configure-scoring

Configure Scoring

Use this skill once per ICP / use case to wire up native scoring. After this runs, every assigned company or contact will have fit and urgency scores that auto-recompute when signals fire.

For concept definitions (dimensions, profiles, rules, assignments, point-value shapes, auto-trigger behaviour), see _shared/scoring.md.

Saber CLI check

Run saber --help. If the CLI isn't installed, scoring requires it — point the user to saber.app and stop. The platform doesn't expose a non-CLI fallback for scoring configuration.

Step 1 — Choose object type

Scoring is typed: a profile scores either companies (keyed by domain) or contacts (keyed by LinkedIn profile URL). You'll need a separate profile per type.

Ask:

"Are we scoring companies, contacts, or both? Each needs its own profile."

For the rest of the flow, do one type end-to-end before starting the next.

Step 2 — Create the profile

saber scoring profile create --type company --name "<descriptive name>" \
  [--description "<optional context>"]

Naming guidance: include the use case (e.g. New business — EMEA mid-market, Expansion — existing customers, Enterprise outbound). Profile type cannot be changed later; a typo means deleting and recreating.

Capture the returned profileId for the next steps.

Step 3 — Gather inputs

Before writing rules, collect:

1. Signal templates already in Saber. Rules are keyed on signal template IDs — without templates there's nothing to point at. Three states to handle:
   • No templates and no signals → route to signal-discovery then create-company-signals / create-contact-signals first.
   • Templates exist → list them and decide which to include in the profile.
   • Lots of historical ad-hoc signal data, few templates → route to extract-signal-templates first to consolidate ad-hoc executions into reusable templates. Without this, historical signal data won't contribute to scores.
2. The weighted model from generate-signals (categories + weights), or whatever the user has informally.

Step 4 — Translate signals to rules

For each signal template in scope, decide:

1. Dimension — fit or urgency. Default mapping from generate-signals categories:

   generate-signals category	Default dimension
   icp_fit	fit
   buying_signal	urgency
   urgency	urgency

   Override if the user has a strong opinion. Disqualifiers (e.g. "is a competitor") usually go in fit with strongly negative false points.

2. Answer type — must match the signal template's answer type exactly. Pulling the wrong shape returns 422 INVALID_POINT_VALUES.

3. Point values — translate weight (1–3) into points. Default scaling is points = weight × 10:

   Answer type	CLI
   boolean	--true 20 --false 0 (disqualifier: --false -100)
   number / percentage / currency	--range "0:500:5" --range "500:5000:15" (max exclusive, repeatable)
   list	--choice "Salesforce:10" --choice "HubSpot:8" (repeatable)

Then upsert each rule:

saber scoring rule upsert <profileId> \
  --signal-template <signalTemplateId> \
  --dimension fit \
  --answer-type boolean \
  --true 20 --false 0

For complex shapes you can pass JSON instead:

saber scoring rule upsert <profileId> \
  --signal-template <id> --dimension urgency --answer-type number \
  --points-file ranges.json

Repeat per signal template. Show the user the running list of rules as you go and confirm before assignments.

Step 5 — Verify rules with one object

Pick one company / contact you can sanity-check (an existing customer is ideal). Assign the profile to just that one object first:

saber scoring assignment create --profile <profileId> --type company --object <domain>

The assignment triggers compute. Read the result with contributions:

saber scoring scores --type company --object <domain> --detailed

Walk through with the user: are the contributing signals what you expected? Any rule firing the wrong way? If a rule is wrong, rule upsert again — every upsert recomputes.

Step 6 — Bulk-assign the full list

Once the rules look right, bulk-assign the rest of the target list:

saber list company companies <listId>
# extract domains, then:
saber scoring assignment bulk --profile <profileId> --type company \
  --object <domain1> --object <domain2> ...

For contact profiles, use LinkedIn profile URLs as object IDs:

saber scoring assignment bulk --profile <profileId> --type contact \
  --object https://linkedin.com/in/jane --object https://linkedin.com/in/bob

Bulk assign is idempotent — duplicates are skipped silently.

Step 7 — Confirm and hand off

Show a summary:

• Profile name + id
• Rule count by dimension
• Assigned object count

Tell the user:

• Scores are live. Read any score with saber scoring scores --type ... --object ....
• Auto-trigger is on. Scores recompute automatically when underlying signals complete; you don't need to call compute manually.
• Next steps:
  • score-accounts — rank the assigned list by current fit + urgency
  • manage-scoring — tune rules, add or remove assignments later
  • qualify-inbound — qualify a new inbound lead against this profile

Troubleshooting

• 422 INVALID_POINT_VALUES — the --answer-type doesn't match the shape you provided. Fix the shape and retry the upsert.
• Score is 0 with no contributions — no signal data exists yet for that object, or no rules cover the signals it has. Check signalCoverage vs. totalRules from --detailed output.
• Wrong type — profile type is immutable. Delete and recreate with the correct type.