The signals competitors can't see

Why generic ICP filters fail

Your ICP doc says "Series B SaaS, 100-500 employees." So does everyone else's. There is no differentiation in the signals you score on.

The question that matters: what is actually different between your closed-won accounts and your closed-lost accounts? Not what you think. What the data shows.

How to discover your niche signals

Step 1: Export two lists

Pull from Salesforce or HubSpot:

• closed-won.csv (50 accounts that closed)
• closed-lost.csv (50 accounts that did not)

Step 2: Run the niche signal discovery skill

The skill authors one durable Play that enriches both cohorts, preserves evidence, and compares the same frozen contract across won and lost accounts:

deepline plays search "won lost signal discovery firmographics technographics hiring" --json
# Claude Code authors signal-discovery.play.ts with one frozen contract for both cohorts.
deepline plays check signal-discovery.play.ts
deepline plays run --file signal-discovery.play.ts --input @pilot-cohorts.json --watch

Step 3: Review the signals

Claude analyzes both cohorts and surfaces the 3-5 attributes that statistically separate winners from losers.

Step 4: Deploy as a scoring model

The discovered signals become a webhook-bound scoring Play. Its definePlay options declare and authenticate the trigger:

{
  webhook: {
    auth: {
      type: 'standard-webhooks',
      headerFamily: 'standard',
      signingSecrets: ['INBOUND_RELAY_WEBHOOK_SECRET'],
      toleranceSeconds: 300,
    },
  },
}

Use an authenticated relay that verifies HubSpot's v3 signature, then signs the forwarded body with Standard Webhooks v1 headers. Reuse the same webhook-id when retrying the same event. Then store the Play-scoped signing secret, publish, and verify the Play:

deepline secrets set INBOUND_RELAY_WEBHOOK_SECRET --scope play --play icp-scoring
deepline plays check icp-scoring.play.ts
deepline plays publish icp-scoring.play.ts
deepline plays describe icp-scoring --json
deepline plays versions --name icp-scoring --json

Sample analysis: Cybersecurity ICP

A sample / illustrative fraud prevention company ran niche signal discovery on 84 closed-won and 85 closed-lost accounts. Here's what the data showed.

Positive-fit signals (predict closed-won)

Anti-fit signals (predict closed-lost)

The scoring model

Signals get weighted into a lead score:

From signals to outreach

Once you have the signals, Deepline finds the companies and contacts that match.

Step 1: Resolve the stage graph

deepline plays search "company to contacts" --json
deepline plays describe prebuilt/company-to-contact --json
deepline plays describe prebuilt/name-and-domain-to-email-waterfall --json
deepline plays bootstrap company-people-email --from csv:tam-companies.csv \
  --people play:prebuilt/company-to-contact \
  --email play:prebuilt/name-and-domain-to-email-waterfall \
  --limit 500 --out signal-outreach.play.ts
# Complete every generated TODO mapping, select the people output rows field,
# and remove the scaffold's guard throws before checking or running it.

Step 2: Check and pilot the owned Play

deepline plays check signal-outreach.play.ts
deepline plays run --file signal-outreach.play.ts --limit 3 --watch

Step 3: Run the accepted route

deepline plays run --file signal-outreach.play.ts --limit 500 --watch

What gets enriched

Cost breakdown

From raw CRM export to scored outreach list for under $40.

Who uses this

• Founders who want to stop guessing which accounts to prioritize
• RevOps leads building scoring models that predict revenue
• AEs who want to know which accounts are worth their time

Deploy ongoing scoring

Once you have the signals, deploy them as an always-on workflow:

Claude Code adds the schedule to the same definePlay options before republishing:

{
  cron: { schedule: '0 8 * * *', timezone: 'America/New_York' },
}

deepline plays check icp-scoring.play.ts
deepline plays publish icp-scoring.play.ts
deepline plays describe icp-scoring --json
deepline plays versions --name icp-scoring --json

New leads get scored, enriched, and sequenced automatically. Signals stay tied to real outcomes, not assumptions.

Common questions