What Claude contributes
Claude is Anthropic’s assistant and model API. In this enrichment pilot, its useful job is narrow: extract candidate company facts from supplied source text and draft an opener after validation. It is not the source of a company’s headcount, and a fluent answer is not evidence.
Choose Claude for interpreting unstructured text against explicit instructions. Use ordinary code for fixed scoring rules, date checks and exact comparisons. If your CRM already holds the required structured facts, adding a model to copy them creates another failure point without resolving a data gap.
Select the right execution path
A Claude Code skill packages reusable instructions and supporting files. An API structured output constrains the response schema. These solve different setup problems; neither proves that a returned fact is true.
For a manual task, review extracted facts beside their sources. For repeated API processing, validate the output contract, retain evidence and record the model and prompt version. For Clay, configure Use AI and its model account separately. Installing a local skill does not configure a Clay table.
Subscription and API costs
Claude pricing lists Pro at $20/month or $200/year, Max at $100/month and Ultra at $200/month, alongside Free and team plans. Usage limits apply.
Eligible Max, Ultra and Team plans include monthly API credits. Claiming requires an active eligible plan for 7 days and linking a Console organization. Credits expire each billing cycle and do not cover interactive Claude Code. Check eligibility and the linked balance before treating these credits as available budget.
For API work, estimate uncached input tokens times the input rate plus output tokens times the output rate, using the selected model’s current rates. Add applicable tool, caching and retry costs. Divide actual total processing and review cost by accepted records. A subscription price alone cannot price an automated enrichment run.
Try the downloadable example first
The Clay and Claude workflow includes 20 synthetic records, extraction and opener prompts, a schema and a local validator. Run python3 pipeline.py demo from the extracted bundle. This uses supplied fixtures and deterministic code: it calls neither Claude nor Clay and incurs no API spend.
For the optional Claude Code path, retain the bundle’s supporting files under .claude/skills/lead-enrichment-clay-claude/, with SKILL.md at that directory’s root. Follow the workflow’s setup before running a live model. For an API or Clay adaptation, map model output to the documented JSONL fields rather than assuming the skill directory is an importable integration.
What must still be checked
- Schema validity is not factual validity. Require source IDs and exact quotes. Check missing, stale and conflicting evidence before scoring; never ask the model to fill a gap with a plausible estimate.
- Instructions do not replace execution controls. Keep CRM writes and sequence enrollment outside the extraction step. Review any draft before using it in outreach.
- One successful answer is not an evaluation. Keep a fixed test set containing missing and contradictory inputs. Compare rejection behavior, correction time and cost before changing prompts or models.
Use the enrichment evaluation guide for that comparison. Apollo supplies prospecting data and outreach functions; it fills a different role from a text model. Read Clay vs Apollo to choose the surrounding process. Keep your existing assistant if it meets the same evidence checks at a lower measured cost.
Evidence status: official documentation was checked on November 1, 2026. No live Claude extraction benchmark or model comparison was performed for this profile. The offline synthetic demo tests the local rules, not model accuracy.