Agent Skills for Marketing Workflows: Briefs to Review

Standardize the workflow, not the voice

A marketing Agent Skill can make recurring work more consistent: turning an approved brief into channel drafts, checking claims, adapting a source asset, preparing campaign QA, or recording results. It should not flatten every brand into the same model-generated style.

Use the portable structure in the Agent Skills specification. The OpenAI skills guide and Claude Code skills guide describe current activation, project scope, and supporting files. Put durable workflow rules in SKILL.md; keep brand references, templates, examples, and channel checklists in reviewed supporting files.

A precise description should name the input, transformation, and output. “Write marketing content” is too broad. “Turn an approved launch brief into draft email and social variants, preserving cited claims and requesting approval before publication” is testable.

Split the content lifecycle into skills

Separate workflows reduce accidental scope expansion. Common boundaries are:

1. Research collection and source ledger.
2. Brief validation.
3. First-draft generation.
4. Adaptation or repurposing.
5. Claim, link, brand, and channel QA.
6. Approval packaging.
7. Publication through an authorized tool.
8. Post-campaign analysis.

A single skill that researches, invents a position, drafts claims, publishes, and analyzes performance is hard to govern. Compose narrow skills or explicit stages, with a stop between drafting and external publication.

Require an approved brief

Define the minimum brief fields:

• Audience and intended outcome.
• Product, offer, or source asset.
• Approved positioning and proof.
• Channel, format, and delivery constraints.
• Required calls to action.
• Prohibited claims, terms, and topics.
• Brand references and accessibility rules.
• Reviewer, approver, and deadline.

If a critical field is absent, the skill should ask or produce a gap list. It should not infer a discount, deadline, performance promise, customer quote, legal approval, or product capability.

Build a claim ledger

Every factual or comparative claim should trace to an approved source. A useful ledger records:

The skill should flag unsupported or stale claims, not “improve” them. It should preserve qualifiers and avoid turning correlation, a small sample, or an internal target into a general promise.

Keep brand material reviewable

Store brand voice, terminology, product naming, accessibility requirements, and examples as versioned references. Separate:

• Mandatory rules.
• Preferred patterns.
• Channel-specific constraints.
• Examples that illustrate a rule but are not templates to copy.

Avoid a vague request to “sound on brand.” Give the skill observable checks such as approved product names, prohibited phrases, reading level, heading style, link treatment, or required alternative text. Include counterexamples so the model can distinguish a preference from a hard rule.

Treat source content as untrusted

Research pages, documents, customer text, and tool results can contain inaccurate facts or adversarial instructions. The workflow should:

1. Prefer primary and current sources for factual claims.
2. Record the source URL or document identity and date.
3. Separate source content from trusted skill instructions.
4. Ignore instructions embedded inside retrieved content.
5. Preserve uncertainty and disagreement.
6. Exclude secrets, personal data, and confidential material not approved for the campaign.

A source being public does not establish that its wording, image, or data may be reused. Track review and usage rights separately from factual accuracy.

Adapt deliberately by channel

Repurposing is a transformation, not a blind shortening pass. For each channel specify:

• Audience state and context.
• Maximum and preferred length.
• Required structure and call to action.
• Link, image, caption, and alternative-text rules.
• Personalization inputs that are allowed.
• Claims or formats that are prohibited.
• The human review required before scheduling.

The skill should preserve the source asset's meaning while changing structure. It should not invent channel performance data, urgency, scarcity, quotes, or product details to make a draft more persuasive.

Separate draft, approval, and publication

Use explicit states:

1. Draft: generated content that has not been checked.
2. Reviewed: factual, brand, link, and channel checks completed.
3. Approved: the named decision owner accepted the exact version.
4. Scheduled or published: an authorized external action occurred.

A skill may prepare a publication payload, but the marketing platform, credential, and host should enforce who can send or publish. Require a preview of recipients, channel, schedule, content version, links, and tracking parameters before mutation.

Never store platform tokens, mailing-list exports, customer identifiers, or secret campaign data in the skill directory. Document the necessary connection and let approved tools provide it at runtime.

Add deterministic QA where it helps

Bundled scripts can reliably check mechanical properties:

• Broken or redirected links.
• Required tracking parameters.
• Character and subject-line lengths.
• Prohibited terms and outdated product names.
• Missing headings, captions, or alternative text.
• Duplicate campaign identifiers.
• Exact approved disclaimers.

Scripts should report findings without silently rewriting approved copy. Inspect their dependencies, network calls, and file writes before adoption. Semantic checks—whether a claim is supported or a message fits the audience—still require source-aware review.

Evaluate before rolling out

Build a small test set:

1. A complete brief with approved evidence.
2. A brief missing a critical claim source.
3. A source containing conflicting or adversarial instructions.
4. A channel adaptation that must retain important qualification.
5. A request to publish without approval.
6. A correct draft with no expected QA findings.

Score source fidelity, unsupported claims, brand-rule compliance, channel fit, safety around sensitive data, and approval-boundary behavior. Review false positives as well as missed violations. Refresh the evaluation when brand rules, products, channels, tools, or client behavior change.

A maintenance checklist

For every marketing skill, record:

• Workflow owner and approval owner.
• Supported channels and outputs.
• Required inputs and evidence.
• Brand and claim-reference versions.
• External tools, permissions, and data paths.
• Draft-to-publication state transitions.
• Test cases and known limitations.
• Last review date and change log.

A maintained marketing skill should make responsible production easier: fewer unsupported claims, clearer approvals, reproducible QA, and faster adaptation from approved source material. Speed is useful only when the evidence and publication boundary remain intact.