Iterate, Break Tasks Apart and Reuse What Works

Most people use an AI assistant as a vending machine: insert request, collect answer, and if it is wrong, insert a different request. Skilled users treat it more like a pair of hands in a workshop. They look at what came out, say what to change, split the job into parts and keep the jigs that worked.

This article is about that second way of working. It is where "I tried AI and it was disappointing" becomes "I know how to get this done."

• How to give feedback that moves an answer closer to what you want.
• How to use critique and alternatives without being misled by agreement.
• How to break a large task into steps, with checks.
• How to build a small, reliable prompt library.
• How to tell when prompting is not the real problem.

Google's prompting guide opens with the observation that prompt design is an iterative process, and lists ways to vary a prompt: rephrasing, switching to an analogous task and changing the order of the content. That is the machine's side. Yours is the feedback side.

Compare these follow-ups to the same draft:

Strong feedback has three parts: what to keep, what to change and why. The why matters because it lets the model apply the principle elsewhere in the text.

1. Read the answer as an editor, not as an author. Highlight what is useful, what is missing, what is wrong.
2. Say what to keep. Without this, the next version may discard the good parts.
3. Name the change precisely, pointing at a paragraph, a step or a sentence.
4. Change one or two things at a time. A dozen changes make it hard to see which one helped.
5. Re-check anything that changed. New text brings new chances for invented details.

If you have corrected the same issue three times and it keeps returning, the conversation has probably become muddled. Copy your best version and your learnings into a fresh chat with a cleaner prompt that states all the requirements up front. This is not defeat. It is often the fastest route, and the improved prompt becomes a template.

You can use the model to widen your view, not just to produce the first draft.

Ask for alternatives. "Give me three different approaches to structuring this essay, with one sentence on the trade-offs of each." You choose, which keeps you in charge.

Ask for a critique with criteria. "Review the draft against this checklist: clear goal in the first two lines, no jargon, every claim has a source, under 300 words. For each item say pass or fail, and quote the lines that fail."

Ask for the strongest objection. "What is the best argument that this recommendation is wrong?" This is a direct counter to the agreeable tone models can fall into.

Ask what is missing. "What questions would a sceptical manager ask after reading this?"

A model reviewing its own answer is not a reliable check. It may share the same blind spots, repeat the same wrong fact as "verified", or invent flaws to appear helpful. Treat critique as a source of hypotheses to test, not as a verdict. Three practical rules:

• Give criteria. A review without criteria tends to be generic.
• Separate generation from verification. Generate in one step, then verify in another, ideally against a source or by running something. For important facts, go to an independent source. How to Fact-Check AI-Generated Answers shows how.
• Watch for flip-flopping. If "are you sure?" makes the model reverse a correct answer, the lesson is that agreement is not evidence.

Large tasks fail in large ways, and you cannot easily see where. Split them. Anthropic's article on building effective agents calls this basic pattern prompt chaining: decomposing a task into a sequence of steps, with each step processing the output of the previous one, and with checks in between.

Goal. Turn raw survey comments into a one-page summary for your manager.

One giant prompt. "Here are 200 comments. Write a report with themes, quotes, recommendations and a conclusion." You get a plausible report. Are the themes real? Are the quotes real? Does the recommendation follow? You cannot tell.

Chain instead.

Each step is small, so you can judge it quickly. Mistakes are caught before they flow into the next step. Notice also that you made the decisions at the points that need judgment: the themes and the actions.

• The task has more than one kind of work (extracting, judging, writing).
• You will reuse parts of it.
• The output must be correct, not just plausible.
• You notice the model dropping requirements from a long list.

Do not decompose when the task is simple and the answer is easy to judge. Splitting a two-line email into five steps is wasted effort.

If you solve a problem once and have to solve it again from scratch, you have lost the learning. A prompt library is a notes file or document holding your best prompts, organised for reuse. It is also the first step towards automation, because a prompt with named blanks is something a program can fill.

• A name and purpose ("Summarise a meeting transcript for people who were absent").
• The template, with clearly marked blanks.
• Example input and output, so you remember what good looks like.
• How to check the result ("compare action items with the transcript").
• Notes: which tool, when it last worked, known weak spots.

Here is a tiny program that fills a template and refuses to run if a blank is missing. It shows the idea in the simplest possible form. The "document" is a made-up sentence, and nothing calls an AI service.

# File: prompt_template.pyfrom string import TemplateSUMMARY_PROMPT = Template(    "Task: summarise the document below for $audience.\n"    "Length: at most $max_points bullet points.\n"    "Rules: use only the document; if something is not stated, write 'not stated'.\n"    "Document:\n$document")prompt = SUMMARY_PROMPT.substitute(    audience="a manager with two minutes",    max_points=3,    document="Q3 support tickets rose 12% after the app update. Most were login errors.",)print(prompt)try:    SUMMARY_PROMPT.substitute(audience="a manager")except KeyError as missing:    print("Template refused to run, missing value:", missing)

Task: summarise the document below for a manager with two minutes.Length: at most 3 bullet points.Rules: use only the document; if something is not stated, write 'not stated'.Document:Q3 support tickets rose 12% after the app update. Most were login errors.Template refused to run, missing value: 'max_points'

You do not need a program to benefit. A text file with [AUDIENCE] and [PASTE TEXT] markers does the same job by hand. The principle is the same: a named blank is a decision you cannot forget to make.

• Start small. Ten prompts you use are worth more than a hundred you do not.
• Name by job, not by cleverness. "Rewrite for a nervous first-time reader" beats "Prompt 17".
• Date and test. Note the date and the tool, and re-run a prompt after a big model change.
• Delete what you never use, every month or so.
• Never store secrets in it. Prompts with real names, keys or confidential details should not live in a shared document.

Sometimes five rewrites do not help, because the prompt was never the issue. Anthropic's prompt engineering overview makes a related point: it assumes you already have clear success criteria and a way to test against them, and notes that not every failing result is best fixed by prompt engineering. Check this list before you try a sixth rewrite:

The last-but-one row is the most overlooked. If you cannot judge the output, no prompting will fix that. The fix is learning, and AI can help with that too, as Module 3 shows.

• Replying with vague feedback like "improve this".
• Changing everything at once, so you do not learn which change worked.
• Letting the model grade its own homework and calling it verified.
• Making one giant request for a multi-part job.
• Never saving what worked, then rediscovering it each month.
• Keeping a stale library, with prompts nobody has tested since the tool changed.
• Blaming the prompt when the problem is the tool, the information or your own undecided goal.

Keep the first two sentences and the closing thanks. Cut the third paragraph,which repeats the deadline. Replace "I would like to bring to your attention"and similar phrases with plain wording, because the reader is a friendlycolleague. Target about 80 words.

Task: write a polite follow-up email to [RECIPIENT_NAME], a [RELATIONSHIP](for example client or manager), after [EVENT, e.g. a meeting on date].Goal of the email: [ONE_ASK].Facts to include, and only these: [NOTES].Tone: [TONE]. Length: about [WORDS] words.Do not add any dates, commitments or numbers that are not in the facts.

• Next: How to Use AI to Learn a Difficult Technical Subject starts Module 3 and puts these skills to work on learning.
• To go from templates to automation, see Chatbots, Workflows, Automation and Agents: Choose the Simplest System.
• To build your own library over the course, continue with Design Your Personal AI Operating System.