Step 3: Build the key-terms table for the file
Technique Extraction schema
Structured extraction: define the schema, allow 'not present', require a clause reference per cell, define severity and confidence labels, and ask for full coverage with a count instead of 'only material'
The brief
Learning objective
Write an extraction prompt that produces a reliable, auditable table: fixed columns and rows, a clause reference for every cell, an explicit null value for absent terms, defined High/Medium/Low labels so the reviewer can filter instead of the model pre-filtering, and a coverage instruction with a closing count so the model lists every relevant clause and never silently narrows scope.
Scenario
The partner wants a one-page key-terms table of the lease on the file before the client call, with each term flagged for how much it matters to the tenant so she can scan it in a minute. Three things typically go wrong when AI does this: it invents terms that are not in the lease (a governing-law clause, a rent review) because leases 'usually' have them; it stops after the first relevant clause when several apply; and, if told 'only the material terms', it quietly drops rows. You will paste the lease extract into the assistant.
Your task
Write the prompt that produces the key-terms table. The rubric checks whether your prompt defines the columns and rows, allows and requires 'not present' where a term is missing, demands a clause reference for every value, defines a significance label and a confidence label with one-line definitions, and instructs the model to cover the whole extract, say how many clauses it found where more than one applies, and finish with a count of items found.
Materials
What you have on file
Extract from Lease dated 1 March 2021, Orrin Property Holdings (Landlord) / Brackwater Foods Ltd (Tenant), Unit 4 Kestrel Park. 1.1 'Break Date' means 1 March 2027. 'Term' means ten years from 1 March 2021. 'Rent' means the annual rent and all other sums reserved as rent. 3.1 The Tenant shall pay Rent of 610,000 per year by equal quarterly instalments in advance. 5.2 The Tenant shall keep the Premises, including the Refrigeration Plant, in good and substantial repair and condition. 8.1 The Tenant shall not make structural alterations. Non-structural alterations require the Landlord's prior written consent, not to be unreasonably withheld. 12.1 The Tenant may terminate this Lease on the Break Date by serving on the Landlord not less than nine months' written notice. 12.2 A notice under clause 12.1 is of no effect unless on the Break Date (a) the Tenant gives vacant possession, (b) all Rent due has been paid, and (c) no material breach of the Tenant's covenants is subsisting. 12.3 Time is of the essence for clause 12. 14.1 The Tenant shall not assign or underlet the whole without the Landlord's prior written consent, not to be unreasonably withheld. Assignment or underletting of part is prohibited. 19.1 Notices shall be in writing and delivered by hand or sent by recorded delivery post to the recipient's registered office; a posted notice is deemed served two working days after posting.
Paste these into your prompt where the task calls for it, or refer to them with a placeholder such as [paste lease extract]. When you run a prompt that uses a placeholder, the Lab appends the materials so the model has something to work on. Where the task asks you to paste an edited copy, nothing is appended: what you paste is what runs.
How you will be graded
Rubric (100 points)
The grader scores your prompt, not the output. Length and formatting earn nothing; a short prompt that hits every criterion beats a long one that misses one.
| Criterion | Pts | What good looks like |
|---|---|---|
| Defines the table explicitly: named columns and a list of required rows (at least parties, term dates, rent, repair, alterations, alienation, break date, break notice period, break conditions, notices, governing law, rent review) | 20 | A column list such as Term | Value | Clause | Quote | Significance | Confidence, and an enumerated row list that includes rows the extract does NOT contain (governing law, rent review) so absence is tested. |
| Permits and requires an explicit null value for absent terms and forbids inferring from market practice or general knowledge | 20 | 'If a term is absent write not present in extract and do not infer it'; the prompt does not say 'fill in every cell' without a null option. |
| Requires a clause reference and a verbatim quote or close paraphrase for every value | 15 | Each cell must be traceable: clause number plus the words relied on; no unreferenced values. |
| Asks for full coverage: whole extract, every relevant clause, a count where several apply, a catch-all for clauses not captured, and a closing count of items found | 20 | 'Where more than one clause bears on a row, list each and state how many you found'; 'after the table list any clause not reflected above'; 'one row per break condition'; 'finish with: Terms found: [n]; clauses not captured: [n]'. |
| Significance and confidence labels are requested and defined with a scale and a one-line definition each, anchored to the tenant's position and to how explicit the wording is | 15 | 'Significance for the tenant: High = affects the ability to exit on the Break Date or exposes the tenant to a claim; Medium = a negotiable cost or restriction; Low = neutral or market standard. Confidence: High = explicit wording in one clause; Medium = requires reading two clauses together; Low = requires interpretation or the wording is ambiguous.' |
| Avoids scope-narrowing language and provides a place for ambiguities and notes | 10 | No 'only material', 'key', 'be conservative' or 'brief summary' filters on the extraction; a Notes section for ambiguities with clause references. |
Hints
Stuck? Open one at a time.
Hint 1
Name the columns and every row you want; if a row can have several entries (break conditions), say 'one row per condition'. Give the model a legitimate way to say nothing is there: 'write not present in extract; do not infer from market practice'.Hint 2
Define the labels in one line each: what High/Medium/Low significance means for the tenant, and what High/Medium/Low confidence means (explicit wording vs. requires reading clauses together vs. interpretation). Labels let you filter; 'only material terms' makes the model filter for you and drop rows.Hint 3
Ask for a closing section listing any clause not captured in the table and a final count of items found, so nothing is quietly dropped.
Your prompt