agentleFS
ghazanfar111212ghazanfar111212/claude-skills-performance-marketing/skills/clay-enrichment-architect/SKILL.md

clay-enrichment-architect

Skill0 starsChanged 38 days agoMITView on GitHub

Designs Clay tables that turn a raw prospect list into personalized outreach at volume, column architecture, waterfall enrichment, AI/Claygent columns, scraping a prospect's site for signals, conditional run rules, credit control, and the export into HeyReach/Instantly/Smartlead. Use whenever an outreach plan depends on a per-lead fact that no prospecting filter returns (a client name, tech stack, portfolio detail, recent post, pricing page, job posting), or when the user says "Clay table", "enrich this list", "Claygent", "waterfall enrichment", "scrape their site", "personalize at scale", "find their clients", "my personalization doesn't scale", "how do I get [attribute] for 500 leads". Also use to audit an existing Clay table that is burning credits or returning junk. For choosing WHO to target and the search filters, use icp-filter-architect. For the sequence and message copy, use linkedin-outreach-architect or cold-email-outreach.

The file

We split this skill into the rules it gives an agent

6 rules. About 0 are common, 6 are distinctive, and 0 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈6

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 0

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    5. Test on 20 rows before running 2,000: Read the actual output, not the success rate.

    line 51
  2. Distinctive:

    6. Score and filter: Add a qualification column that decides Tier A (full personalization) vs Tier B (category-level fallback) vs drop.

    line 56
  3. Distinctive:

    Ask for one fact, not a paragraph.

    line 85
  4. Distinctive:

    Put the cheapest disqualifier as early as possible so bad rows exit before the expensive columns.

    line 102
  5. Distinctive:

    Use waterfalls for contact data (provider 1 → only if empty, provider 2), never parallel providers on every row.

    line 104
  6. Distinctive:

    No fallback opener for rows where enrichment came back empty.

    line 114

≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Clay Enrichment Architect

Clay is not a data source. It is an orchestration layer that takes a seed list and adds columns other tools cannot filter on. Its whole value is closing the gap between the ICP attribute you want and the attribute a prospecting tool can return.

The rule this skill exists to enforce

Never write outreach copy containing a merge variable until the column that populates it exists and has been tested on 20 real rows.

The most common failure in outbound is a sequence built around a personalized finding with no pipeline behind it. The copy reads beautifully and cannot be sent. If a message template contains {{clientName}}, {{techStack}}, {{recentPost}} or similar, the enrichment table is a prerequisite, not a later optimization. Say so before writing the copy.

Corollary: every enrichment column needs a fallback path for rows where it comes back empty. Coverage is never 100%. Design for the miss rate up front (see COVERAGE.md).

Workflow

1. Name the target variable. What single fact does the opener need? One per campaign. "Their biggest client's name and main ranking keyword" is one variable set. "Client name, tech stack, funding, and recent post" is four campaigns pretending to be one.

2. Trace the source. Where does that fact physically live? A portfolio page, a case-study URL, an og:description, a job post, a pricing page, their LinkedIn featured section. If it lives nowhere public, the variable is dead , change the opener instead of building a table. See SOURCES.md for where common attributes actually live.

3. Design the column chain. Cheap and deterministic first, expensive and probabilistic last. Every column after the first should be gated on the previous one succeeding. See COLUMNS.md for the standard chains.

4. Set run conditions on every paid column. This is the single biggest credit lever. A column with no run condition fires on every row including the 900 that will never qualify.

5. Test on 20 rows before running 2,000. Read the actual output, not the success rate. A Claygent column that returns a plausible wrong client name is worse than one that returns nothing, because the wrong name ships to the prospect.

6. Score and filter. Add a qualification column that decides Tier A (full personalization) vs Tier B (category-level fallback) vs drop. Export tiers as separate lists into separate sequences.

7. Export. Map fields deliberately. See EXPORT.md.

Column chain: the default shape

Seed rows            (Sales Nav / Apollo import, name, title, company, LI URL)
  → Find company domain          [cheap, deterministic]
  → HTTP request: homepage       [free-ish, gate: domain exists]
  → Find portfolio/clients URL   [AI, gate: homepage fetched]
  → Scrape that page             [gate: URL found]
  → Claygent: extract the fact   [expensive, gate: page scraped]
  → Validate / normalize         [formula, catches hallucination]
  → Tier score                   [formula: A / B / drop]
  → Write opener line            [AI, gate: tier = A]

Two properties matter more than the specific tools: each step is gated on the last succeeding, and the expensive AI step runs on the smallest possible subset.

Claygent prompting

Claygent is an LLM with a browser. It fails the same ways any LLM fails, so prompt it accordingly:

  • Ask for one fact, not a paragraph.
  • Demand a strict output format and a literal NOT_FOUND when absent.
  • Forbid inference explicitly: "Only return a name that appears verbatim on the page. Do not guess from the industry."
  • Give a negative example of what a wrong answer looks like.
  • Return the source URL alongside the fact so it can be spot-checked.

An LLM asked to find a client name will invent a plausible one rather than return nothing. The NOT_FOUND instruction and the source-URL column are what stop a hallucinated client name reaching a prospect's inbox. Full prompt patterns in CLAYGENT.md.

Credit control

Costs run away through unconditioned columns and re-runs on edited tables.

  • Gate every paid column with a run condition.
  • Put the cheapest disqualifier as early as possible so bad rows exit before the expensive columns.
  • Use waterfalls for contact data (provider 1 → only if empty, provider 2), never parallel providers on every row.
  • Turn off auto-update while building. Auto-update plus a column edit re-runs the whole table.
  • Budget check: rows × paid columns × credits before pressing run.

Anti-patterns

  • Writing the sequence copy before the enrichment table exists.
  • No NOT_FOUND path in a Claygent prompt.
  • No fallback opener for rows where enrichment came back empty.
  • Running the AI column on 100% of rows instead of the qualified subset.
  • Treating Clay as a data source rather than an orchestration layer.
  • Trusting a first-run output without reading 20 rows by hand.
  • One table serving several campaigns with different target variables.
  • Personalizing a fact the prospect does not care about (correct but useless).

References

  • SOURCES.md, where each common attribute actually lives, and realistic coverage rates
  • COLUMNS.md, standard column chains for the common campaign shapes, with run conditions
  • CLAYGENT.md, prompt patterns, output contracts, anti-hallucination guards
  • COVERAGE.md, tiering, fallback openers, what to do with the 40% that enrich empty
  • EXPORT.md, field mapping into HeyReach, Instantly, Smartlead, and the webhook/HTTP alternative

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.