Practical guide

How to create a knowledge base article with AI

An AI-assisted knowledge base article should help a defined reader complete one task using approved, current evidence. AI can help organize source material, expose missing steps, generate test questions, and improve clarity. A named human remains responsible for source selection, factual accuracy, permissions, accessibility, publication, and maintenance. The durable artifact is not just prose:.

By Two Prune

An AI-assisted knowledge base article should help a defined reader complete one task using approved, current evidence. AI can help organize source material, expose missing steps, generate test questions, and improve clarity. A named human remains responsible for source selection, factual accuracy, permissions, accessibility, publication, and maintenance. The durable artifact is not just prose: it is the article plus its source map, test record, owner, and review trigger.

Freeze the reader task and source packet

Write the reader’s situation as an observable task: “A workspace administrator needs to restore access after a verified domain change,” not “Learn about access.” State the starting condition, intended outcome, prerequisites, exclusions, and the point at which the reader must stop and escalate.

Assemble a versioned source packet before drafting. It should contain the approved product behavior, policy, screenshots or interface labels, known exceptions, support escalation route, and last verified date. For every source, record an owner and whether it is authoritative, supporting, superseded, or unresolved. Do not ask an AI system to fill gaps from plausible general knowledge.

Digital.gov’s content guidance recommends plain language, structured and findable content, regular accuracy reviews, clear authority, appropriate language access, and manual validation of machine translation. It is federal digital-content guidance rather than a universal rule, but it offers a useful primary-source standard for task-focused web content. Digital.gov: An introduction to content

Sources: [1]

Use the TASK article method

TASK separates article construction from acceptance.

### T: Target one outcome

Give the article one primary task and one audience. Put the outcome, prerequisites, time-sensitive constraints, and safety or escalation boundary near the beginning. If the page needs several unrelated outcomes, split the navigation or create a parent guide instead of hiding several procedures under one title.

Write the opening answer so a reader can decide whether the page applies before following any steps. Use the same labels the interface uses. Define uncommon terms at first use. Avoid role labels that imply every reader has the same permissions.

### A: Anchor every consequential statement

Create a claim-to-source table before generating prose. Each policy statement, product limit, price, permission, deadline, data-handling rule, and irreversible action needs a current source. Distinguish source fact, instruction, inference, and local recommendation.

AI may propose an outline, compare source versions, flag inconsistent terminology, or identify claims lacking links. It should not decide which conflicting source is authoritative. Route contradictions to the product, policy, security, legal, or support owner named for that source.

### S: Sequence action and recovery

Order steps from the reader’s actual starting point. Each step should contain one action, its expected result, and the next decision. Show alternate paths only where the evidence changes the action. Put warnings before the action they qualify.

Include recovery for the most likely failure points: unavailable control, insufficient permission, stale screen, validation error, partial completion, or a result that differs from the expected state. Give the reader a stop condition and an escalation packet, such as the error text, timestamp, account identifier, and steps already attempted. Never request secrets or sensitive content that support does not need.

### K: Keep it verifiable and maintained

Test the draft with representative readers who were not involved in authoring. Ask each reader to locate the page, predict whether it applies, complete the task in a safe environment, explain the expected outcome, and recover from a seeded problem. Record where they hesitate, choose the wrong path, or need undocumented knowledge.

W3C’s supplemental cognitive-accessibility guidance recommends understandable words, short sentences and text blocks, literal language, unambiguous formatting, separated instructions, and explanations of implied content. The guidance is supplemental rather than a claim that one writing style satisfies every accessibility need. W3C: Use clear and understandable content

Assign an article owner, source owners, review date, and event triggers. Triggers can include a product release, policy change, recurring support failure, renamed interface control, source withdrawal, or accessibility finding. A page without an owner and trigger is a snapshot, not a maintained knowledge-base control.

Sources: [2]

Copyable article evidence card

| Field | Required entry | | --- | --- | | reader task | one starting situation and outcome | | audience and permission | who can act and required access | | prerequisites | verified state before step one | | approved sources | owner, version, authority, link | | consequential claims | claim, source, verification date | | procedure | action, expected result, next decision | | recovery | symptom, safe check, stop condition, escalation | | acceptance test | representative reader, environment, result, defects | | publication owner | named human accountable for release | | maintenance | review date and change triggers |

Keep this card with the article revision. It allows a reviewer to reproduce why the page said what it said, rather than relying on the author’s memory.

Worked example: restore access after a domain change

A support team asks AI to draft an article from six tickets and an old internal note. The draft looks polished but invents a 24-hour propagation window, uses an obsolete button label, and tells every reader to change organization settings.

Target: the team narrows the task to verified organization administrators restoring access after an approved domain change. Ordinary members are directed to an administrator. The page excludes identity-provider failures and suspected account compromise, which follow different routes.

Anchor: the author replaces ticket anecdotes with the current product runbook, permission matrix, and support escalation policy. The source map exposes a conflict between the runbook and the interface label. The product owner confirms the current label and schedules a runbook correction. The invented timing claim is removed because no approved source supports it.

Sequence: the article starts with prerequisites, then shows one action per step and the expected verified-domain status. A missing control routes to the permission check. A validation error routes to a safe retry once, then to support with the error text and timestamp. The article warns readers not to include passwords or authentication codes.

Keep: two administrators and one support agent test the article in a staging organization. One administrator follows an ambiguous alternate path, so the author replaces it with a decision table. The final record includes the three test outcomes, one repaired defect, the product owner, a 60-day review date, and triggers for permission or domain-flow changes. The article is accepted for this bounded task, not treated as proof that every access problem is solved.

Acceptance and hold rules

Accept the article only when every consequential claim traces to an approved source, representative readers can complete the bounded task, warnings precede risky actions, failure routes work, accessibility findings are resolved or explicitly dispositioned, and an owner accepts maintenance.

Hold publication when the product behavior is disputed, a required permission is unknown, the test environment cannot represent the procedure, the article requests unnecessary sensitive data, a serious failure has no safe escalation, or a translated version lacks human review. Record missing evidence rather than smoothing it into confident prose.

Two Prune can use knowledge-base authoring as a realistic assessment task and describe how a participant frames the reader need, navigates sources, uses AI, verifies claims, exercises judgment, and communicates instructions. It does not certify support quality, promise resolution rates, or make employment decisions.

Sources

Digital.gov: An introduction to content

  • W3C: Use clear and understandable content

Sources: [1] [2]

Sources

  1. 1.Digital.gov: An introduction to content · U.S. General Services Administration
  2. 2.W3C: Use clear and understandable content · World Wide Web Consortium

Explore this topic