Frameworks & Templates4 min read

Change One Thing: A Prompt Experiment Template for Work

A ten-minute controlled test shows which prompt instruction improved a reviewed work result—and which change not to save.

Bokili EditorialPublished August 21, 2026· Verified August 21, 2026

A professional compares a baseline prompt with one changed prompt using the same source and success checks.

Prompt improvement often looks like trial and error: add a role, change the tone, insert an example, shorten the format and ask for more detail—all in one edit. If the result improves, you do not know why. If it gets worse, you do not know what to undo. A small prompt experiment replaces guesswork with evidence. Keep the task and safe source material fixed, define what a useful output must do, run a baseline, change one instruction and compare the two results. The aim is not to discover a universal “perfect prompt”. It is to learn which instruction improves one repeated work task.

Why change only one prompt element?

Current guidance from OpenAI describes prompting as an iterative process: start with a prompt, review the result and refine it. Google Cloud calls prompt engineering test-driven and says objectives and expected outcomes should be clear before systematic testing. Anthropic similarly advises defining success criteria and a way to test them before improving a draft prompt. For everyday work, the simplest usable version of that idea is a controlled comparison. One changed element gives you a plausible explanation for the difference.

The ONE prompt experiment

Outcome

Write two or three checks for the work result before you test. Make them observable: facts match the source; all owners appear; the update stays under 150 words.

Neutral baseline

Run the prompt you use now with one approved, non-sensitive input. Save the complete prompt and output.

Edit one thing

Change one component only: context, constraint, example, sequence or output format. Keep the tool, model, input and settings unchanged.

Evaluate

Review both outputs against the same checks. Record pass, partial or fail, plus any new error or extra review work.

Note the decision

Keep the change only if it produces a more useful reviewed result. Save the winning version with its task, date and limits.

Choose a task you can judge

Use a repeated, low-risk task with a clear source and a human reviewer: turning approved notes into a project update, extracting actions from a meeting record or rewriting a public announcement for a defined audience. Avoid personal data, confidential material and decisions that affect people. The experiment tests wording, not whether the AI can replace judgement. It also needs a task where “better” means more than “I prefer this version”.

Worked example: a weekly project update

Baseline:
Using the approved notes below, draft a weekly project update.

One-change version:
Using the approved notes below, draft a weekly project update. Use exactly three headings: Progress, Risks, Next actions. Under Next actions, name the owner and due date only when both appear in the notes. Do not invent missing details.

Review both drafts with the same checks: every factual claim is traceable to the notes; every named owner and date is present in the source; the three headings are used; missing information is left missing rather than guessed.

Only the output-format and evidence constraint changed. The source notes, tool and review checks stayed the same.

Reading is a start. Practice makes it stick.

Start learning

Record the result, not just the preference

A prompt can sound clearer and still create more work. Count what happened in the reviewed output. Did it pass the checks? How many corrections were needed? Did the changed instruction remove one error but introduce another? If both versions fail, keep the baseline and test a different single change. Do not combine several unproven edits into a new “best” prompt.

	Baseline	One-change version
Source fidelity	One unsupported owner added	No unsupported owner added
Required structure	Mixed paragraphs	Three requested headings
Missing details	Due date inferred	Gap left visible
Human corrections	Three corrections	One correction
Decision	Keep as reference	Save for this task after a second test

Run a second example before you standardise

One successful input is a clue, not proof. Repeat the winning prompt on another representative example. Use the same criteria and a reviewer who understands the work. If it holds up, store the prompt beside a short usage note: the task it supports, the inputs it may receive, the checks a person must complete and the conditions that require escalation. This prevents a useful local result from being copied into unrelated work.

A 10-minute prompt experiment

Pick one safe, repeated task and one representative source.
Write three pass checks before opening the AI tool.
Run your current prompt and save the output as the baseline.
Change one instruction only, then run it with the same source.
Score both outputs against the same checks and count corrections.
Record what changed. Keep the new version only if the reviewed result improved.

What this template is—and is not

This template is for learning from a prompt change. It is not a benchmark between models, a security test or permission to use sensitive data. Model behaviour can change, so even a saved prompt needs periodic rechecking. For a broader acceptance brief before delegating work, use Bokili’s Define Done before AI delegation. For a general output review routine, see Verify AI output. The experiment sits between those two habits: define success, isolate one change, then verify what actually happened.