Iterate: bad to better
Diagnose before editing
A weak answer becomes easier to improve when you name the failure instead of asking vaguely for “better.”
Before you start
Why this matters
Suppose you ask an assistant to draft an update for a customer. The response is polished, but it hides the delay, invents a recovery date, and sounds more cheerful than the situation warrants. “Make it better” gives the model no reliable direction. Better could mean shorter, warmer, more detailed, more persuasive, or more accurate. Before changing the prompt, inspect the output as if you were reviewing someone else’s draft.
Iteration begins with a diagnosis. Compare the answer with the job you intended it to do. What must the reader know, feel, or do? Which facts came from your context? Which requirements were explicit? A useful diagnosis describes the gap between the target and the current result.
1Learn the idea
Read
Separate failure types
Most disappointing outputs contain one or more of these failures:
- Task failure: The answer performs the wrong job. You asked for a decision memo and received a general summary.
- Content failure: Important information is missing, unsupported, or invented.
- Audience failure: The vocabulary, background knowledge, or emphasis does not fit the reader.
- Format failure: The result ignores the requested structure, length, fields, or ordering.
- Tone failure: The voice is too formal, casual, promotional, apologetic, or forceful.
- Reasoning failure: The conclusion does not follow from the evidence, alternatives were ignored, or assumptions stayed hidden.
- Safety failure: The output exposes private data, overstates certainty, or suggests an action beyond the tool’s appropriate role.
These categories can overlap, but naming the primary failure helps you choose the next edit. If a report contains a fabricated statistic, shortening it is not the priority. If every fact is right but the manager needs a three-line decision, adding more context will not help.
Read
Trace requirements to the answer
Create a small requirement check:
- Required outcome: present the delay and request approval for a revised plan.
- Required facts: original date, current status, known cause, next decision.
- Prohibited content: guessed completion dates and blame.
- Format: subject plus fewer than 120 words.
- Tone: direct, calm, accountable.
Mark each item met, partly met, or missed. Quote the exact sentence that caused concern. “Tone is bad” is difficult to act on; “the phrase ‘fantastic progress’ conflicts with a two-week delay” is precise. Evidence-based feedback also prevents you from changing a section that already works.
Read
Distinguish prompt problems from evidence problems
Sometimes the prompt is unclear. Other times the necessary information simply does not exist. If your notes contain no owner or deadline, no wording trick can produce a trustworthy owner or deadline. Ask the model to label the field “TBD” or ask you a question. Do not polish a guess until it looks authoritative.
Likewise, an assistant may lack current information. A confident answer about today’s policy, price, or schedule needs a current source. Iteration inside the same conversation does not make stale knowledge fresh. Pause, gather evidence, then continue with the source and its date.
Read
Write a one-line diagnosis
Use this pattern:
The draft currently [observable problem], but the output needs to [requirement], because [reader or risk reason].
For example: “The draft announces the recommendation before explaining its evidence, but the output needs to separate facts from assumptions because the approver must judge uncertainty.” That sentence points toward a structural revision. Another diagnosis might be: “The summary assigns Priya a Friday deadline that does not appear in the notes; preserve only stated commitments and mark missing fields TBD.”
Avoid diagnoses based only on taste. “I don’t like it” is real feedback, but it needs translation. Ask what specifically feels wrong: sentence length, jargon, emotional intensity, repetition, weak evidence, or lack of a clear action.
Read
Prioritize the next repair
Fix high-impact problems first: unsafe disclosure, invented facts, misunderstood goals, and missing decisions. Then address structure, clarity, and tone. Cosmetic edits can hide deeper defects by making an incorrect answer sound finished.
When several issues exist, list them but select one as the next experiment. Save the original output. Controlled comparison is much easier when you can see what changed.
Checking tutor…