Gtm Engineering

---
name: gtm-engineering
description: "Build and evaluate GTM workflows in Clay: ICP, signals, list building, qualification, enrichment, sequencing, CRM routing, and outcome measurement. Use for operating or learning Clay and testing GTM automation reliability."
---

# GTM engineering

Start with the business outcome, its denominator, maturity window, and budget. Read existing client instructions and approved ICP before designing a workflow. Keep customer data and research corpora private.

## Reuse before building

Use list-audit and list-gate for inherited contacts. Locate the installed official Clay plugin and read its command help and relevant skills before assuming capabilities. Probe account and workspace identity. Native workflow APIs, table reads, browser operations, and paid enrichments have different capabilities and costs. Use the browser-use skill for verified gaps. Search the private research corpus with `gtme-library --corpus PATH search 'query'`; retrieval is not evidence that the entire corpus was studied.

Maintain separate identity, execution, and evidence graphs. An account match does not prove current employment. A successful HTTP response does not prove the destination contains the intended rows. A source citation does not prove its claim is true.

Read the graph-engineering skill and its task-graphs reference for substantial multi-source work. Queue independent jobs with named owners, dependencies, expected artifacts, and acceptance checks. Use the runtime's actual concurrency limit; keep ready jobs queued when slots are full. Reserve a separate verifier context before merging implementation or research claims. Do not add dependencies between jobs that do not consume each other's results. Serialize shared-file writes and browser mutations.

## Execute bounded tasks

Use `gtme-graph --help` to validate a frozen plan, initialize private state, select ready nodes, begin an attempt, and finish it with evidence. Each attempt reserves its maximum credits. Failed attempts also consume actual cost. Evidence must bind to the current run and attempt, be captured during that attempt, and include immutable source JSON with the required observations. Keep all attempt artifacts, including failures. This CLI enforces its own workflow state; it is not a global browser hook or an authenticated external witness.

Separate Actions, Data Credits, and external provider dollars. Never add unlike units. The graph budget tracks one explicitly chosen meter; account for other meters separately. Inspect auto-run and schedules before imports or edits. Start with a small fixture, then independently read back row count, stable IDs, mapped columns, and values. Retry only after inspecting whether the previous attempt partly succeeded.

Sending requires the user's actual authorization and approved campaign conditions. The graph tool intentionally never makes send nodes runnable. Drafting is not sending.

## Learn and measure

Read [domains.md](references/domains.md) before choosing what to evaluate, and [evaluation.md](references/evaluation.md) before claiming competence. Use `gtme-math order` only under its stated fixed-cost independence assumptions. Use `gtme-math funnel` for matured binary outcomes and `gtme-math evaluate` for sealed qualification labels. None of these commands establishes causal business uplift.

Apply [the decision standards](../../docs/DECISION-STANDARDS.md) to list selection, column assembly, enrichment ordering, copy, model routing and measurement. Compare a deterministic baseline with bounded alternatives; use dependency graphs for column order, constrained routing for enrichment and cost-sensitive classification where justified. Test learned selection policies offline or in shadow before promotion. Measure consequential errors, abstentions and cost per verified usable result. A sourced firm fact does not prove personal investment ownership or mandate fit. Five-row fixtures test function, not business uplift. Use Jev only for narrow typed judgments with validated evidence and abstention; it cannot authorize spending or sending.

After each substantial phase, compare results against fixed acceptance limits and choose continue, revise, or stop. Record this with `gtme-graph review` when running a phased graph. Keep verified claims, failures, client data and outcome labels private; publish only authorized sanitized procedures. Novelty and a public integration do not establish an exclusive advantage.

Update the relevant map or procedure only from observed behavior. Distinguish documented, inspected, tested, and independently verified capabilities. Record coverage gaps and failures; do not graduate a domain because its tutorial was read.

Use `clay-fixture-check` to compare an exported CSV with a frozen synthetic fixture. It checks content, not export authenticity or workspace identity. Use `gtme-learning evaluate` for paired holdout results and regression preservation before proposing promotion. Its local declarations do not prove evaluator independence or that a test was sealed in advance. Obtain those receipts separately.

Training note, 2026-09-23: user required mathematical, creative and proprietary standards throughout the workflow after recorded lessons failed to establish reliable transfer. Measure changed behavior and preserve private evidence, rather than equating added instructions with expertise.


## Clay correction retained — 2026-09-23

The user requires native Clay enrichment and dynamic, per-row personalized copy
through the built-in UX engine. Never use Sculptor. Do not replace configured
Clay columns with agent research or manually drafted copy. Inspect the live UI,
insert actual source-column tokens, disable each column's Auto-run before saving,
and verify the saved configuration. Test no more than five selected rows per
live test; inspect scope before running. Do not send, launch, activate, or
schedule campaigns. Keep incomplete configuration and unverified outputs explicit.

Retrieve the Clay navigation skill, map, and procedure before exploring again.
Reuse the existing `docs/LEARNING-TO-ACT.md` design: retain procedure, map,
preference, and strategy separately; replay before promoting a procedure; infer
parameters only from multiple observed instances. A UX receipt or graph route
is a partial learning aid, not a free-form recorder, automatic distiller,
retrieval engine, registry, or proof of mastery. Preserve failed attempts and
corrections with private evidence, and share only reviewed generalized lessons.