Write a prompt contract for a document assistant
Specify evidence, missing-data behavior, and output requirements for a document assistant, then test the prompt against concrete support questions.
Foundation
The journal / Ideas in progress
Engineering & everyday learningNotes from
the work.
Software, systems, and the things I’m learning along the way. From first principles to AI in production.
Find something to readAll notes 9
A little theory. A lot of practice.
Build a data inventory before a handbook assistant
Map source authority, permitted uses, audience and derived copies in a fictional handbook assistant, then design a concrete document-withdrawal exercise.
Foundation
Define data contracts beyond column types
Validate units, reporting scope, batch completeness and execution identity before computing a job-duration metric, while preserving legitimate empty results.
Foundation
Split an LLM evaluation by source document and time
Prevent evaluation leakage by grouping related questions and document revisions, using time-aware boundaries, and keeping prompt development away from the final holdout.
Foundation
Forward a result before it reaches the register file
Select the newest available producer for each operand, trace two writes to the same register, and distinguish forwarding from waiting for an unavailable result.
Foundation
Preserve raw values before cleaning a CSV
A messy quantity column retains original tokens and rejection reasons. A worked analytical example with explicit assumptions and checked results.
Foundation
Declare the grain of an analytical table
Define what one analytical row represents, keep receipt and line measures separate, and distinguish a missing line from an observed zero with executable SQL.
Foundation
Distinguish empty, unknown, and not applicable data
A commute survey separates an observed zero, declined answer, not-applicable record, and collection failure. A worked analytical example with explicit assumptions and checked results.
Foundation
Use primary keys and foreign keys to preserve an order
Model order identity, reject orphan references, inspect cascades and separate database relationships from the policy for deleting a paid order.
Foundation