Good AI prompting is not about discovering a magic phrase. It is the practical skill of defining a job clearly enough that an assistant can produce something you can inspect and use.

The strongest prompts resemble good briefs. They explain the outcome, supply the necessary context, identify constraints and define the shape of the finished work.

Quick verdict: A useful prompt needs an outcome, context, source material, instructions, constraints, output format and quality check. Add examples only when the desired pattern is difficult to describe.

Why vague prompts produce generic work

Consider this request:

Write a LinkedIn post about productivity.

The assistant must guess the audience, purpose, point of view, evidence, length and tone. A fluent result may still be useless.

Now compare it with:

Write a 170-word LinkedIn post for owners of small service businesses.

Argument: productivity problems often come from unclear hand-offs, not lazy people.
Use the attached example from our onboarding process as the evidence.
Open with a recognizable work situation, explain one practical change and end with a question.
Use direct, calm English. Avoid hype, invented statistics and hashtags.

This prompt does not control every sentence. It removes unnecessary ambiguity while leaving room for a good draft.

Google's prompt design guide describes the same fundamentals: clear instructions, relevant context, constraints, examples and structured decomposition. Prompting remains iterative; the first version is a test, not a contract with guaranteed results.

The seven-part prompt

1. Outcome

Describe what successful work enables.

Weak:

Summarize this report.

Better:

Create a decision brief that helps the management team decide whether to continue the pilot.

The second request gives the summary a purpose. That purpose influences what should be emphasized.

2. Audience and situation

The same information needs a different explanation for a customer, engineer, executive or new employee.

Include only context that changes the answer:

Avoid decorative personas such as “You are the world's greatest genius.” A precise job and audience are more useful than exaggerated role-play.

3. Source material

Provide the facts the assistant needs. If accuracy matters, explicitly limit it to those sources:

Use only the material between <sources> tags. If a required fact is missing, write “not provided” rather than filling the gap from memory.

<sources>
[paste or attach approved material]
</sources>

Clear delimiters help separate instructions from data. This is especially important when source documents contain their own commands, templates or quoted prompts.

4. Instructions

Use verbs that describe observable work:

For complex work, separate the stages. Ask the assistant to first identify claims, then evaluate evidence, then write the conclusion. One enormous instruction often hides which stage failed.

5. Constraints

Constraints prevent common but unwanted behavior:

- Do not invent quotes, examples or numbers.
- Preserve product names exactly.
- Separate facts from recommendations.
- Flag contradictions instead of resolving them silently.
- Keep the response under 700 words.

Prefer positive instructions when possible. “Use short paragraphs and concrete verbs” is clearer than “do not be verbose.”

6. Output format

Specify structure when the result must enter another workflow:

Return:
1. a five-sentence executive summary;
2. a table with claim, evidence, source and confidence;
3. three unresolved questions;
4. a recommended next action.

For machine-to-machine workflows, use schema-enforced structured output when the platform provides it. Prompting a complex JSON shape is less reliable than a real schema.

7. Quality check

Tell the assistant what to inspect before returning the answer:

Before answering, verify that every number appears in the supplied sources, every recommendation follows from stated evidence and every requested section is present.

This does not guarantee correctness, but it makes the acceptance criteria explicit.

A reusable master template

# Outcome
[What should this work achieve?]

# Audience and situation
[Who will use it, what do they know, and what happens next?]

# Sources
[Paste or attach the authoritative material.]

# Instructions
1. [First operation]
2. [Second operation]
3. [Third operation]

# Constraints
- [What must be preserved?]
- [What must not be assumed or invented?]
- [Length, tone or policy limits]

# Output format
[Headings, table columns, file type or schema]

# Quality check
[Facts, calculations, citations and completeness to verify]

Delete sections that do not help. The shortest prompt that reliably produces an acceptable result is better than a giant template nobody maintains.

Four practical prompt patterns

Research

Research how passkeys affect account recovery for small organizations.

Prioritize standards bodies and primary vendor documentation. Distinguish passkey authentication from recovery procedures. Include publication dates and links. Do not treat a vendor claim as independent evidence.

Return a 700-word brief with: key findings, implementation trade-offs, unresolved risks and a source table.

For current or multi-source work, enable the assistant's web or research tool. A model cannot cite information it was not given or allowed to retrieve.

Editing

Act as a careful editor. Preserve the author's argument and verified facts.

Audience: non-technical founders.
Goal: make the draft clearer, shorter and more specific.

First list the three biggest editorial problems. Then provide a revision. Finally list any factual claims that still need a source.

Analysis

Analyze the attached customer-feedback export.

Treat each row as one response. Identify recurring themes, but keep positive and negative feedback separate. Include the number of supporting responses for each theme. Quote no more than one short representative sentence per theme and remove identifying details.

Return a table, three implications and the limitations of the dataset.

Decision support

Compare the attached proposals against these weighted criteria:
- implementation risk: 30%;
- total cost: 25%;
- required internal time: 20%;
- support: 15%;
- optional features: 10%.

Use only explicit proposal content. Show the evidence and calculation for each score. Mark missing information instead of awarding an average score. End with questions that could change the ranking.

When examples help

Examples are valuable when format, tone or classification boundaries are hard to explain.

Provide one or two representative examples:

Input: “The export takes forever but support fixed it quickly.”
Output: {"product": "negative", "support": "positive"}

Do not provide ten nearly identical examples. They consume context and may cause the model to copy superficial phrasing instead of learning the intended rule.

Improve prompts through error analysis

When an answer fails, do not immediately add more words. Identify the failure type:

Failure Better correction
Wrong facts Supply authoritative sources and require citations
Wrong emphasis Clarify audience, outcome and decision
Missing section Define an explicit output structure
Invented detail Add a grounding constraint and unknown-state label
Generic language Supply concrete examples and prohibited clichés
Inconsistent classification Define categories and boundary examples
Calculation error Require code/calculation steps and manual verification

Change one or two variables, rerun the same test cases and compare results. This is more informative than continually rewriting the whole prompt.

Build a small evaluation set

If a prompt will be reused, test it against five to ten inputs:

Score factual accuracy, format compliance, edit effort and safe handling of uncertainty. OpenAI, Anthropic and Google all emphasize iteration in their prompting guidance because model behavior must be evaluated on the task you actually care about.

Common prompting mistakes

Asking for hidden certainty

“Give me the definitive answer” does not create evidence. Ask for sources, assumptions, uncertainty and alternative explanations.

Mixing unrelated jobs

Researching, deciding, writing, formatting and publishing in one request makes failures difficult to diagnose. Split high-value work into stages with review between them.

Treating length as quality

Long prompts can contain conflicting instructions. Prioritize the outcome, sources, constraints and output. Remove background that does not change the work.

Skipping the final human check

Prompt quality reduces errors; it does not remove responsibility. Verify important facts, numbers, citations and external communication.

Final recommendation

Write prompts like compact operating briefs. Give the assistant the facts and boundaries it needs, make the deliverable observable and define how quality will be judged.

The goal is not a prompt that sounds sophisticated. It is a prompt that repeatedly turns known inputs into reviewable, useful work.

Sources and further reading