You will be able to turn vague requests into complete briefs and compare the results side by side.
If you look back through your assistant chat history, most of it is probably short one-line requests such as "summarise this", "make this more professional", "ideas for team bonding" and "reply to this email". You probably rewrote most of the answers to those requests yourself. That history is useful evidence. Take five prompts from it and rewrite each one as a proper brief. Then run both versions, score what comes back and note what made the difference.
The rewrites use the four parts of a brief from lesson 1.1, "The assistant only knows what you put in front of it". The four parts are context, task, constraints and format, and lessons 1.2 and 1.3 cover each one. Context is who you are, who is involved and why it matters. The task is what to do, for example revise. Constraints cover length, what to keep and what not to invent. Format is the shape of the output, for example email body only.
Use prompts that were actually sent, either by you or by colleagues who are happy to share them. A prompt you write specially for the exercise comes out too tidy. Real prompts are short, a bit vague and written in a hurry, and those are the faults this exercise is meant to fix.
Choose a spread of task types, because the four parts do not matter equally for each type. A good mix is a writing task, a summary, a request for ideas and a reply to someone. If all five prompts are emails, you learn a lot about emails and not much else.
Copy each prompt word for word into a document, together with any text or file it was used on. Both runs must use the same input, or the comparison means nothing.
Rewrite each prompt using all four parts, and keep it as short as the job allows. Tidying a two-line message needs no paragraph of background. A client proposal might need five sentences before the task.
Priya is the HR executive at a logistics company in Singapore you met in lesson 1.2. One of her five prompts was one she had sent the week before: "make this more professional", followed by a draft reply to a job applicant. He had interviewed for a warehouse supervisor role and was not selected. He was polite and well qualified, and the company may want to approach him for future openings.
Her rewrite had all four parts. For context, she said she works in HR at a logistics company in Singapore, explained who the candidate is and said why the relationship matters. The task was to revise her draft reply. She set four constraints. The reply had to be courteous and clear without sounding like a form letter, and it had to stay under 120 words (an example figure). It also had to keep her point about future openings and add no reasons for the decision that were not in her draft. The format was email body only. The rewrite took her about two minutes to write.
Run both versions in the same assistant and the same mode. Use a fresh chat for each run so neither run affects the other. If memory or custom instructions are switched on, you can leave them as they are for both runs. Module 6 covers them.
Give each output a usability score, which measures how close it is to something you would actually send or use. Use this scale:
1: unusable, you would start again yourself 2: a few usable lines, mostly rewritten 3: usable after substantial editing 4: usable after light edits 5: usable as it is
You are scoring the output, so how well the prompt is worded does not count. Judge each output before you look at its partner. Better still, ask a colleague to score the outputs blind.
Priya's original prompt produced a stiff, longer version of her draft that dropped the line about future openings. She scored it 2. Her rewrite kept that line, cut her draft by a third and read as if a person had written it. She scored it 4 because she still changed one phrase.
Most people skip this step, and it is where the learning happens. For each pair, write one line on which of the four parts changed the result most. Then look for patterns across all five pairs.
For Priya, the part that mattered most changed from task to task. On the candidate reply, constraints made the biggest difference, because "keep my point about future openings" stopped the model dropping the most important sentence. On a request for team-bonding ideas, context mattered most. Once the model knew the team worked shifts across two depots, it stopped suggesting weekday dinners. On a meeting summary, format made the biggest difference. Asking for "decisions, then actions with owners" turned a page of prose into something she could forward. Her five originals scored 2, 1, 3, 2 and 2, an average of 2.0. Her five rewrites scored 4, 4, 5, 3 and 5, an average of 4.2.
Your results will differ from hers. Working out which part matters most for the kind of work you do is more useful than any general rule. If a rewrite does no better than the original, record that too. It usually means one of two things. Either the task was simple enough that a short prompt was fine, or the missing piece was something the four parts do not cover, such as an example of good output, which module 2 covers.
Put your results in a table with one row per prompt and six columns. The columns include the original prompt, the rewrite, the score for the original, the score for the rewrite and a one-line note on what made the difference. Keep the rewrites. The best ones become the first entries in the prompt library you build in module 9, and you will be surprised how often you reuse them.
Set aside about twenty-five minutes, open a document for your table, and start with the prompt from your history that annoyed you most.
Make a before-and-after table of five prompts. Add the scores and a one-line note on which part of the brief changed the result most.
Junxiong-WFG Organisation is an authorised representative of AIA Financial Advisers Private Limited (Reg. No. 201715016G).