Agent Skills: README Generator

README Generator

Prerequisites: Run schema-enricher first if schema.yaml is missing descriptions; ensure metadata.yaml is present and complete. When to use: Creating or updating README.md for any shared dataset, derived table, or table with multiple downstream consumers

🚨 REQUIRED READING - Start Here

BEFORE generating any README, review the following:

1. Layout conventions: READ references/layout_conventions.md

   • Section order, conciseness rules, anti-patterns to avoid
   • Information sources (which file to read for which section)

2. README template: READ assets/readme_template.md and COPY its structure

   • Fill every {placeholder} from the source files
   • Do not skip or reorder sections

Workflow

Step 1: Read source files

Read all three files before writing anything:

sql/<project>/<dataset>/<table>/query.sql     → source tables, GROUP BY dimensions, metrics, @param
sql/<project>/<dataset>/<table>/metadata.yaml → DAG, partitioning, clustering, retention, owners
sql/<project>/<dataset>/<table>/schema.yaml   → field names, types, descriptions for Key Fields section

If only query.py exists (no query.sql): note it — the Data Flow and How It Works sections may be incomplete or require manual input. Fill what is possible from metadata.yaml and schema.yaml.

Extract and record:

• FROM clause — source table(s) with fully qualified name
• GROUP BY fields — these become Dimensions
• Aggregated fields — SUM/COUNT/DISTINCT targets become Metrics
• WHERE clause — @param_name for Implementation Notes
• DAG name, partition field, cluster fields, owners — for Overview table
• Table version — from directory name (e.g., _v1)

Step 2: Check if README.md already exists

ls sql/<project>/<dataset>/<table>/README.md

• Exists → read it, identify sections to update or add (do not remove existing content without noting it)
• Does not exist → generate from template

Step 3: Write README.md

READ assets/readme_template.md and fill every placeholder:

📌 Overview table — use metadata.yaml for DAG/partition/cluster/retention/owner; derive Version from directory name.

🗺️ Data Flow — Mermaid flowchart TD with exactly 3 nodes:

• Node A: source table(s) with short label + fully qualified name
• Node B: **This query** with filter and GROUP BY description
• Node C: Partitioned table with time and cluster annotation
• For multiple sources: A1, A2 → B

🧠 How It Works — 4–5 numbered steps. Step 5 MUST explicitly state data inclusion/exclusion policy:

• "All records from source are included; no exclusions applied at this layer."
• OR list specific exclusions (bots, synthetic clients, test populations)

🧾 Key Fields — two sub-tables (Dimensions, Metrics). Use {a\|b\|c} shorthand for related field families. Group dimensions by: Date & Geo, Browser, Search, [Product] config, User. Omit dimension rows not applicable to this table.

🧩 Example Queries — exactly 3, graduated:

1. Basic aggregation — date filter + 1–2 GROUP BY dimensions
2. Segmentation — GROUP BY a user/product dimension with SAFE_DIVIDE ratio
3. Attribution/Advanced — multi-metric, WHERE filter on a dimension, SAFE_DIVIDE

Rules:

• Always use SAFE_DIVIDE() for ratios — never raw division
• Use GROUP BY 1, 2 shorthand
• Comment each: -- N. Description
• Fully qualified table name in FROM

🔧 Implementation Notes — 3–5 bullets extracted from query.sql logic.

📌 Notes & Conventions — bullet definitions for key fields from schema.yaml descriptions.

🗃️ Schema & Related Tables — one section; combine schema.yaml link + upstream + downstream.

Step 4: Conciseness check

Before finalizing, verify:

• [ ] Total line count ≤ 170
• [ ] No separate sections for Scheduling, Storage, Owners, Retention (all in Overview table)
• [ ] No separate "Schema Reference" + "Related Tables" (merged into 🗃️)
• [ ] SQL examples use GROUP BY 1, 2 shorthand
• [ ] How It Works uses single-line numbered steps (no multi-paragraph blocks)
• [ ] How It Works Step 5 explicitly states data inclusion/exclusion policy

If over 170 lines, trim by: shortening SQL examples, collapsing Notes & Conventions bullets, abbreviating How It Works steps.

Step 5: Write and report

Write the README.md to:

sql/<project>/<dataset>/<table>/README.md

Then read back the written file and confirm:

• [ ] All sections from the template are present and in order
• [ ] No {placeholder} tokens remain unfilled (exception: if only query.py exists, Data Flow and How It Works may be partially filled — note which sections and why)
• [ ] Line count is within target (≤ 170)
• [ ] Mermaid block renders valid flowchart TD syntax

Report:

• Path written
• Line count
• Sections included
• Any placeholders left unfilled (with reason)

Integration with Other Skills

| Skill | When to invoke | |---|---| | schema-enricher | Run first if schema.yaml is missing descriptions — needed for Notes & Conventions | | create-pr | After README.md is written — stages, commits, and opens a draft PR |

Decision Tree: Rich vs. Minimal Style

Table has multiple downstream consumers OR is a shared dataset?
  → Rich style (this skill)

Table is a UDF, static reference, or simple single-consumer table?
  → Minimal style: title + ## Description with 5–10 bullet points
  → Do not use this skill for minimal style

Example Invocations

Create a README.md for telemetry_derived.newtab_daily_interactions_aggregates_v1
Update the README.md for firefox_desktop_derived.newtab_clients_daily_v2 — add missing example queries
Generate README for ads_derived.impressions_v1