Lesson objective

Write constraints that stop a draft from making a pesticide, irrigation, tree or design decision, and refine a prompt by fixing its cause rather than editing each output.

Plain-language explanation

Why this matters now

Most prompts go wrong in one of two ways. Either the draft invents a fact, or the draft makes a decision nobody asked it to make. The first is caught by checking against sources. The second is harder, because a decision often looks like helpful advice: "consider a spring pre-emergent," "increase run times in zone 3," "the limb looks stable," "move the patio two feet east."

In landscaping, each of those sentences sits inside someone's qualified work. A draft that makes them has done so on paper, and paper is where customers read promises and crews read instructions.

A question to sit with

A crew leader's notes say "grubs found in two spots." The draft customer update says "we recommend a grub treatment this month." Nobody asked for a recommendation. Who made that decision, and who will the customer think made it?

What goes wrong

Constraints that are too vague. "Be careful with chemical advice" is not a constraint a model can follow or a reviewer can check.

Constraints only in the reviewer's head. The account manager knows not to promise treatments; the prompt does not say so.

Negative rules without a place for the question. Told only "do not recommend treatments," the model may drop the grub observation entirely, and the customer never hears about it.

Fixing the output, not the prompt. Each week someone deletes the same invented recommendation. The prompt never changes, so the error never stops.

Refining by adding more words. A prompt grows longer and vaguer with every fix until nobody knows which instruction is doing the work.

Believing refinement moves the frontier. A prompt refined ten times still cannot make a model a reliable judge of a tree, a pest or a label.

The practical answer

Boundary constraints: name the decision, then give it a destination. Each constraint has two halves. The first says what the draft must not decide. The second says where the question goes instead.

Boundary "Must not" half "Instead" half Qualified owner
Pesticides and fertilizers Name a product, rate, timing, site or conditions, or say a treatment is needed List under "Questions for the certified applicator" Certified pesticide applicator
Irrigation Give run times, schedules, design changes or repairs List under "Questions for the irrigation lead" Licensed irrigator where the state licenses the work
Trees Say a tree or limb is safe, unsafe, healthy or diseased List under "Questions for the arborist" Licensed tree expert or arborist where the state licenses tree work
Design Change grades, structures or drainage, or produce a plan that falls in a state's landscape architecture practice List under "Questions for the designer" The designer, or a licensed landscape architect where required
Digging Say where it is safe to dig List under "Questions for the crew leader before any dig" The person who authorises the dig
Promises Offer work, dates or prices not in the sources List under "Questions for the account manager" The person who signs the proposal

Which of these a state licenses varies. Check the rules that apply to your company with your own state authorities.

Refinement: fix the cause, once. When a draft needs the same correction twice, change the prompt, not the next output. Use a short loop.

  1. Name the error exactly. "Draft recommended a grub treatment from a crew observation."
  2. Find the layer that allowed it. Missing boundary constraint (layer 1), missing observation label (layer 3), or no "Questions for" list (layer 4).
  3. Make one change. Add the pesticide boundary with its destination.
  4. Rerun on the same inputs. Keep the notes identical so you can see the effect of the change alone.
  5. Check two things. Did the error go? Did anything else break, such as the grub observation disappearing?
  6. Record the change. Module 03 covers versions; at minimum, date it and note why.

Where refinement stops. Research, by analogy: in the Dell'Acqua experiment (Organization Science, March 2026; 758 consultants at Boston Consulting Group, using GPT-4 in 2023), people using AI on a task outside the capability frontier "were 19% less likely to produce correct solutions compared with those without AI." The frontier is invisible from outside. Refinement makes a drafting task more reliable; it does not make identification, treatment choice, tree assessment or irrigation design into drafting tasks.

A confident draft about a boundary is still a boundary. Professional guidance (NIST AI 600-1, July 2024; voluntary): models "sometimes provide logical steps for how they arrived at an answer even when the answer itself is incorrect." A recommendation that arrives with a neat explanation is not more trustworthy for it.

Remember this

  • A boundary constraint has two halves: what the draft must not decide, and where the question goes instead.
  • Pesticide, irrigation, tree, design, dig and promise boundaries each have a named, qualified owner.
  • When the same correction appears twice, fix the prompt, not the next output.
  • Change one thing, rerun on the same inputs, check for side effects, record it.
  • Refinement improves drafting; it never turns a qualified decision into a drafting task.

Landscaping example

Marlowe Property Services is a fictional grounds company serving three school campuses. The office lead, Colm Adebayo, notices that the weekly campus update prompt keeps producing two kinds of sentence reviewers delete: "we suggest adjusting the irrigation in the athletic field zones" and "the oak by the bus loop appears healthy." He runs the refinement loop. Step 1: he names both errors exactly. Step 2: he finds that the prompt's constraint block says only "no chemical advice," so irrigation and trees have no boundary at all. Step 3: he adds one change first, the irrigation boundary with its destination, "list under Questions for the irrigation lead." Step 4: he reruns on last week's notes. Step 5: the irrigation suggestion is gone and the crew's note about "dry patches in zone 5" now appears as a question. The oak sentence is still there. Step 6: he records version 1.1 with the reason. He then repeats the loop for the tree boundary: "Do not say a tree is safe, unsafe, healthy or diseased. Report what the notes say and list it under Questions for the arborist." On the same notes, the draft now reports "crew noted peeling bark on the oak by the bus loop [S2]" and lists it for the company's arborist, Tove Lindqvist. Colm records version 1.2. Over the next month, reviewers delete no boundary sentences, the irrigation lead receives three zone questions, and the arborist receives two tree questions she would otherwise not have seen. Neither question list contains a decision: each item reports what the crew wrote and waits for the qualified person.

Responsible-use warning

Do not write a boundary without a destination. A bare "do not recommend treatments" can make a model drop the observation altogether, so the customer and the qualified person never hear about it. Do not fix the same error in the output week after week while the prompt stays the same, and do not stack several changes at once, or you will not know which one worked or what it broke. No amount of refinement makes a model fit to choose a product, judge a tree or set an irrigation schedule.

Human-review requirement

The prompt owner approves every boundary block and every refinement, and checks that each change was rerun on the same inputs and recorded. The certified pesticide applicator reviews the pesticide boundary wording and receives every question it routes. The irrigation lead, licensed where the state requires it, and the arborist, licensed where the state licenses tree work, do the same for their boundaries. The account manager or crew leader reviewing each draft confirms that observations were reported, not dropped, and that no decision appears in the text.

Practical exercise

Use a fictional company, property and notes only.

  1. Write fictional crew notes containing one pest observation, one irrigation observation and one tree observation.
  2. Write a one-line constraint, "no chemical advice," and a draft a model might produce from it that breaks two boundaries.
  3. Name each error exactly, as in step 1 of the refinement loop.
  4. Identify which layer allowed each error.
  5. Write the two-half boundary constraint for each of the three observations, naming the destination and the qualified owner.
  6. Apply one change only and write the draft you would expect on the same notes.
  7. Check for side effects: did any observation disappear?
  8. Apply the next change and repeat steps 6 and 7.
  9. Write the version record for each change, with the date and the reason.
  10. List two tasks no refinement would make suitable for a model, and say who owns each.
  11. Write the "Questions for a qualified person" list your final draft would produce, and name the person who receives each item.
  12. Ask a colleague to run your final prompt on a second set of fictional notes and report whether any boundary sentence reappeared.
  13. Record in your workbook how many changes it took and what each one fixed.

Knowledge check

  1. What are the two halves of a boundary constraint?

    Answer guidance: What the draft must not decide, and where the question goes instead, with the qualified owner named.

  2. Why can a bare "do not recommend treatments" constraint cause harm?

    Answer guidance: The model may drop the observation entirely, so the customer and the certified applicator never hear about it.

  3. What are the six steps of the refinement loop?

    Answer guidance: Name the error exactly; find the layer that allowed it; make one change; rerun on the same inputs; check whether the error went and whether anything else broke; record the change.

  4. What does the OWASP 2026 project leads' statement say about fluent output?

    Answer guidance: When a model's fluent, confident output drives a decision or a tool call, a wrong answer turns into a wrong action.

Sources and further reading