Lesson objective

Run an assisted landscaping task through seven fixed steps, from naming the output and its qualified owner to recording the check, so every draft is built from supplied facts and checked before anyone acts on it.

Plain-language explanation

Why this matters now

Triage decides whether a task may use AI and on what terms. This lesson is what happens next: the same seven steps, in the same order, every time, so two office staff doing the same task produce work an account manager can check the same way.

The steps are fixed because landscaping errors are quiet on paper. A bed measured at 340 square feet looks the same as one measured at 430. A plant name that is one letter off reads as smoothly as the right one. A sentence that turns "we will monitor the patches" into "we will treat the patches" reads as a small edit and is really a promise of a pesticide application nobody decided on.

A question to sit with

An account manager turns a site walk into a customer update in three minutes and spends ten checking it against the notes and photographs. A colleague produces the same update and sends it. A month later the customer says the company promised to replace three shrubs. Which of the two can show what was actually said, and why?

What goes wrong

Starting with the question instead of the facts. "Write an update on the Harlow property" asks the model to supply facts it was never given. It will supply them.

Using stale facts. The notes are from the first visit, the measurements were corrected on the second, and the paste came from the first.

Letting the draft become the source. After two rounds of edits nobody has looked back at the notes or the photographs.

Checking the easy items only. The date and the customer's name get checked; the sentence that quietly added "and we will replace the two dead boxwoods" does not.

Accepting instead of challenging. Checking asks whether the draft matches the notes. Challenging asks what it left out, what it assumed and what the customer would read into it.

Drifting into qualified work. A draft update that "recommends a pre-emergent in early spring" or "suggests running zone 4 longer" has made a pesticide or irrigation decision on paper, whether or not anyone meant it to.

Nobody assigned to act. The update is filed, but no one is told to send it, so the customer hears nothing, or hears two different versions.

The practical answer

Seven steps. They are this programme's recommendation.

Step What you do Must be true before moving on
1. Name the output and owner One sentence: what will exist, and which qualified person owns it A named person, not a role
2. Assemble the facts Site notes, measurements, photographs, the signed scope, the label text if a label is involved Every item dated, and the latest version confirmed
3. Write the instruction Task, audience, format, constraints, and "not stated" when the facts are silent "Use only the supplied material" is in it
4. Generate the draft In the approved system only System approved for this data class
5. Check and challenge Check every item against its source, then look for what is missing or implied Every checkable item checked, not sampled
6. Owner's decision and assignment The owner decides and says who acts on it Anything pesticide, irrigation, tree or dig related has gone to its qualified owner
7. Record Log the task, sources, checks, corrections and owner Written now, not reconstructed

Step 2: the facts are the point. Whatever the model did not receive, it may invent. Whatever it did receive, you can check against. Identify each item precisely: "Site-walk notes, R. Quint, 12 April; bed measurements corrected 15 April; photographs 1 to 14 from 12 April; signed maintenance scope dated 3 February."

Step 3: four constraints travel with every instruction.

  1. Use only the supplied material; where it is silent, write "not stated" rather than filling the gap.
  2. Tie every factual sentence to a note, measurement, photograph or scope line.
  3. Recommend no product, rate, timing, irrigation setting or tree action; if the notes raise one, write "question for the qualified person" instead.
  4. Promise nothing the signed scope does not already include.

Module 02 builds these into a full prompt architecture.

Step 5: check, then challenge. Checking is a list.

Item Checked against
Measurements, areas, counts The corrected measurements, recomputed by a person, never taken from the draft
Plant names The notes and the photographs; an unfamiliar name is a question, not a fact
Work described as done The crew record for that date
Work promised The signed scope
Anything a label says The label itself, by the certified applicator
Dates and visit times The schedule

An AI output is never proof. A plant name, a measurement, a label statement or a product name produced by a model is checked against its source before anyone relies on it.

Challenge is three questions: what do the notes say that the draft does not, what does the draft say that the notes do not, and what would the customer take as a promise?

Step 6: decide and assign. Workload is exactly when review weakens.

Spring start-up and the first mowing weeks are the normal condition of this trade, so the owner's decision is a named step with a named person. The owner also says who sends or acts on the output, and which version.

Remember this

  • Supply the dated facts; never ask the model to recall a property, a plant or a label.
  • Four constraints: supplied material only, tie each fact to a source, no product, rate, irrigation or tree decisions, no promise beyond the signed scope.
  • Check every item, then challenge for what is missing or implied.
  • The owner decides and names who acts, on which version.
  • Record it now, while the details still exist.

Landscaping example

Rillwater Landscape & Irrigation is a fictional design-build and maintenance company. The account manager, Priya Castellano, is asked for a written update to the fictional Ashcombe Court HOA board after a spring walk. Step 1: the output is a one-page board update, owned by Priya. Step 2: she supplies her notes from 9 April, the corrected bed measurements from 11 April, photographs 1 to 18, and the signed maintenance scope dated 20 January. She notices her first paste used the uncorrected measurements and replaces them. Step 3: her instruction says use only the supplied material, tie each fact to a source, recommend no product, rate, irrigation setting or tree action, and promise nothing beyond the scope. Step 4: she runs it in the approved assistant. Step 5: checking finds three problems. The draft gives the entry bed as 520 square feet; the corrected figure is 250. It names the shrubs along the pool fence as "Japanese holly"; the notes say "unidentified evergreen, confirm". And it says the company "will apply a fungicide to the lawn spots next week", which nobody decided. Challenge finds that the draft never mentions the board's question about the dead crabapple. Priya corrects the area, restores "to be confirmed", removes the fungicide sentence and sends the lawn photographs to the certified applicator, Owen Farr. Step 6: Owen decides the spots need another look before any decision; Priya adds a line saying the lawn will be inspected, and assigns herself to send the update and ask the owner about the crabapple. Step 7: she records the task, sources, four corrections and both decisions. Checking took 14 of 20 minutes.

Responsible-use warning

Do not let the draft replace the facts. After the first round of edits it is easy to work only from the text on screen, and from then on nothing is being checked. Checking means returning to the notes, the corrected measurements, the photographs and the signed scope for every item. A fluent sentence that promises a treatment nobody decided on, or turns an unidentified shrub into a named species, is how a wrong fact leaves the company under its own name.

Human-review requirement

The named owner decides before the output is used: the account manager for customer updates and replies, the estimator and the person who signs the proposal for anything priced or promised. Any mention of a pesticide, rate or timing goes to the certified pesticide applicator, who checks it against the label or removes it. Irrigation settings and tree-safety statements go to the licensed irrigator or arborist where the state licenses that work. The owner checks the list rather than repeating it blind, confirms the challenge questions were asked, names who sends or acts on it, and confirms the record entry exists.

Practical exercise

Use a fictional company, property and customer only.

  1. Choose one recurring written task from your role, such as a board update, and describe it in one sentence.
  2. Write step 1: the output and its named, qualified owner.
  3. List the step 2 facts, each with a fictional date, and note how you confirmed each is the latest.
  4. Write the step 3 instruction with all four constraints in your own words.
  5. Write a fictional draft yourself containing one wrong measurement, one plant name the notes never gave, one promise beyond the scope and one product mention.
  6. Build the step 5 checking list: every checkable item and its source.
  7. Write the three challenge questions and answer them for your draft.
  8. Write the owner's decision note, including where the product mention went and who sends the output.
  9. Recompute every measurement in your draft by hand and record the arithmetic.
  10. Write the step 7 record entry.
  11. Write a short note to the person you assigned, naming the version they must send and what to do if the facts change before it goes out.
  12. Swap drafts with a colleague, repeat steps 6 and 7 on theirs, and list what each of you caught that the other did not. Record everything in your workbook.

Knowledge check

  1. Name the four constraints that travel with every instruction in this workflow.

    Answer guidance: Use only the supplied material and write "not stated" where it is silent; tie every fact to a source; recommend no product, rate, timing, irrigation setting or tree action; promise nothing beyond the signed scope.

  2. How does challenging differ from checking?

    Answer guidance: Checking compares each item with its source. Challenging asks what the notes say that the draft omits, what the draft says that the notes do not, and what the customer would take as a promise.

  3. What happens at step 6 if a draft mentions a pesticide product or rate?

    Answer guidance: It goes to the certified applicator, who checks it against the label or removes it. The model's wording never becomes the instruction.

  4. Which factors did the 2012 automation bias review find made over-reliance worse?

    Answer guidance: Workload, task complexity and time constraint, among others. The review concerned clinical decision support, so its use here is by analogy.

Sources and further reading