Research report · 21 October 2026
Copywriting: five pre-install evidence gates
For one approved, non-personal landing-page draft, an owner-maintained brief and reviewed prompt remain the first route. Copywriting is a conditional, source-inspected draft or edit candidate—not an installation recommendation. This ledger brings together the earlier single-skill documentation, source comparison, and host documentation findings. No new source or host was queried for this synthesis.
Read each gate at its evidence level
1. Intended files, not a reachable source page
The selected Copywriting text displayed 2.3.1 on 20 October and names three in-folder references and conditional marketing-context paths. The maintainer README advertises named-skill selection; its ZIP design describes a named folder with authored resources. These are source instructions and packaging claims, not an inventory of selected files. Before installation: choose an immutable revision and route; inspect the exact folder or archive manifest, frontmatter, references, scripts, assets, external paths, license and hashes. Without those bytes, stop.
2. Destination and invocation, not format alone
Codex local-skills and Claude Code skills documentation describe different placement and invocation routes. Neither documents this Copywriting payload as installed in a chosen host. Before claiming compatibility: a technical reviewer chooses one host and route, then verifies the destination, version/configuration, discovery, invocation and reference resolution for the intended bytes. No selected install or invocation was observed.
3. Effective access, not assumed prompts
The selected text conditionally asks for local marketing-context files, including Learnings; that instruction is not consent. Its observed frontmatter lacks allowed-tools, but the selected file is not a package security audit. Codex approval and Claude Code permission documentation show that local reads or tools can be available without a prompt for each action. Before any read or tool use: the data owner approves each context path and contents; the host administrator checks actual filesystem, script, connector, network and approval scope. No effective target policy was tested.
4. Brief, claims and rights, not a source self-check
Observed 2.3.1 offers a blank-brief draft, real-voice draft or point edit and directs supportable claims and an AI-tells check. This is source intent, not observed output quality. Its numeric efficacy statements and 3× example lacked checkable support in the earlier bounded review. Before a draft is accepted: use an authorized non-personal brief; a named marketing reviewer checks material claims, proof and qualifiers, use rights, brand fit and the exact CTA destination. Unsupported figures, testimonials, comparisons or guarantees stay out. AI-tells is optional style review, not validated detection or buyer-response evidence.
5. Publication needs a separate decision
The inspected selected-file text did not direct a CMS publish or account action; this does not constrain an uninspected package or host. A named reviewer must inspect the rendered page and its accessibility, claims, rights and CTA; the site owner then authorizes any exact publication change separately. No rendered page or publishing action was tested in this research. Keep output local and draft-only until that decision.
Research library · One-skill documentation · Selected-source delta · Host documentation