polym-sdr-leads-qualification

name	polym-sdr-leads-qualification
description	Score and qualify Trigify leads using LLM-powered evidence extraction and deterministic routing. Supports Anthropic, OpenAI, Gemini, BytePlus Ark, and any LiteLLM-compatible provider.

SDR Leads Qualification

Score Trigify lead exports with a single LLM call per lead: evidence extraction, sub-score assignment, and deterministic Python routing into qualified / review / disqualified queues.

Runbook: docs/RUNBOOK.md · Examples: docs/EXAMPLES.md

---

Invocation Protocol

When this skill is triggered, follow the steps below in order. Do not skip steps.

Step 1 — Locate the skill directory

The skill lives at ~/.claude/skills/polym-sdr-leads-qualification/ when installed, or at the repo path if running from source. Resolve the correct path before proceeding.

SKILL_DIR=~/.claude/skills/polym-sdr-leads-qualification
# fallback for repo-local runs:
# SKILL_DIR=<repo-root>/skills/polym-sdr-leads-qualification

Step 2 — Ask for the CSV file

If the user has not already provided a CSV path, ask:

"What is the path to your Trigify CSV file?"

Step 3 — Campaign selection

List available campaigns by scanning $SKILL_DIR/campaigns/ for subdirectories that contain a config.json file (this naturally excludes common/ and any other utility directories). Then ask the user:

"Which campaign should I use to score these leads?

Available campaigns:

1. seedance — ByteDance Seedance AI video (default)
2. kling — Kuaishou Kling AI video [… any other detected campaign folders …]
3. Create a new campaign

Press Enter to use seedance, or type a number / campaign name."

• If the user picks an existing campaign → skip to Step 5.
• If the user picks "Create a new campaign" → proceed to Step 4.

Step 4 — New campaign creation

4a. Gather minimal inputs

Ask the user these questions (can be in one message):

1. Product name  (e.g. "Runway", "Adobe Firefly", "HeyGen")
2. One-line tagline  (e.g. "AI video creation for creative teams")
3. What does the product do?  (2–4 sentences — key capabilities, differentiators)
4. Target customer types  (which industries / company types are ideal buyers?)
5. Target buyer personas  (job titles most likely to buy — list them)
6. What pain does the product solve?  (speed, cost, scale, quality — be specific)
7. Hard disqualifiers  (industries or roles that should NEVER qualify)
8. Campaign ID (slug, lowercase, no spaces — e.g. "runway", "heygen")
   Leave blank to auto-generate from product name.

Tell the user: "You don't need to be exhaustive — I'll fill in the details using the Seedance campaign as a structural template."

4b. Generate campaign files

Using the user's answers plus the seedance campaign files as structural references, generate all six files. Write them to $SKILL_DIR/campaigns/<campaign-id>/.

config.json

{
    "name": "<Product Name>",
    "tagline": "<tagline>",
    "score_weights": {
        "segment_fit": 20,
        "buyer_fit": 20,
        "pain_fit": 25,
        "workflow_fit": 20,
        "commercial_fit": 15
    },
    "qualified_threshold": 70,
    "disqualified_threshold": 60
}

product.md — Describe the product: key capabilities, differentiators, primary use cases by segment. Model the structure on campaigns/seedance/product.md.

icp.md — Define the ICP: primary target verticals (table with segment ID, description, why-fit), target personas (Tier 1 decision makers, Tier 2 influencers, Tier 3 low priority), anti-ICP list, segment → use case mapping. Model on campaigns/seedance/icp.md. Derive segment IDs and verticals from the user's stated target customer types — adapt, do not copy Seedance segments verbatim.

signals.md — Define what constitutes a high-intent, medium-intent signal for this product. A POST about the product's problem space = high intent. A LIKE = low signal. List product-specific buying signals. Model on campaigns/seedance/signals.md.

qualification.md — Full 100-point scoring rubric using the same five sub-scores (segment_fit 20, buyer_fit 20, pain_fit 25, workflow_fit 20, commercial_fit 15). Adapt score bands to match the product's ICP. Model on campaigns/seedance/qualification.md.

disqualification.md — Hard disqualifiers: wrong industry, junior roles, irrelevant functions, confirmed competitors. Pull from the user's stated hard disqualifiers plus infer obvious ones from the product description. Model on campaigns/seedance/disqualification.md. Follow the same IMPORTANT rule: only confirmed negatives belong in hard_disqualifiers; uncertain cases go in unknowns.

4c. Show a confirmation summary

After generating the files, show the user a structured summary:

Campaign: <id>
Product: <name> — <tagline>

Target segments: <list the top 4-5 segment IDs>
Tier 1 buyers: <list top 3-4 job titles>
Hard disqualifiers: <list top 3-4>
Scoring thresholds: qualified ≥ 70 | review 60–69 | disqualified < 60

Files written:
  campaigns/<id>/config.json
  campaigns/<id>/product.md
  campaigns/<id>/icp.md
  campaigns/<id>/signals.md
  campaigns/<id>/qualification.md
  campaigns/<id>/disqualification.md

Then ask:

"Does this look right? Type yes to score the leads with this campaign, edit to change anything, or cancel to stop."

• If "yes" → proceed to Step 5 with --campaign <id>.
• If "edit" → ask which file or field to change, apply the edit, re-show the summary.
• If "cancel" → stop.

Step 5 — Confirm env vars

Check that LLM_API_KEY and LLM_MODEL are set. If not, prompt:

"Please set LLM_API_KEY and LLM_MODEL (and optionally LLM_PROVIDER / LLM_API_BASE) before scoring."

Step 6 — Run the scorer

python3 $SKILL_DIR/scripts/score_csv.py \
  --input <csv_path> \
  --output-dir ./output \
  --campaign <campaign_id>

Add --limit N if the user asked to process only a subset. Add --disable-web-enrichment if they asked to skip website fetches.

After the run, report:

• Total leads processed
• Qualified / Review / Disqualified counts from output/summary.json
• Path to output directory

---

Pre-flight Check

Set in .env or as environment variables:

Variable	Required	Example
LLM_API_KEY	Yes	sk-ant-... / sk-... / AIza...
LLM_MODEL	Yes	claude-sonnet-4-6, gpt-4o, gemini-2.0-flash, ep-xxxx
LLM_PROVIDER	No	anthropic, openai, gemini — auto-detected if omitted
LLM_API_BASE	No	BytePlus Ark URL or any custom endpoint

Quick Start (manual CLI)

# Anthropic
LLM_API_KEY=sk-ant-... LLM_MODEL=claude-sonnet-4-6 LLM_PROVIDER=anthropic \
python3 ~/.claude/skills/polym-sdr-leads-qualification/scripts/score_csv.py \
  --input leads.csv --output-dir ./output --campaign seedance

# BytePlus Ark
LLM_API_KEY=<ark-key> LLM_MODEL=ep-xxxx LLM_PROVIDER=openai \
LLM_API_BASE=https://ark.ap-southeast.bytepluses.com/api/v3 \
python3 ~/.claude/skills/polym-sdr-leads-qualification/scripts/score_csv.py \
  --input leads.csv --output-dir ./output --campaign seedance

Parameter Reference

Flag	Description	Default
--input	Path to Trigify CSV	Required
--output-dir	Output directory	Required
--campaign	Campaign pack: seedance, kling, or any custom id	seedance
--limit N	Process only first N rows	all
--disable-web-enrichment	Skip website fetch per lead	off
--sleep-ms N	Pause between leads (rate limit control)	0
--provider	Override LLM_PROVIDER	env
--api-base	Override LLM_API_BASE	env
--model	Override LLM_MODEL	env
--api-key	Override LLM_API_KEY	env

Output Files

output/
  all_scored_leads.csv
  qualified_leads.csv
  review_leads.csv
  disqualified_leads.csv
  summary.json
  artifacts/{lead_id}_input.json
  artifacts/{lead_id}_enrichment.json
  artifacts/{lead_id}_evidence.json
  artifacts/{lead_id}_decision.json

Available Campaigns

Campaign	Product	Qualified ≥	Review band
seedance	ByteDance Seedance (AI video)	70	60–69
kling	Kuaishou Kling (AI video)	70	60–69

Custom campaigns created via the interactive flow are stored in campaigns/<id>/ and are immediately available via --campaign <id>.

Scoring Logic

Each lead goes through: normalize → parallel web enrichment → single LLM call (evidence + 5 sub-scores) → Python clamping + routing. Hard disqualifiers and signal_strength=low override the score regardless of total. See docs/RUNBOOK.md for the full decision protocol.

When To Invoke

• Process or inspect a Trigify lead export
• Score a single company or person for campaign fit
• Explain why a lead was routed to review or disqualified
• Work with or extend campaign prompts under campaigns/
• Create a new campaign for a product not yet in the library
• Prepare outreach output from scored leads