2. Setup: Your Knowledge Base

Steps 1 and 2. Building the APG knowledge base that everything else reads from. Everyone starts here.

Why this comes first

You are producing two descriptions: the kind of company that buys from you, and the kind of person inside it who decides. Every later step reads them. Lightfield uses the first to find and rank companies, the second to pick who to contact, and both to write the copy in the sequences.

So if these are wrong, everything downstream is wrong - the wrong companies, ranked badly, emailed by the wrong person about the wrong thing - and you cannot fix it in the copy. When something needs to change later, you change it here, once, and everything built afterward picks it up.

Start with a dedicated APG knowledge base

Lightfield already has a general knowledge base that the whole CRM runs on. APG gets its own, separate from that. You do not set it up by hand - running Build ICP below creates a dedicated APG folder and writes the files into it.

Keep it separate on purpose. The APG knowledge base holds your outbound context: who you sell to, who buys, what disqualifies them, and what signals to act on. That context changes far more often than the general knowledge the rest of your CRM depends on, and keeping it in its own folder means you can update your targeting and messaging without disturbing everything else. It is your sequences-specific knowledge base, and the skill reads from it every time it sources companies, scores them, picks contacts, and writes copy.

Do you have enough closed-deal history?

Build ICP reads the won and lost deals in your CRM. A hundred closed deals give it plenty to reason from. Five do not, and it will hand you a confident description built on almost nothing.

Every Lightfield customer has the CRM. What varies is how much win-loss history sits in it. A founder running outbound to land the first customers has a nearly empty one, and that is the common case, not the exception. Be honest about which you are, and do not mistake a thin-history output for fact.

If your history is thin, do not run a bare build my ICP and trust it. Instead:

• Feed it everything with real signal: the deals you closed, the ones you lost and why, named companies you know are a perfect fit even if they never entered your pipeline, plus any positioning or research docs.

• Tell it you are early and want a hypothesis, not a summary. Ask it to reason from the best-fit customers you named and from who should buy, rather than averaging a handful of deals.

• Treat the first version as a draft you correct hard after your first 50 sends. Cold outbound is partly how you find your real ICP.

Step 1: Build ICP

Type build my ICP.

What it does

Lightfield reads your CRM instead of asking you to fill in a form. It pulls the deals you won, lost, and churned, joins them to what it knows about each company, and reads the actual history of 20 to 30 lost deals to work out why they really died. That last part matters: the "reason lost" dropdown is half empty and half wrong, so Lightfield reads the emails and call notes instead.

You can also hand it other material: positioning docs, win-loss research, a board deck. Anything with real information in it.

What it produces

When it finishes

Read the files properly, not a skim. Twenty minutes here saves weeks later. Lightfield also flags the analysis it could not finish because your CRM data was patchy, and suggests fixes: recording loss reasons, filling the industry field, adding a competitor-lost-to field. Each one makes the next refresh better.

Pay particular attention to anti-icp.md. Who not to sell to, and which buyer you were talking to when you lost, is usually where a team learns something new, and it protects your reply rate more than any rewriting.

Refresh cadence

Re-run quarterly, or whenever you add roughly ten closed-won deals. A stale ICP quietly degrades everything built on it.

Step 2: Build your scoring rubric

Type build my scoring rubric.

What it does

Turns the ICP into a 0 to 100 score for any account, plus a signal catalog. It reads the step 1 files, then asks two questions. Answer both properly - this is where your judgment enters the system, and a shrug produces a rubric that cannot tell companies apart.

Question 1: strategic factors beyond fit. Logo prestige, geography, partner-ecosystem fit, public-company status, customer-of-customer signals.

Question 2: fit versus timing. How much fresh trigger events count against baseline fit. 70/30 leans toward accounts that look right, 50/50 toward accounts that look ready. Unsure, start at 70/30 and weight timing more once you trust your signals.

What it produces

scoring-rubric.md holds the dimensions across three pillars - ICP fit, outbound timing, strategic - weights summing to 100, bands at A 80+, B 60 to 79, C 40 to 59, D below 40.

signals.md holds the signal catalog: labels attached to accounts (recently funded, hiring spike, leadership change) with the buyer pain each implies. That pain mapping tells your copy what to lead with.

How scoring works

Scoring is additive. No penalties, no disqualifiers - a weak account earns low marks rather than getting knocked down. So anti-icp.md shapes the low end of the bands rather than filtering; clearly wrong accounts get screened out earlier, during sourcing.

Changing things later

When positioning shifts, or you learn something new about who buys, change it here, in the knowledge base. Then re-score, and let new sequences inherit it.

Two ways to change it: edit a description directly, or hand the agent new evidence and have it fold it in. When you land a strong result or a new use case, point the agent at the calls or paste it in and say enrich the APG articles with this. It updates the descriptions in place, and your next find-accounts and next sequence read the new version. You are adding to the root, not rebuilding it.

Do not hand-edit account scores or patch one sequence to work around a stale knowledge base. Patch downstream and the root drifts while every future sequence inherits the stale version.

Checklist

• ICP built

• All five files actually read, anti-icp.md especially

• CRM hygiene gaps the agent flagged written down where you will see them

• Scoring rubric built, with real answers to both questions

• Fit versus timing split chosen deliberately

• Signals catalog reviewed, and the pain mapping makes sense to you

Where to go next

The Setup: Your Target List article picks up at step 3. Running warm plays first, which most teams should? Go straight to the Play Library and come back for the target list when you are ready for cold. Several plays run entirely off the two docs you just built.