Lesson objective
Build the source block of a reusable plant prompt so the model works only from screened, released documents at stated revisions, each registered with an identifier the draft must cite.
Plain-language explanation
Why this matters now
Level 1 introduced a structured prompt for plant work and the screen that happens before it. Module 01 of this level put that prompt inside a seven-step workflow. This module takes the prompt itself apart and rebuilds it as an architecture: a set of blocks that a plant can reuse across people, shifts and tasks, and that a reviewer can inspect block by block when something goes wrong.
The first block, and the one that matters most, is the source block. It answers one question: what is the model allowed to work from?
No source in this programme shows that any particular prompt structure improves accuracy on manufacturing work. The architecture in this module is taught as this programme's recommendation because it makes the model's inputs visible and checkable, not because it is proven to make the model right.
A question to sit with
A quality engineer writes: "Using the customer's latest quality requirements, draft the receiving inspection steps for part 5540." She has not attached anything. Where, exactly, will the model find "the customer's latest quality requirements"?
What goes wrong
Asking for recall. "Per the customer's requirements", "per the standard", "per our usual process": each phrase invites the model to invent the content it refers to.
Supplying a document without its identity. A pasted table with no document number, revision or date cannot be checked later, because nobody can say which version it came from.
Supplying too much. A whole quote package attached "for context" puts every page through the screen and every page in front of the model, when the task needed two.
Supplying the wrong revision. The attachment came from an old email thread, not from document control.
Mixing sources without ranking them. The print says one thing, the customer email says another, and the prompt gives the model no rule for which governs, so it quietly picks one.
Screening after assembling. The source block is assembled first and screened second, by which time a controlled file may already be in the tool.
The practical answer
The source block has five parts. They are this programme's recommendation.
| Part | What it contains | Why |
|---|---|---|
| 1. Screen statement | A line confirming each document's data class and export status were settled, and by whom | Nothing enters the block unscreened |
| 2. Document register | One line per document: an identifier you assign, title, number, revision, release date | The draft can cite, and the reviewer can trace |
| 3. Extracts | Only the pages, sections or rows the task needs, each labelled with its register identifier | Less exposure, less noise |
| 4. Precedence rule | Which document governs when two disagree | Conflicts are surfaced, not settled silently |
| 5. Closed-source instruction | "Use only the documents in this block. Do not use general knowledge of any standard, customer or process." | Recall is ruled out in writing |
Part 1: the screen statement comes first because the screen comes first. Every document is checked before it is placed in the block.
Part 2: register every document. Give each one a short identifier the draft must use.
| ID | Document | Number | Revision | Released |
|---|---|---|---|---|
| D1 | Customer drawing | 5540-A | F | 18 Aug |
| D2 | Router | R-5540 | 4 | 25 Aug |
| D3 | Material specification | MS-22 | C | 2 Jun |
| D4 | Customer quality clauses (receiving section only) | CQ-11 | dated 1 Jul | 1 Jul |
The numbers and dates above are fictional. In a real prompt every line comes from document control, not from memory or from the file name on your desktop.
Part 3: extract, do not attach. Ask what the task needs. Receiving inspection steps need the drawing's inspection characteristics and notes, the material certificate requirements, and the receiving clauses. They do not need the pricing pages, the tooling drawings or the full process sheet. Label each extract with its identifier and location: "D1, sheet 2, notes 4 to 7."
Part 4: state precedence, and ask for conflicts. A simple rule works for most tasks: the released drawing governs dimensions and notes; the customer's dated clauses govern what the customer requires; the router governs sequence. Then instruct: "If two documents disagree, do not choose. List the conflict with both identifiers." A conflict found by the model is a question for the engineer or the quality authority, never something the draft resolves.
Part 5: close the source. Write the closed-source instruction in plain words. Add: "Where the documents are silent, write NOT STATED." A gap marked NOT STATED is a finding; a gap filled with a plausible value is an error waiting for a machine.
What an AI output is worth. An AI output is never proof. A dimension, clause or revision that appears in the draft is checked against the cited document before anyone relies on it, even when the citation looks exact. Models can produce citations that look right and point to the wrong place; NIST notes outputs "may also include confabulated logic or citations that purport to justify or explain the system's answer."
Remember this
- The source block decides what the model may work from; a prompt without one invites recall.
- Screen every document before it enters the block; undetermined export status keeps it out.
- Register each document with an identifier, number, revision and release date from document control.
- Extract only what the task needs, state which document governs, and ask for conflicts rather than choices.
- Require a citation on every value, and check the cited location, not the citation.
Manufacturing example
Kestrel Precision Machining is a fictional job shop. The quality engineer, Beatrix Omondi, needs receiving inspection steps for a fictional customer part, 5540, a machined aluminium housing. Her first attempt reads "Using the customer's latest quality requirements, draft receiving inspection steps for 5540." The draft cites a certificate retention period and a sampling plan that appear in no document Kestrel holds. She rebuilds the prompt with a source block. Screen statement: drawing 5540-A revision F recorded as not controlled by the shop's export contact, Ruth Salazar, on record EC-131; all four documents are in a class the approved assistant is cleared for. Register: D1 drawing 5540-A revision F, D2 router R-5540 revision 4, D3 material specification MS-22 revision C, D4 customer quality clauses CQ-11 dated 1 July, receiving section only. Extracts: D1 sheet 2 notes 4 to 7 and the inspection characteristics table; D3 section on certificates; D4 receiving section. Precedence: D1 for dimensions and notes, D4 for customer requirements. Closed-source instruction and NOT STATED rule added. The new draft cites every value. It reports one conflict: D1 note 6 calls for a certificate of analysis per heat, while D4 asks for one per shipment. It marks the sampling plan NOT STATED. Beatrix checks each citation against the documents and finds one wrong zone reference, which she corrects. She sends the conflict to the quality manager, Anders Moyo, who confirms with the customer before the steps are issued.
Responsible-use warning
Do not write a prompt that points at a document instead of supplying it. Phrases such as "per the customer's requirements" or "per our standard process" ask the model to invent content that looks authoritative. Do not attach whole packages for context, and never let a document into the source block before its export status and data class are settled. When two documents disagree, the model must report the conflict, not resolve it: the choice belongs to the engineer or the quality authority.
Human-review requirement
The named reviewer, usually the quality engineer or the engineer responsible for the document, reads the source block before the draft. They confirm that the screen statement names who settled export status, that every revision matches document control on the day, that the extracts cover what the task needs and no more, and that the precedence rule is right for this customer. They check every cited value at its cited location and send every reported conflict to the person with authority to resolve it, recording the outcome.
Practical exercise
Use fictional documents, parts and customers only.
- Choose a recurring drafting task from your role that relies on at least three documents.
- Write a one-sentence version of the prompt you might have used before this lesson, and underline every phrase that asks the model to recall something.
- Write the screen statement for a fictional document set, naming who settled export status and on what fictional record.
- Build a document register of three to five fictional documents with identifiers, numbers, revisions and release dates.
- For each document, list the extract the task needs and one part you would leave out, with the reason.
- Write a precedence rule for the set.
- Write the closed-source instruction and the NOT STATED rule in your own words.
- Invent one conflict between two of your documents and write how the prompt asks the model to report it.
- Write a short fictional output with three cited values, one of which cites the wrong location, and describe how the reviewer would find it.
- Name who in your fictional plant would resolve the conflict, and how the outcome would be recorded.
- Ask a colleague to check your register against a fictional document control list and note any revision they would query.
- Save the finished block in your workbook as a template, with a note of the date you last confirmed each revision.
Knowledge check
-
What are the five parts of the source block taught in this lesson?
Answer guidance: A screen statement, a document register, labelled extracts, a precedence rule, and a closed-source instruction that rules out general knowledge.
-
Why should a prompt say NOT STATED rather than let the model fill a gap?
Answer guidance: A gap marked NOT STATED is a visible finding for a person to resolve. A gap filled with a plausible value is an unsupported number that may reach a machine.
-
What should the model do when two supplied documents disagree?
Answer guidance: List the conflict with both document identifiers and not choose between them. The engineer or quality authority resolves it.
-
Does this programme claim that the block structure makes a model more accurate on plant work?
Answer guidance: No. No source shows that a prompt structure improves accuracy on manufacturing work; the structure is recommended because it makes inputs visible and checkable.
Sources and further reading
- National Institute of Standards and Technology — Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, NIST AI 600-1 (2024-07-26). Accessed 2026-09-15.
- Bureau of Industry and Security (U.S. Department of Commerce) — 15 CFR 734.13 - Export. Accessed 2026-09-15.
- United States Congress (via U.S. Government Publishing Office, govinfo) — 18 U.S.C. 1839 - Definitions (Defend Trade Secrets Act / Economic Espionage Act). Accessed 2026-09-15.
- AI Coalition Network — Certification Educational Standard: Human Review of AI-Assisted Professional Work. Accessed 2026-09-15.