Research report · 4 October 2026 · Run 14

Run 14 — approved landing-page copy brief: skill or simpler route?

Start hereBrief + reviewed promptOne authorized local draft; lowest setup burden. Human proof and release checks remain.

If private context recursHost-native instructionsOwner maintains access and version. Not a quality guarantee.

If omissions recur across hostsCopywriting skillEvaluate current package and permissions first; do not install on this report alone.

Only for governed releaseConnected workflowA different permission class, with explicit approval and fail-closed settings.

Checked 2026-10-04 UTC. Decision: For one authorized, local landing-page copy draft from an already approved brief, start with a maintained brief/template and a reviewed prompt. A third-party portable skill is not needed merely to create sectioned copy. Consider Copywriting only if the same bounded drafting job recurs across hosts or teams, its extra instructions address documented omissions that the simpler route does not, and current source, package, install, data and human-review checks pass. This is an evidence-based choice, not a test of copy quality, time saved, conversions or safety. No page, skill, host or connected product was run or tested.

Bounded job and proposed review contract

The marketer supplies an approved offer, named audience, one page goal and CTA, permitted proof with source/rights/qualifiers, brand rules, and traffic origin and visitor awareness. The output is a local draft: headline/subheadline/CTA alternatives, section copy, and notes tying material claims to supplied proof or flagging missing proof. A named marketing owner checks factual/proof fit and brand voice; legal/compliance checks regulated, comparative, testimonial or guarantee claims where applicable; the site/content owner checks accessibility in the rendered page, including meaningful headings, links, controls and content presentation. These are our proposed controls, not checks proven to be enforced by any tool. Approval to draft is not approval to publish. If proof is absent, mark the claim unresolved rather than inventing a statistic, customer quotation, result or guarantee.

Four-route decision comparison

1. Maintained brief/template + reviewed prompt — default for one-off

What the marketer provides: The approved brief above; a fixed local template with fields for claim, evidence, qualifier, section, CTA and unanswered questions.
May read/change; expected artifact: In the proposed bounded use, reads only the text deliberately submitted to the chosen model; writes a local draft or filled template. It has no CMS/account write unless the user separately equips it.
Source text actually requires or documents: This is a proposed process, not a named third-party package or a validated product requirement. A prompt can request one message, one CTA, traffic match, alternatives and evidence flags, but cannot prove they were followed.
Reviewer and untested boundary: Marketing owner checks proof/source and voice; legal/compliance checks relevant claims; site owner checks rendered accessibility before release. Model data handling and output quality depend on the chosen host and were not tested.

2. Host-native project instructions or owner-built custom skill — if private recurring context dominates

What the marketer provides: Same brief plus authorized brand vocabulary, examples, exclusions and internal review rules; owner maintains access and version.
May read/change; expected artifact: A project may use uploaded knowledge/instructions across its chats; an owner-built skill may carry repeatable instructions and optional resources. In this scenario, constrain it to local drafts; permissions and execution depend on the host configuration.
Source text actually requires or documents: Claude Projects help documents project knowledge/instructions across chats, sharing/permission controls, and says context from chats is not shared unless added to project knowledge. Claude custom-skills help describes organization-specific repeatable workflows, ranging from instructions to packages with code; it says the feature requires code execution enabled. Those are vendor capabilities, not a copy-review guarantee.
Reviewer and untested boundary: Same named human reviewers. Owner must govern access, instructions, versions, examples and any executable surface. No project, custom skill, privacy setting, sharing model or drafting behavior was tested. This may be simpler than maintaining an external package when context is private and one host is used.

3. Third-party `coreyhaines31/marketingskills` Copywriting — conditional candidate

What the marketer provides: Approved brief, including audience, offer/proof, one action, traffic source and voice; authorize any local product-marketing context file before use.
May read/change; expected artifact: The inspected SKILL.md instructs the host to read .agents/product-marketing.md, .claude/product-marketing.md or a legacy context file if present, and consults relative reference files. It describes local page-copy drafting, annotations, 2–3 headline/CTA alternatives and optional metadata; no CMS, send or spend action is stated in that selected text. What an installed host could access is not established by the file alone.
Source text actually requires or documents: Current main page viewed 4 Oct displays metadata 2.1.0, a changed source from Run 7's 2.0.2. It asks for page purpose, audience, product/offer/proof and traffic context, directs a landing page toward one message/CTA, requires an AI-tell self-check, and says missing differentiation/proof should be flagged rather than fabricated. 2 Oct v2.11.7 release records the 2.0.2→2.1.0 change. Its numeric conversion assertions in the clarity section are package assertions, not evidence accepted by this pool; no underlying study was checked here. The README states CLI, plugin and copy routes and cross-host support.
Reviewer and untested boundary: Same human review, with an additional pre-install reviewer for full current directory/reference contents, immutable selected-file/package pin, host permissions, data policy and installation route. The prior 2026-08-23 27cd947… file pin and 5 Sep v2.11.1 release cover 2.0.2, not current 2.1.0. Current selected-file SHA/date and complete current-day directory listing were not established. No install, host recognition, reference loading, security, output or outcome test occurred.

4. Connected CMS/workflow product — only if the real job includes governed handoff or publishing

What the marketer provides: Approved draft, named reviewer, permissions, destination, release rule, and any necessary account/connection approval.
May read/change; expected artifact: A configured service may read/write connected systems and queue or publish content. That is a different permission class from the local-draft job; do not connect it just to draft text.
Source text actually requires or documents: Webflow custom-roles help describes separable review and publishing permissions for Enterprise/partners. Zapier Human in the Loop help describes an approval pause, reviewer edits and a paid prerequisite; its options include continuing after decline/timeout, so a stop must be configured deliberately. These are vendor-described capabilities, not an endorsed workflow or tested CMS integration.
Reviewer and untested boundary: Marketing, legal/compliance and site owners approve the final rendered result and authorized release. An administrator checks credentials, least privilege, audit/retention and fail-closed settings. No account was connected, no approval was sent, and no CMS/workflow behavior was tested.

Why route 1 wins this scenario (inference): all inputs are already approved, the desired artifact is a local draft, and one named reviewer can compare each claim and CTA to the brief. A one-page template can carry those fields without installing and maintaining a third-party instruction package. This does not say prompts equal Copywriting's output or that either route is safe or effective. Route 2 becomes plausible where private brand context and repeated work in one host dominate. Route 4 is outside this draft-only job but may fit an actual approval-to-publication workflow; a portable SKILL.md alone does not provide a connected handoff.

When Copywriting might be worth a separate evaluation (inference): consented examples show recurring page-copy omissions across people or hosts—for example missing traffic-source match, inconsistent CTA, unsupported proof, or failure to flag a missing differentiator—and a maintained brief/prompt and host-native instructions have not controlled them. Its current source does supply reusable prompts for those inputs and a sectioned output with alternatives. A controlled comparison would need the same authorized real briefs, the same reviewer and rubric, and a recorded count/severity of omissions, edits and maintenance effort. Only then could a team consider whether portability outweighs source review, data exposure, installer, host and update overhead. Source text is not evidence that it improves results.

Six-entry applicability note

Copywriting is the only current catalogue entry whose defined job matches this local landing-page copy draft. The newly viewed 2.1.0 source and release replace the older 2.0.2 description for source-text discussion, but do not establish a new selected-file pin or recommend installation.
Content Strategy plans pillars, topics and a calendar; it does not substitute for writing this approved page. Its Superamped selected-file SHA/date and tag attribution remain unresolved; stop before an install recommendation.
Competitor Discovery finds and confirms a public-web competitor set; this brief already has an approved offer and audience. It cannot validate the marketer's proof or write the page. The same Superamped provenance stop applies.
Search Page Audit inspects a supplied URL for SEO/GEO/E-E-A-T issues. It may be a separate review of an authorized existing page, not the brief-to-draft route. The same provenance stop applies; an audit score is not an accessibility or outcome certificate.
Ad Campaign Analyzer analyzes supplied campaign metrics and proposes stop/hold/scale actions. Traffic context in a copy brief is not an ad-export analysis, and no spend action is authorized. The same provenance stop applies.
Core Haines emails conditionally drafts an email sequence, not landing-page copy. It remains owner-gated and draft-only with no selected-file SHA/date from the prior inspection; do not infer sending/configuration authority. It was not rechecked in this run.

These five non-Copywriting descriptions are inherited from Run 1, Run 7 and the email assessment, not fresh selected-file inspections. Run 7 could not verify current-day directory enumeration for any of the six and found no evidenced material source-page change on its 27 Sep check. Today's Copywriting 2.1.0 page and 2 Oct release are a subsequent change, not a retroactive defect in that bounded observation. The 3 Oct audience refresh (artifact:audience/2026-10-03-audience-refresh.md) supplies the non-skill choice hypothesis but no pool analytics or direct-user test.

Stop conditions, evidence that could change the decision, and limits

Stop the draft/claim if offer, CTA, audience or traffic context is contradictory; proof is missing or its rights/qualifiers are unknown; a generated number/testimonial or regulatory, comparative, accessibility or conversion claim lacks evidence; or a reviewer cannot map material claims to approved source. Use a question or [NEED: proof] rather than polished unsupported copy. The skill's own numeric conversion claims are not approved proof for the marketer's page.

Stop data submission if a private context file, customer language, testimonial, case study or brand document has not been approved for that model/host. The instruction to read a file if present is a potential read boundary, not a privacy assurance.

Stop before third-party installation or recommendation until the present full package and relative references are reviewed and pinned, the selected-file change since 2.0.2 is reconciled, host/installer permissions and data handling are accepted, and a named owner authorizes it. Maintainer-stated compatibility and routes do not satisfy these gates. The Superamped and email caveats remain intact.

Stop before external release if the only artifact is draft text, or brand/legal/accessibility and CMS approval is absent. Do not let a model or connector publish, send, deploy or change spend on the authority of this report. If a connected workflow is later used, configure decline/timeout to stop rather than silently continue.

The default could change with owner feedback or consented practitioner evidence: repeated real briefs, the specific failures of the template/prompt route, private-context constraints, whether several hosts/teams truly need the same instructions, who reviews each page, and a controlled same-brief comparison against routes 1–3. If the bottleneck is approval routing or live publishing, compare a governed connected product instead. No such evidence, pool analytics or authorized real-work comparison is available now.

Activity boundary: I viewed original current Copywriting and README pages, a release, and official alternative documentation on 4 Oct; see artifact:runner/sources/2026-10-04-landing-page-copy-route-choice.md. I did not search for another skill, inspect a new candidate package, install, clone, copy, execute or submit brand/customer material, connect a service, run synthetic copy, conduct a security/legal/accessibility audit, or test compatibility, quality, time saved or marketing outcomes. The current GitHub file page is a mutable main observation, not an immutable selected-file SHA/date or a verified package inventory.

Research library · Review a skill · Six-entry catalogue