AI Writing Guardrails

What this page covers

This page is about AI in content work: drafting microcopy, error messages, KB articles, marketing copy, and any text that ends up in front of a person. It covers humans using AI tools to draft content, and it covers JMF’s own agents (Ask Hubert and successors) that generate content as part of a product.

For AI guardrails on broader topics — data handling, sanctioned tools, model selection, security review — see AI Toolkit › AI Guardrails. That page is the enterprise-level policy; this one is the writing-specific operationalization.

What AI can ship unreviewed

The rule for every content type is the same shape: who can finalize, and what review applies. JMF’s default leans conservative until measured experience shows we can loosen.

The matrix is intentionally conservative. Promote a content type to a looser column only after a documented pattern of clean AI output for that content type, not after a single good draft.

Prompt patterns

A good prompt to the AI tool includes the voice anchor, the structural pattern, and the constraints. Without those, the model defaults to generic SaaS copy. With them, the draft lands close to JMF voice on the first try.

Template

Context: I'm writing a [content type] for [product / surface].
Voice: JM Family — direct, plain, respectful. No em dashes. Sentence case for UI strings.
Pattern: [the structure: e.g., for errors, "what happened / why if it helps / what to do"]
Constraint: Under [word count] words. No marketing-speak.
Don't: [common drift modes to avoid — see Section 6]
Draft a [content type] for [specific scenario].

Examples

• Button label

  Context: I'm writing a button label for a vendor list view. Voice: JM Family — direct, plain. Pattern: action verb + object, two words. Constraint: must name the outcome, not “Submit” or “OK”. Draft three options.

• Error message

  Context: I'm writing an error message for an expired session. Voice: JM Family — calm, helpful, never blames the user. Pattern: what happened / why / what to do, in that order. Constraint: under 25 words. Draft one.

• Empty state

  Context: I'm writing an empty state for a watchlist with zero vendors. Voice: JM Family — direct, optimistic. Pattern: explain what would be here + how to get the first one. Don't say “No items.” Draft one.

Working with AI output

What to feed the AI

A model needs context to land close. Paste these into the prompt when relevant.

• Voice & Tone

  The prose section on pillars (Direct, Plain, Respectful), and the tone matrix row for the current context.

  Open Voice & Tone →
• Grammar & Mechanics

  The punctuation rules and the contractions table.

  Open Grammar & Mechanics →
• Anti-patterns (content-and-voice)

  Catalog entries that describe failure modes for voice and tone specifically.

  Open the catalog →
• The editorial export (planned)

  A structured JSON bundle of voice rules and examples Hubert ingests. Reference once shipped.

• Five on-voice examples from the same surface

  Models anchor better on five examples than five paragraphs of rules.

What to watch for in AI output

Common drift modes JMF rejects. Reviewers should flag these regardless of how well the rest of the draft reads.

• Em dash overuse. Models love them. JMF uses em dashes sparingly in prose and never in UI strings.
• Hedge words. “May,” “might,” “perhaps,” “consider.” Strip and rewrite as a direct statement or a real question.
• Cliches. “Seamless,” “robust,” “leverage,” “best-in-class,” “synergize.” Banned in the plain-words list.
• Overly formal phrasing. “In order to,” “prior to,” “at this time,” “please be advised.” Rewrite plain.
• Marketing-speak in product UI. “Unlock,” “supercharge,” “world-class.” Product UI is a workplace, not a launch event.
• Hallucinated product names or features. “Ask Hubie” instead of “Ask Hubert.” “JMF Portal” instead of “The Hub.” A fact-check pass beats a tone pass for these.
• Generic empty-state copy. “No items found” is a model default. Voice & Tone has the JMF replacement pattern.
• Disclaimers and apology padding. “We apologize for any inconvenience” adds nothing. Drop or replace with a recovery action.

Approved tools and data handling

The specific list of sanctioned AI tools and the data-handling rules around them are owned by JMF IT and security, not this page. The rules below are conservative defaults; verify against current enterprise policy before relying on them for sensitive work.

Review and revision protocol

For every content type where the matrix requires human review, a lightweight protocol keeps quality up without slowing teams down.

1. Draft. AI generates the draft from a prompt that includes the voice and pattern context.
2. Self-edit. The person who prompted the AI cleans drift modes (see “Working with AI output”) before sending for review.
3. Reviewer pass. A second person reads the draft against the relevant standards page: Voice & Tone, Grammar & Mechanics, or the anti-patterns catalog.
4. Send back vs. fix. If the draft has a structural problem (wrong pattern, wrong tone), send back to the author. If it is wording-level, the reviewer fixes in place and checks the item off.
5. Approve. One reviewer is enough for low-stakes content. Two reviewers for legal, compliance, or external-facing copy.

Turnaround targets

Guardrails for agents inside JMF products

The same content rules bind agents that generate text as part of a product. Ask Hubert is the first; future MCP consumers and custom agents will follow.

Where the rules live for agents

• kit-templates/AGENTS.md

  The canonical “read these before generating output” instructions that ship with the Agent Kit. Authoritative for agents inside consumer product repos.

  Open the Agent Kit page →
• Content & Voice anti-pattern catalog

  Agents should refuse to generate content that matches a known anti-pattern.

  Open the catalog →
• The editorial export (planned)

  A structured artifact downstream of Voice & Tone that Ask Hubert ingests. Ships in a follow-up.

Concrete expectations

1. Agents introduce themselves with their name (for example, “I’m Hubert”) only when context calls for it, not in every response.
2. Agents do not invent product names, feature names, or policy text. When asked about an unknown JMF term, the agent says so.
3. Agents apply the JMF voice (direct, plain, respectful) to their own utterances, in addition to content they generate on behalf of the user.
4. Agents disclose AI-generated content per the disclosure rules when the surface requires it.

Disclosure and traceability

When AI helps draft content, JMF’s posture is straightforward: be honest where honesty matters, and skip the labels where they would add friction without value.

Audit trail: every content type covered by this page should be traceable to a reviewer signoff. The mechanism (PR comment, ticket, document property) varies by surface; the requirement does not.

For auditors and risk reviewers

If you arrived to verify how JMF prevents AI from shipping bad content, three places hold the rules:

• This page

  The operational rules for content work.

• The Decision Log

  Records what we chose and why. Filter to content-and-voice for the writing-specific decisions.

  Open the Decision Log →
• The Anti-Pattern Catalog

  Documents the failure modes we actively reject. Filter to content-and-voice.

  Open the catalog →

Compliance escalation paths and approved-tool lists live in AI Toolkit › AI Guardrails (enterprise-policy scope). This page is the writing-specific operationalization of those rules.