AI-Assisted Writing Governance

AI Prompt Context Worksheet

Prepare a bounded AI-assisted writing task with approved source context, explicit instructions, prohibited inputs, output constraints, and human verification.

Free editable Markdown · Writers, editors, and AI governance teams · Updated July 30, 2026

Download Markdown

Accessible HTML preview

Blank template

The downloaded file contains the same fields in editable Markdown.

Task definition

Task name: [Enter a short identifier]
Responsible person: [Enter name]
Intended reader: [Describe audience and situation]
Permitted AI task: [Describe one bounded operation]
Final use: [Internal exploration / Draft input / Publishable after review / Other]
Successful output: [List observable requirements]
AI must not decide: [List judgments reserved for people]

Context inventory

Duplicate the source fields for each item.

Source ID and title: [Example S1 — Product guide]
Owner and permission: [Who authorized this use?]
Version or date: [Enter version/date]
Relevant passage: [Paste only necessary material]
Role: [Authoritative / Background / Disputed / Style example]
Required attribution: [Describe or write None]
Known limitation: [Describe scope or uncertainty]
Terms to define: [List domain terms]

Data and rights check

Prohibited inputs for this task: [List categories]
Approved service or environment: [Name and policy reference]

No passwords, access tokens, private keys, or security details are included.
Personal, customer, employee, and health information was removed or approved.
Confidential business information is permitted in the selected service.
Copyright, license, and contractual limits were checked.
The minimum necessary source text is included.

Prompt specification

Role or perspective: [Describe useful expertise without inventing authority]
Instruction: [Write the bounded task]
Facts that must remain unchanged: [List source IDs and facts]
Voice and terminology: [List applicable rules]
Scope boundaries: [State exclusions, dates, locales, or versions]
Required output format: [Table / Alternatives / Outline / Annotated rewrite]
Missing-information behavior: [Ask questions / Mark unknown / Stop]
Prohibited behavior: [No invented citations, no unsupported guarantees, other]

Verification plan

Claims to compare with sources: [List]
Calculations to reproduce: [List or None]
Independent checks: [List product, policy, or external checks]
Expert reviewer: [Name and question]
Disclosure requirement: [Needed / Not needed / To review]
Final approver: [Name]
Record location: [Where prompt, output, and decisions will be saved]

How to use this template

Describe one bounded task, reader, use case, and observable definition of success.
Inventory approved context and remove information that must not enter the selected service.
Specify facts, voice rules, scope, output format, and actions the system must avoid.
Plan source comparison, factual checks, expert escalation, and final human approval.
Save the completed worksheet with the resulting output and provenance record.

Define a bounded task with a testable result

Begin with the smallest useful job: propose three outlines from approved notes, identify unclear sentences, or rewrite one paragraph for a specified reading level. Avoid vague instructions such as “make this better,” which force the system to guess the reader, evidence, tone, and success criteria. State what the output will be used for, who will read it, which facts must remain unchanged, and what a successful response contains. Separate generation from decision-making. An AI system may offer alternatives, but a qualified person must choose the evidence, interpret policy, assess sensitive implications, and approve publication. A narrow task is easier to review and produces a clearer record when something goes wrong.

Curate context instead of uploading everything

Context should be necessary, authorized, current, and understandable without hidden assumptions. List each source and mark which passages are authoritative, background, disputed, or merely examples of style. Remove personal data, secrets, unpublished financial information, credentials, and licensed material unless the responsible policy and service terms explicitly permit their use. Do not assume that a paid account makes every input safe. If a task cannot be completed without restricted material, use an approved alternative workflow or stop. Explain important terms, dates, product versions, and geographical limits so the output does not merge facts that belong to different situations.

Design verification before requesting output

Write the review plan while the sources are visible. Identify facts that must be checked word for word, calculations to reproduce, citations to open, and claims that require an expert. Ask the system to distinguish source-supported statements from suggestions and to mark missing information rather than invent it. Output formatting can make verification easier: require a table connecting proposed claims to source IDs, a list of assumptions, or tracked alternatives instead of seamless prose. After generation, compare the response with the supplied material and independent evidence. Never treat fluent wording, citations, or a confidence statement as proof that a result is accurate.

See the fields in context

Fictional example: rewrite a setup warning

BlueHarbor is an invented application. The source text and workflow below are illustrative, not real product instructions.

Task: Produce two plain-language alternatives for one fictional setup warning without changing its required sequence.
Context: Approved source S1 says users must save an export before disconnecting; no customer information is included.
Boundary: The system may simplify wording but must not add compatibility claims or promise that data loss is impossible.
Output format: Two alternatives followed by a table showing how each instruction maps to S1.
Verification: A product owner checks the sequence, and an editor approves the final wording.

Frequently asked questions

Is a longer prompt always better?

No. Include the context needed to remove meaningful ambiguity. Excess, irrelevant, or conflicting material can make review harder and may expose information unnecessarily.

Can this worksheet be pasted into the AI tool?

Use it first as a human planning record. Paste only the approved prompt and permitted context, not internal risk notes or prohibited information.

What should happen when the sources conflict?

Label the conflict and require the output to preserve it. A person with the appropriate expertise should resolve or accurately communicate the uncertainty.

Does requesting citations prevent fabricated sources?

No. Open every citation, confirm that it exists, and check that it supports the exact statement. Prefer source IDs from a curated context packet when possible.