Governed, self-service GTM analytics stack. dbt-core + MetricFlow + DuckDB/Snowflake + a versioned skills/ folder consumed by Claude-style assistants.
A reconstructed, runnable GTM warehouse: 23 public metrics across 7 semantic models, column-level PII tags that propagate from source YAML through the manifest and get enforced in CI, and a skills/ markdown folder that lets an LLM assistant pick the right metric and emit a deterministic MetricFlow query instead of hallucinating SQL. The point is not another dbt project, it is the governance surface around it: tagged columns, singular reconciliation tests, a skill-to-metric crosswalk the build fails on, and a freshness check so stale metric docs cannot ship. Built solo against open sources (dbt-labs canonical layering, standard GTM formulas) as a sincere reconstruction of a published pattern.
git clone https://github.com/JayDS22/gtm-semantic-layer
cd gtm-semantic-layer
make up # pip install, dbt deps/seed/build, mf validate-configs
make q METRIC=arr GROUP=metric_time__monthRequires Python 3.11+. No cloud credentials needed, DuckDB is the default target.
Seeds → staging (PII-tagged, source-tested) → intermediate (incremental MRR daily, account hierarchy reconciliation, cohort builder) → marts (dims + facts) → semantic layer (MetricFlow YAML) → skills/ markdown consumed by an LLM router. See docs/architecture.md for the Mermaid diagram, docs/metric-glossary.md for all 23 metrics with formulas and confidence tiers, docs/skills-pattern.md for the skills folder convention, and docs/example-trace.md for a Slack-shaped user question walked end-to-end: skill selection → mf query → answer.
| Layer | What's here |
|---|---|
seeds/gtm/ |
8 realistic GTM seeds (SFDC + Stripe + product events + quotas + financials) |
models/staging/ |
6 staging models, PII tags, source tests |
models/intermediate/ |
7 ints incl. int_subscription_mrr_daily (incremental + merge), int_account_arr_monthly, int_account_cohort, int_account_hierarchy_reconciliation (hierarchy-detector adversarial mod) |
models/marts/core/ |
5 dims + 7 facts incl. fct_mrr_movement, fct_cohort_arr_monthly, fct_arr_qoq_growth, fct_opportunity, fct_funnel_event, fct_financials, fct_quota |
models/semantic/ |
7 semantic models, 23 public metrics + ~30 supporting |
skills/ |
23 metric files + 3 debug runbooks + 4 FAQs, each with frontmatter |
scripts/ |
pii_check.py (manifest walker), skill_metric_crosswalk.py, skill_freshness.py |
tests/singular/ |
account-hierarchy reconciliation, subscription-period overlap, MRR-movement reconciliation, ARR↔billed reconciliation (adversarial-pass mod), NRR cohort denom nonzero |
- Revenue (7):
arr,mrr,new_arr,net_new_arr,grr,nrr,arr_qoq_annualized_growth_pct - Pipeline (5):
pipeline_coverage,pipeline_coverage_qtd,win_rate,avg_sales_cycle_days,weighted_pipeline_arr - Funnel (4):
mql_to_sql_conv,sql_to_won_conv,mql_to_won_conv,activation_rate - Retention (2):
expansion_rate,logo_retention - Efficiency (5):
cac_payback_months,magic_number,burn_multiple,gross_margin,rule_of_40
Full table with grain, formula, dimensions, and confidence_tier in docs/metric-glossary.md. Honest tiers after Day 6: cac_payback_months and rule_of_40 both lifted to stable (segment-weighted CAC denominator with real new_arr by segment; Rule-of-40 now carries both growth + margin terms). The 3 funnel conversions remain evolving (same-period proxy aggregation, true per-user cohort tracking is Day 7+).
A plain-English GTM question routed through the skills/ folder to a deterministic MetricFlow query. The script at scripts/ask_claude.py stands in for the LLM step (keyword scorer over the 23 skill cards) so you can see the control flow without an API key; in production the scorer is any model that can rank skill files. The emitted mf query command and the returned tabular output are the real contract.
Try it locally after make up:
./scripts/ask_claude.py "show ARR trending by month" --run
./scripts/ask_claude.py "NRR by cohort monthly" --run
./scripts/ask_claude.py "rule of 40 by quarter" --run
Scripted terminal demo rendered via charmbracelet/vhs from .vhs/demo.tape. No live recording; re-render after any user-facing change with vhs .vhs/demo.tape.
Static captures (same source, individual frames):
mf query --metrics arr --group-by metric_time__month returning ARR by month from the live DuckDB warehouse.
tree skills/ showing the 23 metric files + 3 debug runbooks + 4 FAQs under their standard categories.
CI status: the green badge at the top of this README is live from GitHub Actions on the main branch.
- Day 1, Skeleton, seeds, staging, source YAML, CI green
- Day 2, Marts + first 5 metrics
- Day 3, Remaining 15 metrics + skills folder + CI crosswalk
- Day 4, PII enforcement + 4 singular tests + skill freshness + docs
- Day 5, segment-weighted CAC payback (stable), demo GIF via vhs, screenshots, repo public
- Day 6, Rule-of-40 growth term (stable), pii_check + skill_freshness CI gates flipped to blocking
- Day 7, ask-Claude pattern demo (scripts/ask_claude.py + .vhs/ask-claude.tape) + README embed
- Day 8+, SCD2 opportunity snapshot, YoY growth once >= 5 quarters of history, resume + application link
The skills/ folder pattern and column-level PII taxonomy are a sincere reconstruction from Anthropic's June 2026 post How Anthropic enables self-service data analytics with Claude (Chen Chang et al.); the architecture here is a reconstruction of the published pattern, not a copy of unpublished internals. Metric definitions, warehouse schema, dbt macros, singular tests, and CI are mine, written against open sources (dbt-labs canonical layering, standard GTM metric formulas). The repo was originally named anthropic-gtm-semantic-layer and renamed to gtm-semantic-layer for portability, the pattern applies to any data-forward org, not just one.
Day-by-day progress in BUILD-LOG.md. The project was built from a 4-day design spec with adversarial self-review in the parent workspace (_designs/03-anthropic-gtm-semantic-layer.md); the adversarial mods are explicitly folded in: incremental + merge strategy on int_subscription_mrr_daily, int_account_hierarchy_reconciliation as a hierarchy-detector probe, pipeline_coverage_qtd as a sibling metric, and the ARR↔billed reconciliation singular test.