“Make it answer from your document.”
Wednesday you saw the four-line pattern ground a model in the truth. Today you build it yourself — first on a practice document, then on one from your own life — and you set the trap that catches a model sneaking outside the fence.
Answer visitor questions about the gallery. #### THE WREN GALLERY · VISITOR ONE-PAGER The Wren Gallery is the campus art museum of a small Virginia college. Hours: Tuesday through Sunday, 10:00 AM to 5:00 PM; closed Mondays and university holidays. Admission is free for students with a valid campus ID; guests of students pay $5 at the door. Guided tours run every Saturday at 2:00 PM and last about 45 minutes; no reservation is needed for groups smaller than eight. Groups of eight or more must email the gallery at least two weeks in advance. Photography is welcome in all permanent-collection rooms, but flash photography and tripods are not permitted anywhere in the building. Food and drink stay in the lobby. The Rotating Wing changes exhibitions three times a year; current listings are posted at the entrance each term. Lost and found is at the front desk. The gallery closes early, at 3:00 PM, on the last Friday of each month for member events. #### Answer only from the document above. If it is not in the document, say exactly: It is not in the document. Question: When do guided tours run?
Practice documents (all fictional, invented for this lab): the Wren Gallery · the Colonial Cinema · the Tribe Trails one-pager
1Build the pattern on the practice document
- The box above is the whole pattern, already built: the task, the fictional Wren Gallery
document between
####fences, both magic sentences, and Q1 (findable) at the bottom. - Press RUN and read the answer underneath.
- Press ADD TO MY LOG, choosing Fixture questions Q1–Q3 as the destination.
2Trigger the fallback
In the same box, change only the last line to Q2 (absent): “Question: How much is parking at the gallery?” — leave the document and the two magic sentences exactly as they are — then press RUN again and add it to the same log box.
The document travels with every question. That is the grounding pattern: the model is never asked to remember, only to read what is in front of it.
3Set the trap
Change the last line once more to Q3 (the trap): “Question: What time does the gallery open on Mondays?” and press RUN. The document says closed Mondays — but models WANT to be helpful and may invent an opening time, or answer from general museum knowledge.
- If it said “closed Mondays” (from the document): the fence held. Log it.
- If it invented hours: you caught a leak. Tighten the prompt (“say exactly: ‘It is not in the document.’ Do not use any outside knowledge”) and press RUN again until the fence holds. Add both runs and your fix to the log.
4Your document
Swap in the document you brought (no document? use a spare practice copy above — and remember the document rules: public or course material only). Keep it to about a page — the paragraph, not the textbook; excerpt if it’s long. Rebuild the four-line prompt for THIS document, then run three questions through it — one RUN each, the document travelling with every question:
- findable — the answer is in the document;
- absent — the fallback must fire;
- your trap — plausible, related, NOT in the document. Think of something your document makes impossible to answer but a helpful model would be tempted to guess — your version of Step 3’s “Monday hours.” Log whether the fence held, and your fix if it leaked.
5Persona A/B on your best question
Take your best findable question. Press RUN on the SAME grounded prompt twice:
- A: prepend “You are a patient guide for someone brand-new to this.”
- B: prepend “You are a terse assistant for someone in a hurry.”
Log both. Two sentences: what changed (delivery) — and what didn’t (the facts, still from the document)?
✎Journal · reflection · submit
- Prompt journal: 3 real prompts from your week, one line each.
- Reflection (4 sentences): When is “I can’t find that in the document” more honest — and more useful — than an answer?
- AI Use Statement (50–150 words, required): tools; outputs quoted verbatim (word-for-word); what you checked. Missing → returned ungraded.
There is no file and nothing to upload: you do the lab right here in the hands-on session, and the submission panel below is where it all goes. Each piece is checked the moment you type it, and submitting records the lab and unlocks the next one.
Completion, not polish — every item present and genuine, and the points are yours. Nobody grades your prose. The checklist: fixture Q1–Q3 · your-document pattern + 3 Q&A · persona A/B, both outputs · fences visible in your logged prompts · journal · reflection · statement in range. Labs are 10% of the course, across six labs, and they unlock in order — one left undone blocks the next. No late window: finishing late beats not finishing — if you fall behind, tell your instructor rather than skipping ahead. AI use: Level 3 — submit your prompts and working results.
→Go further (optional)
- Pocket card #3 — the pattern (printable, in the Week 3 Go-further pack on Blackboard and under Materials).
- Zamfirescu-Pereira et al., “Why Johnny Can’t Prompt” (CHI 2023) — why this is hard for everyone, not just you.
- Reynolds & McDonell, Prompt Programming for LLMs §§1–3.
- Course reader part 2: Guidelines §§2.3–2.4 (Blackboard).
- Tag-pair fences — <doc> [the material] </doc> — read as structure and let you name the thing; use them when the material has parts.
- Practice: ground a model in the Tribe Trails one-pager — how far can you push before the fence gives?