Prompt Engineering

A young woman with a pixie cut in a magenta turtleneck pins a detailed city map with a cyan route between a yellow and a magenta marker to the wall, while her other hand takes down a sheet of random scribbles. A courier helmet hangs on a hook to the right.

Writing AI task briefs is a four-part skill: describe who uses the result, what data is attached, what format to deliver, and what is off limits. Without these four parts, the person checking the output has no written criteria - only a feeling - and cannot tell a colleague what went wrong.

To apply the method to your own task, do four things: describe the task so a colleague could reproduce it without you, run the model on approved safe material, check the result against criteria you wrote down in advance, and record exactly what did not line up. Everything else - the template, the journal, the failure classes - builds on this cycle rather than replacing it.

The example below was built for this walkthrough. Dates and ticket numbers are invented; no real company data was used.

What does an AI task brief consist of?

An AI task brief is a short document that describes a task so clearly that another person could run it without you. A quick test: if you cannot hand the brief to a colleague, it is not finished.

A brief has four parts. Each one sets a boundary for the person checking the result, but none of them controls model behavior directly.

Goal and recipient. Who uses the result and what they do with it. An operations lead before a weekly meeting will not read an essay about company culture - they need a memo with bullet points for questions. But the recipient alone does not say which data exists and which does not.

Input. What is attached and what is not. Missing data will not appear on its own. Missing data is a data problem, not a wording problem.

Output format. A file, a table, or text with a defined structure. This field also names who receives the result and what they do with it next. Format does not check content: both fields can be filled with vague phrases if the acceptance criteria are not written separately.

Limits and stop conditions. What is not allowed: no forecast, no efficiency rating, no blame. Plus a rule for missing data: "if the data is insufficient for a conclusion, say so plainly - do not supplement from general knowledge." Even with this rule, a soft judgment can still slip into the output, such as "unfortunately, the situation does not allow..."

Training example: writing a task brief for a report

The input

Report on internal request handling, September 2026:

  • Request 01: completed, delayed because the request form lacked a date.
  • Request 02: awaiting confirmation of the responsible party.
  • Request 03: completed, delay reason not recorded.
  • Request 04: in progress, source document not attached.
  • No total request count for the month. No processing times.

The recipient is the head of the operations team. They need to decide which clarifications to request before the weekly meeting.

The completed task

Copy this task and substitute your own data.

Goal and recipient. The operations team lead picks which clarifications to request from handlers before the weekly meeting. The result goes to them, not to a shared chat.

Input. Report on internal request handling, September 2026. Request 01: completed, delayed because the request form lacked a date. Request 02: awaiting confirmation of the responsible party. Request 03: completed, delay reason not recorded. Request 04: in progress, source document not attached. No total request count for the month. No processing times. No other data.

Memo format. Two sections: "What the report states" and "What needs clarification." Each finding is tagged with a request number (01, 02, 03, 04). Do not fill in missing data. Do not calculate sums or percentages. Do not send the memo.

Limits and stop conditions. No forecast for next month. No team efficiency assessment. No blame. If the data is insufficient for a conclusion, say so plainly rather than supplementing from general knowledge.

A failed output and one fix

Here is an example output built for this walkthrough - not a real model run, but a sample for analysis:

"The main cause of delays is that employees forget to fill out request forms. Everyone needs to be required to take training."

Check against the facts in the input. Request 01 has a stated delay reason: missing date. Request 03 has no recorded reason - we do not know what happened. Requests 02 and 04 are unfinished. No monthly total. No processing times.

The "main cause" conclusion does not follow from four requests where only one reason is recorded. The "all employees" conclusion does not follow either: one case with a date issue, one with no recorded reason, two unfinished. The output violated the written condition: "if the data is insufficient - say so plainly, do not supplement from general knowledge."

One fix. Add to the task: "The word 'main' and any claim about a prevailing cause are only acceptable with a complete list of causes and comparable data. Otherwise list the known cases and what is unknown."

Corrected memo (also a composed sample, not a real model output):

Section "What the report states":

  • 01: completed, delay reason - missing date in request form.
  • 02: awaiting confirmation of the responsible party.
  • 03: completed, delay reason not recorded.
  • 04: in progress, source document not attached.

Section "What needs clarification":

  • Request 03: ask the handler for the delay reason.
  • Request 02: confirm who owns it and what exactly is needed.
  • Request 04: request the source document.
  • General: the report contains no monthly request count and no processing times. Those figures are needed for any workload assessment.

The fix makes an already-written condition more explicit. The effect needs to be tested on real runs.

A young man with curly hair in a yellow sweater stands at a standing desk with his back to the viewer. The monitor shows a chat window: a grey message block and a cyan reply, with a cyan check mark beside each of the reply's four lines.

How does a task brief transfer to a new input?

A task written for one document may contain hidden assumptions. To check whether it works in different conditions, take the same task and change one input parameter.

New input. Report on internal request handling, October 2026:

  • Request 01: completed, delayed because the request form lacked a date.
  • Request 02: awaiting confirmation of the responsible party.
  • Request 03: completed, delayed because of missing access rights to the system.
  • Request 04: in progress, source document not attached.
  • No total request count for the month. No processing times.

One part changed: request 03 now has a recorded delay reason. The recipient, format, and limits are the same.

Task for the October input (copy in full):

Goal and recipient. The operations team lead picks which clarifications to request before the weekly meeting. The result goes to them, not to a shared chat.

Input. Report on internal request handling, October 2026. Request 01: completed, delayed because the request form lacked a date. Request 02: awaiting confirmation of the responsible party. Request 03: completed, delayed because of missing access rights to the system. Request 04: in progress, source document not attached. No total request count for the month. No processing times. No other data.

Memo format. Two sections: "What the report states" and "What needs clarification." Each finding is tagged with a request number. Do not fill in missing data. Do not calculate sums or percentages. Do not send the memo.

Limits and stop conditions. No forecast. No efficiency assessment. No blame. The word "main" and any claim about a prevailing cause are only acceptable with a complete list of causes and comparable data; otherwise list the known cases and what is unknown.

Composed solution (training sample, not a real model output):

Section "What the report states":

  • 01: completed, delay reason - missing date in request form.
  • 02: awaiting confirmation of the responsible party.
  • 03: completed, delay reason - missing access rights to the system.
  • 04: in progress, source document not attached.

Section "What needs clarification":

  • Request 02: confirm who owns it and what exactly is needed.
  • Request 04: request the source document.
  • General: the report contains no monthly request count and no processing times. Two recorded causes (01 and 03) are not enough to compare their frequency - that requires data across all requests.

The "clarify" section lost one item: the reason for request 03 is now known. The task structure transferred without changes.

Four reasons AI task briefs fail

When an AI produces an unsuitable result, the cause usually sits in one of four places. This is a diagnostic framework, not an exhaustive list - other technical causes exist.

1. Undefined output. The task does not describe the result shape. The model generates something coherent but not what is needed. In this example: without "two sections, request numbers, do not send," the output could drift into recommendations and efficiency ratings. Fix: describe the structure specifically.

2. Incomplete or contradictory data. If something is absent from the document, the model may fill the gap with an assumption that looks plausible. In the training example: request 03 is completed but the delay reason is not recorded - the model might assume the same reason as request 01. The task says "if data is missing, say so plainly." That is an instruction, not a guarantee. If you need the data, get it from the source. If the document contains contradictory figures, add a stop condition: "on discovering a contradiction, report it and do not draw a conclusion."

3. Wrong tool. If the task requires exact calculation - average processing time from dates, for example - a language model is not the right tool. Calculation belongs in a spreadsheet or calculator. The model extracts data and explains results, but it does not replace arithmetic. If processing times are absent from the input, no tool recovers what is not there.

4. Task too large. When several steps with different inputs are bundled into one task, the model runs them in arbitrary order or mixes results. Fix: split into stages, each with a checkable output.

A separate issue: when key rules live in an attached file such as a policy document. That text is input data, not trusted rules. OWASP (Open Worldwide Application Security Project, a web-application security foundation) classifies this as indirect injection in LLM01:2025: an external document contains instructions the model may execute as commands. State key task rules directly in the brief. Security at the environment level requires separate measures: separating untrusted data from system instructions, restricting access rights, and requiring confirmation for risky actions.

Rewriting the brief does not fix the data and does not grant access to the system that holds missing numbers. When the cause lies there, rewording the brief is pointless.

What does the ten-field template for repeatable tasks include?

For a simple one-off task, four or five fields are enough: task and goal, input materials, required output, acceptance criteria, and forbidden actions. The full list is needed when a task repeats or must be handed to a colleague.

# Field What it describes
1 What to do and why What decision becomes possible once the result exists
2 What is attached and what is missing Versions, dates, file path, explicit note of absent data
3 The model's role in the task Summarizes, extracts, compares, reformats. What stays with the person
4 What output is needed File, table, text of a defined structure. Who uses it and what they do next. Save path
5 What is forbidden Do not add from general knowledge, do not conclude beyond the text, do not calculate without data
6 What makes the result acceptable Specific checkable conditions, not "well written"
7 How to verify the result Which source, calculation, or person confirms or refutes a finding
8 What to do on unknowns or contradictions Clarify before running, flag the problem, stop generation
9 Who accepts Who scores against criteria and decides whether to use the output
10 Where it is stored and how it is passed Task version, output, journal, recipient

How do you build the skill of writing AI task briefs?

The skill builds through a repeated cycle: describe the task, run it, check against criteria, record what happened. Below are the working cycle, the iteration journal, and the transfer check.

The working cycle

  1. Describe before running. Write down: what should appear, who uses it, what criterion makes the result acceptable. This step sometimes reveals that the task is phrased as an intention ("do an analysis") rather than a description of the required result. If you cannot answer - the task is not yet formulated.

  2. Fill the brief and run. Do not skip the "forbidden actions" and "acceptance criteria" fields even for simple tasks. The first training run uses approved safe material in an approved environment. Real working data is used only with explicit permission. Save the result with the tool name and version noted.

  3. Check and record. Compare against written criteria, not a general feeling of quality. A survey of 319 knowledge workers by Lee and colleagues (CHI 2025, a conference on human-computer interaction; 936 examples) records in self-reports a link between high confidence in AI output and lower stated effort on critical review. It is a correlation from self-reports, not proof of causation.

  4. Change one variable and repeat. If you rewrite the wording, add data, and change format at the same time, you cannot tell what mattered. Change one thing at a time. This is not a strict experiment: one repeat of a variable response does not prove cause. Save versions of the task, input, tool, and criteria.

The iteration journal

The journal holds records of each run: what did not work, why, what changed, and what the result was. Fill it after a real run, not on a training example. A minimum entry contains: the specific defect by criterion, a hypothesis about the cause, exactly what changed, where the input and saved output live, who helped, and the check result.

The entry below is a composed sample for illustration. This is not a real journal and not the result of a conducted experiment.

Defect: the word "main" appeared in the output applied to delay causes, though a reason is recorded in only one of four requests. The written condition "do not supplement from general knowledge" was violated. Unverified hypothesis: adding a clarification about the word "main" will help uphold that condition. Changed: added a rule about the word "main" and a complete list of causes. Input and output: the training report text and the first memo are saved as parts of the composed sample. Help: none. Check: the corrected memo lists four requests separately and makes no overall conclusion. Effect of the fix is not established.

When two reviewers disagree

When two people score the same saved output differently against identical criteria, the causes vary: different subject-matter expertise, different readings of one criterion, a missing criterion, or a random checking error.

Procedure: each person writes their decision and grounds before discussion, then both compare on a specific output fragment. If the cause is criterion wording, clarify it and re-check the same saved output. If a criterion is missing, write it down and re-check. If the cause is expertise, bring in a subject-matter expert.

A repeated model answer does not by itself confirm the previous one. A new request cross-checked against a document, source, or verifiable calculation can include a verification step; the basis of the check must be accessible to the person.

How to check whether the skill has taken hold

Ask the employee to write a brief for a real work task. A colleague who was not involved takes the brief, runs the task, and scores the result against the written criteria. Then both compare scores.

A disagreement in scores may point to incomplete or ambiguous criteria - but other causes are possible: different expertise, random error. Inability to run without asking questions may indicate incomplete input description or missing access rights. A short brief alone does not prove the employee lacks the skill.

Transfer check: hand the employee a new, unfamiliar document of the same class and see whether they apply the structure independently. A delayed check - run weeks later when the task recurs - answers whether the person retains the approach over time. Keep these two checks separate.

A sign the skill has transferred: the employee diagnoses the problem class and applies the matching fix, rather than cycling through rewordings.

What the method does not solve

This approach helps diagnose several possible causes: ambiguous output, incomplete data, wrong tool, excessive complexity. The brief does not remove model limitations, does not replace missing data, and does not grant access to the system that holds them. The task and journal structure is a training method for transferring expertise, not a standard with measured error reduction.

The Lee and colleagues study is based on self-reports from 319 workers and records correlations, not causal links. Its results do not prove that AI reduces critical thinking. OWASP LLM01:2025 describes indirect injection as a risk, not as a universally reproducible attack with measured frequency. Security measures require implementation at the environment level.

Overhead view of a white board with two sheets. The left one has uneven squiggles, some struck through and circled in red, and a hand with a red pen strikes out another line. The right one has three neat blocks of lines, each with a cyan check mark.

Frequently asked questions

Do I need to fill all ten fields for every task

No. For a simple one-off task, four or five fields are enough: task and goal, input materials, required output, acceptance criteria, and forbidden actions. The full ten-field list is needed for repeatable tasks that get handed to colleagues or become standing team instructions. Use the shorter version for one-off work and expand the template when tasks grow more complex.

What if input data changes every time

The task describes the input structure and rules for working with it, not the specific data. The variable part - the actual data - is substituted at each run. The fixed part - recipient, format, limits - stays the same. If the data structure itself changes from run to run, that is a signal to split the task into stages: first bring the data to a standard form, then run the analysis. Each stage is a separate brief with its own acceptance criteria.

Can I use an external policy document as a source of rules for AI

A policy text in an attached file is input data, not system rules. The model may follow it, but it may also interpret embedded instructions as commands. OWASP calls this indirect injection (risk code LLM01:2025 in their classification): an external document contains instructions the model executes as commands. State key task rules directly in the brief. Security against these risks is handled at the environment level: separating untrusted data from system instructions, restricting access rights, and requiring confirmation for risky actions.

What if after several iterations the result still does not meet criteria

Diagnose the cause first. The problem may be in the tool, the data itself, the criterion wording, or the task being too large. Several failed iterations are a signal to find the cause, not to rewrite the brief again. Choose the action that matches the cause: refine the brief, change the data, switch tools, or run additional training. Rewriting the brief does not fix the data and does not grant access to the system that holds missing numbers.

How do I know whether an employee has learned to write AI task briefs

Ask them to write a brief for a real work task. Then a colleague who was not involved takes the brief, runs the task, and scores the result against the written criteria. Both compare scores. A disagreement may point to incomplete criteria or ambiguity, but other causes are possible - different expertise, random error. An employee who has learned the skill diagnoses the problem class and applies the right fix, rather than cycling through rewordings.

What is the iteration journal and why keep it

The iteration journal holds records of each run: what did not work, why, what changed, and what came out in the end. A minimum entry contains the specific defect by criterion, a hypothesis about the cause, a description of what changed, where the input and output live, who helped, and the check result. The journal exists to show what actually affected the improvement. Without recorded iterations, it is hard to reconstruct what changed between runs and whether each change had an effect.

How do I handle long input documents when writing AI task briefs

When the input document is long, the model may process different parts unevenly. The brief should specify which fragment of the document is the source for each finding. If the document is very long, split the task into two stages: on the first, the AI extracts the relevant fragments; on the second, it analyzes them. Each stage is a separate brief with its own output and acceptance criteria.

Summary

Writing AI task briefs is a skill that comes with practice. Start with one real task: describe it in four parts, run it on safe material, check the result against written criteria, and record what did not line up. The template, the journal, and the ten-field structure extend this cycle - but only once the basic cycle is working.

To walk through task structure on your team's actual problems or build a training cycle inside your organization, write on Telegram @shimaoz or email hello@majento.ai.

More on building AI training inside an organization in the AI training section.

Sources

  1. microsoft.com publication
  2. genai.owasp.org publication

View as Markdown