The query is a specification,
not a conversation.
Prompt craft is the discipline of constructing a retrieval request with exactly the constraints the question requires — no more, no fewer. An over-constrained query misses evidence. An under-constrained query returns noise. This course teaches the practitioner to find the right boundary every time.
LLX-ACADEMY-CRS-01 · R1
This is the entry course.
Prompt craft is the first course in the Academy curriculum. No prior PARALLAX RC® certification is required, but the practitioner must have access to a sealed corpus for the practicum sessions.
Access to a sealed corpus
The practicum exercises require a live sealed corpus. The corpus does not need to be the practitioner's own — a training corpus provided by the Academy is sufficient for the examination.
Basic familiarity with PARALLAX RC®
The practitioner should understand what PARALLAX RC® does: ingestion, indexation, and retrieval. The architecture courses are not required but are recommended before the practicum.
What the ASL expects from a well-formed query.
The Advanced Semantic Linter decomposes every query into typed retrieval primitives. This session explains what those primitives are and what the ASL does when a query cannot be decomposed — which is the most important failure to understand.
Retrieval primitives
The five primitive types: passage search, metadata filter, spatial constraint, temporal range, entity match. Each maps to exactly one index. A query that does not map cleanly produces an incomplete plan.
The decomposition step
How the ASL reads a natural language query and identifies its primitives. What happens when a query is ambiguous — and why ambiguity is resolved by the practitioner, not the engine.
Plan determinism
The same query produces the same retrieval plan every time. This is a guarantee and a responsibility: if the plan is wrong, it is consistently wrong. Practicum: submit the same question three ways and compare the plans.
Failure modes
What an under-constrained query looks like in the result set. What an over-constrained query looks like. How to read the plan to diagnose which problem you have before reading a single passage.
What you can constrain and what you must not.
Every primitive type accepts a different set of constraints. This session works through each type: what it can target, what it cannot, and what happens when a practitioner tries to constrain something the relevant index does not hold.
Temporal constraints
Date ranges against the XATR index: creation date, issue date, transmission date, received date. Which date type to constrain for a given forensic question. What happens when dates conflict.
Entity constraints
Party names, document numbers, obligation types against the KUNZU index. How to constrain an entity match without accidentally excluding a variant spelling or a related entity.
Spatial constraints
Zone targeting against the spatial index: page range, bounding box, drawing zone, section. When spatial constraints add precision and when they remove relevant evidence.
Leaving constraints open
Not every dimension should be constrained. This session identifies the common error of over-specifying: constraining a date when the question is about a party, constraining a zone when the question is about a topic. Practicum: deliberately over-constrain and observe the result.
The result is a citation set, not an answer.
The ASL returns a set of cited passages, a linter report and a retrieval plan. This session teaches the practitioner to read all three — in order — before making any claim about what the corpus contains.
The passage set
How passages are ranked. What the rank means. Why the top-ranked passage is not necessarily the most evidentially significant — and how to identify which one is.
The citation
Reading the citation: file path, file hash, page, byte-range, ingestion timestamp, SHA-512 seal. What each field proves and what it does not prove. Practicum: locate the cited passage in the source file using only the citation.
The linter report
What contradictions look like in a linter report. What an empty result looks like. What a result with high passage count but low relevance looks like — and how each is a signal about the query, not about the corpus.
The empty result
An empty result from a well-formed query on a complete corpus is a forensic finding. This session explains how to record, present and defend an empty result as evidence of absence — and when it is not safe to do so.
A retrieval task with a verifiable result.
The prompt craft assessment is a set of forensic questions posed against a sealed examination corpus. The practitioner submits a retrieval plan for each question, executes the retrieval, and presents the passage set with its citations. The examiner replays every retrieval.
Retrieval plan submission
For each question, the practitioner writes the retrieval plan before executing it — stating the primitive types, the constraints and the expected index for each primitive. The plan is submitted and recorded before the retrieval runs.
Execution and citation
The practitioner executes the retrieval and presents the top-ranked passages with their citations. Each citation must be verifiable against the sealed corpus.
Examiner replay
The examiner replays each retrieval using the submitted plan against the same sealed corpus. A passing result means the examiner reaches the same passage set.
Passing prompt craft contributes to Certified operator.
Prompt craft is the first module of the Certified operator examination. Passing it alone does not confer the certification — custody and verification must also be passed. It does, however, unlock provisional access to the retrieval interface for supervised practicum work.