so we can see what
is out there. Don't add my usual company-profile filters yet.
Look for the signals themselves.

Write and run the code needed to collect the available data.
For each result, keep the company or account, what you found,
the source link, any visible date and the original text that
explains why it might be relevant.

Show me the results in a list I can review. Tell me which
sources worked, which were noisy and which you couldn't
access. Keep missing information visible rather than guessing.

What to look at: Are there identifiable companies? Can you open the original post, listing or record? Is there enough detail to understand what the company is doing? A large number of scraped records is a starting point for review.

If a source needs an API key or an integration, ask Claude to explain the current access requirements and setup. Keep credentials out of the exported list and any public page. If a source is blocked, have it explain the gap and continue with the others.

2. Review the list, then add your filters

Open the results and give Claude direct feedback. This is the same back-and-forth as the workshop: this signal is interesting, this one is too broad, skip that one, look more closely at this type of request.

Add filters you normally use after building the list, not before: pain-based segmentation is not ABM.

Start by looking at the need. Then apply the filters that make the list useful for your business. Keep the original collection so you can revisit a filter if it removes too much.

Let's review what came back.

These are useful examples: [records or links]
These aren't useful: [records or links]
What makes the difference: [your reasoning]

Keep companies showing [the need or activity that matters].
Remove [irrelevant results or recurring false positives].
Now apply these additional filters: [your usual filters].

Use this feedback to qualify the rest of the list. Show me
which records you kept, removed or are unsure about, with
short reasons and the supporting source links.

Remove duplicate companies while keeping their different
supporting signals. Keep the original data available.
Don't fill gaps by inventing a need, contact or company detail.

Examples of useful feedback:

• Creator campaigns: keep explicit briefs and roles describing production work. A generic marketing vacancy may tell you very little about the need for your particular tool.
• Practice software: review new registrations and hiring records against the actual offer. A billing vacancy may be less useful for a full practice-management product than it first appears.
• Fundraising: distinguish a founder saying they plan to raise from a company simply appearing in an accelerator directory.

The point is to teach Claude what you see in the data. You do not need to write an elaborate qualification system before looking at the first results.

3. Double down on the useful sources

Compare the sources after reviewing their results. Which ones give you named companies, relevant needs and enough detail to act on? Which mainly give you duplicates, vague matches or records you cannot use?

Data-rich sources make the next step easier: the underlying problem is clearer, the company can be identified and the source gives you context for a useful conversation.

Based on the list and my feedback, compare the sources we ran.

Show me which produced useful companies and why. Separate
raw records from unique companies and qualified results.
Point out overlap, weak matches and missing information.

Recommend which sources to double down on, which to refine
and which to drop for now. Use the actual results rather than
estimated platform totals.

Run a second pass on the promising sources using what we
learned. Keep the original source links and explain whether
any broader search or different query would improve the list.

What to look at: Useful yield matters more than the largest raw count. Also distinguish a source's historical backlog from the new records it produces over time. Your first broad run is not a weekly-volume forecast.

Cut sources that are not helping. You can return to them later if you find a better access route or a more specific signal.

4. Turn the results into a list you can work

Have Claude put the useful records into a simple view. A table or downloadable CSV may be enough. If a small dashboard makes it easier to review the list, ask for one.

Keep the reason for including each company next to the company itself. The list should help you understand who is worth investigating and why, without repeating the entire search.

Turn the qualified results into a clear list I can work with.

Show the company, relevant signal, supporting source link,
date when available and a short explanation of why it fits.
Keep people and business contact details if the sources
already provided them. Leave unknown details blank.

Let me review the list by source and signal. Add a simple
way to record my feedback, and let me export the results.
Use [a table / CSV / simple dashboard] for this version.

Show me the result and run it so I can check that it works.

What to look at: Can you quickly spot good matches, open the evidence and flag mistakes? Keep improving the usefulness of the list. You can add contact enrichment and outreach in their own steps once you know which companies are worth working.

5. Make the winning sources recurring, then iterate

Once you know which sources are worth keeping, ask Claude to turn those collection steps into a recurring workflow. Choose a frequency that suits the source and how quickly you can use the results. In the workshop, the workflow ran every two days; use a cadence that makes sense for your campaign.

Let's make the useful sources recurring.

Keep these sources: [sources]
Use the filters and feedback we settled on above.
Desired frequency: [cadence]
Where I want to receive or review the list: [location]

Set up the collection to run on that schedule and add new
results to the list. Avoid adding the same companies again,
keep new supporting signals and preserve my review decisions.

Explain where it will run and whether my computer needs to
be on. If it needs hosting or an API, explain the setup and
costs before adding them. Keep credentials private.

Run it once and show me how to check the last run, see any
errors and pause it. Confirm the scheduled run works too.

A page being online does not mean its collector is running. Check that a later run actually brings back new data or reports what happened.

Then keep feeding your judgment back into it:

• These companies had the problem we expected. Find more like these.
• These looked relevant from the title, but the description told a different story. Tighten that query.
• This source mostly repeats records we already have. Give it less attention.
• This source is giving us better context. Go deeper there.

Speed matters in the early stages. One rhythm from the approach is to build over the weekend, start using the lists on Monday and iterate on Wednesday. The first version gives you something to learn from; each run and each real conversation should make the next list better.