Most workflows focus on prompt engineering. The real leverage is in systematic human review. My stage reduces a 60-minute edit to 15.
**Stage 1: Structural & Logical Integrity (5 min)**
* Verify the argument follows a clear problem -> solution -> evidence -> implication flow.
* Scan for logical leaps or unsupported claims. I tag these with `[LOGIC?]`.
* Check that all data points or examples serve the core thesis. Delete tangentials.
**Stage 2: Technical Accuracy & Specificity (5 min)**
* Replace all vague language. "Works well" becomes "Reduces runtime by ~40% on datasets >10GB."
* Validate all tool names, version numbers, and configuration parameters.
* For code snippets, I run a mental linter:
```sql
-- Bad: SELECT * FROM table
-- Good: SELECT id, event_date, status FROM fact_orders
```
**Stage 3: Tone & Consistency (3 min)**
* Remove any marketing superlatives (e.g., "revolutionary," "seamlessly").
* Ensure terminology is consistent (e.g., stick with "materialize" vs. "create" for dbt models).
* Read the final two paragraphs aloud to catch awkward phrasing.
**Stage 4: Cost & Efficiency Note (2 min)**
* Add a brief, practical note if applicable. Example: "Using this incremental model pattern cut our Snowflake compute costs by ~22% monthly."
This checklist is applied linearly. No backtracking. The draft either passes each stage or is recycled.
EXPLAIN ANALYZE