robomotionio Lead Qualification
The reusable scoring engine. The agent runs intake to synthesize a qualification prompt, enriches each lead with the facts the prompt needs, scores in parallel, and persists the prompt for reuse. Scripts handle prompt persistence + scorecard glue; scoring is the agent's.
When to use
- "Qualify this lead list against our ICP." / "Score these leads."
- Building a reusable ICP qualification prompt; refining one after results.
- The scoring sub-skill inside event/Luma/signal pipelines.
How to run
Mode select
- No prompt referenced → intake: ask structured product/ICP/criteria questions and synthesize a qualification prompt, then save it.
- Prompt referenced → reuse.
- Feedback given → refine the saved prompt.
Save / load / list prompts
python3 ${SKILL_DIR}/scripts/qual_prompt.py --store ${WORKSPACE}/prompts.json \
save --name acme_icp --prompt-text "Qualify a lead as QUALIFIED if ..."
python3 ${SKILL_DIR}/scripts/qual_prompt.py --store ${WORKSPACE}/prompts.json load --name acme_icp
python3 ${SKILL_DIR}/scripts/qual_prompt.py --store ${WORKSPACE}/prompts.json list
Enrich (the data the prompt needs)
python3 ${SKILL_DIR}/scripts/web_research.py --leads ${WORKSPACE}/leads.csv \
--fetch-pages --output ${WORKSPACE}/research.json
Keyless degrade for LinkedIn profile / company facts. With APOLLO_API_KEY or
PHANTOMBUSTER_API_KEY (+ cookie), prefer those for deeper enrichment.
Score (you, the agent)
python3 ${SKILL_DIR}/scripts/qual_prompt.py scaffold --leads ${WORKSPACE}/leads.csv \
--output ${WORKSPACE}/scorecard.json
Fill verdict (qualified/disqualified), confidence, and reasoning per lead against the
prompt. Parallelize scoring; cap concurrency. Reasoning is required per verdict — confidence
without rationale is rejected. Then:
python3 ${SKILL_DIR}/scripts/qual_prompt.py finalize --scorecard ${WORKSPACE}/scorecard.json \
--format csv --output ${WORKSPACE}/scored.csv
Outputs
- A scored list:
[{name, company, linkedin_url, verdict, confidence, reasoning}]. - The saved/updated qualification prompt (in the prompt store).
Credentials / env
- Required: none — scoring runs on the LLM + serp-derived data.
- Optional:
APIFY_API_TOKENorPHANTOMBUSTER_API_KEY+ cookie (LinkedIn enrichment depth);APOLLO_API_KEY(company/person facts);ANTHROPIC_API_KEYonly if the scoring LLM isn't platform-provided.
Notes & edge cases
- Persist prompts so they're reusable (reuse mode) and calibratable (refine mode).
- Parallelize scoring for throughput; cap concurrency to respect rate limits.
- Reasoning must be explicit per verdict.
- Without LinkedIn enrichment keys, score on serp-derived facts and lower confidence; note it.