Choose and review

Choose a route before a skill

Start with the smallest authorized route that yields a reviewable result; compare the four active packages only when they address a documented gap.

Do you need a skill?

1 · One-offBrief/template + reviewed promptApproved input → local draft or table; named owner checks claims and gaps.

2 · Private recurring contextHost-native or owner-builtOwner versions context and access in one host; each output is reviewed.

3 · Portable repeated jobThird-party SKILL.mdOnly assess after documented omissions and a full package, installer, host and data review.

4 · Live operationGoverned connected productAdmin approves least privilege, logging, costs and fail-closed human release.

This route rule is an inference, not a comparative test. Run 15 gives inputs, outputs, reviewers, unknowns and the six-row matrix.

Four-skill chooser

Draft a marketing page

For one approved local landing-page draft, start with a maintained brief/template and reviewed prompt. Consider Copywriting only if recurring, documented omissions across hosts justify its additional package and review burden. Observed 2.3.1 on 16 October asks for a page draft or narrow edit depending on input; a point edit calls for changed lines and brief reasons, not default full-page alternatives. The complete 2.1.0→2.2.0 intermediate delta is unavailable, although full 2.1.0 and observed 2.3.1 endpoints were compared. The source directs claim-support/decline handling and a pre-handoff AI-tells check. Those source instructions are not the pool’s claim/proof/rights/CTA gates or a validated detector; the style check remains optional here. Hold unsupported numbers, proof, testimonials, comparisons, guarantees and CTA links; do not read unapproved local context. A named owner reviews the rendered page and separately approves publication. See Run 22’s claim/proof audit and Run 14’s four-route comparison.

Do not choose: when you need a plan, web research, an audit, campaign analysis, or direct publication.

Plan content work

For one approved first-month plan, begin with a maintained, versioned brief/template. Consider Content Strategy only as a conditional planning scaffold after source, data, host and reviewer gates.

Its 40/30/20/10 weights and X/10 labels are unvalidated preferences. Keyword stage, competitor gap and “top-performing” claims need dated, scoped evidence. Stop before install advice while selected-file SHA/date and current inventory remain unverified. No publishing action follows a draft. Read Run 19.

Find competitors

Consider for this bounded artifact only after source, data, host and human-review gates: Competitor Discovery. Use it for a public-web starting set that a person will confirm. A versioned owner sheet is the first route for one bounded set. The source ranks candidate frequency, but that is not a verified threat or action priority; it normally excludes tangential results. Keep all types provisional and stop before installation advice while provenance and inventory remain open. Read Run 20.

Do not choose: when web fetches are not allowed or current claims will not be reviewed.

Draft a lifecycle-email sequence

Owner template first: For one approved local draft, use an owner-maintained brief and reviewed prompt. Consider emails conditionally for our proposed five-email post-signup welcome hypothesis from non-personal, approved material.

Stop: Timing, exit and metric fields are hypotheses, not configured rules or targets. Omit unsupported proof, urgency, guarantee or promise; a named owner approves claims, terms and exact CTA URL. Tagged 2.1.4 adds reference-before-drafting and local Learnings instructions versus complete tagged 2.1.0; authorize any such read separately. Its implementation URLs point to mutable main. Do not read unapproved local context or follow implementation guides for this draft. Recipient-policy and ESP owners decide actual eligibility, overlap and schedule; every import, setup, test or send needs separate approval. Run 21.

Read the full comparison, access details, and review boundaries.

Before you install or share data

Email boundary: Mutable emails source displayed 2.1.3 on 14 Oct; the 2.1.0 change commit is historical, but its latest selected-file SHA/date, full 2.1.1→2.1.3 diff and inventory remain unknown. The five-email limit is ours, not the source’s fixed length. Timing, exits and metric ideas are unvalidated hypotheses. The conditional emails candidate is draft-only. Do not load local product context without data authorization. The linked Copywriting AI-tells page is style guidance outside emails; the separate tool guides include credentials, recipient data and send routes. GitHub lists an emails-source.zip asset for the 15 Oct v2.11.37 release and reports its SHA-256. Tagged source displays 2.1.4 and names three same-folder references, a Copywriting AI-tells cross-reference and separate implementation URLs. The ZIP bytes and manifest were inaccessible, so inclusion or exclusion of those paths, CLI-selected payload, current mutable file state and host resolution remain unverified. Read Run 27. Read Run 28 selected-file comparison. Source reading authorizes no implementation. A draft never authorizes an email-system action.

First stop: Copywriting’s mutable selected text displayed 2.3.1 on 16 October; 2.2.0 was observed on 15 October. Its latest selected-file SHA/date, exact full 2.1.0→2.2.0 delta and current inventory/payload remain unverified. Full 2.1.0 and observed 2.3.1 endpoints were compared. Read Run 26. The 2 October file-attributable 2.1.0 change commit and older 2.0.2 pin are historical. The current source directs a pre-handoff AI-tells self-check; this catalogue treats it only as optional style review, not detection evidence. The numeric conversion assertions and 3× example lack a checkable underlying study in the inspected passages; no install or output-quality test was run. The 25 September recheck did not establish selected-file SHA/date or tag findings for the four Superamped entries examined then. Two remain active cards and retain that stop; Search Page Audit and Ad Campaign Analyzer are removed and not recommended for installation.

Check	Continue only when	Stop when
Source	You can inspect the original repository, selected files, license, included resources, intended revision, and host documentation.	The selected file cannot be tied to the intended revision. This applies to the two active Superamped candidates; the other two have been removed.
Access and cost	The host, internet use, credentials, tools, and possible cost are documented or recorded as unknown.	It requests a key, account, connector, paid service, credit, or connection without specific approval.
Data and web content	You can provide only authorized, necessary material and treat fetched web content as evidence, not instructions.	It would receive unapproved customer, campaign, revenue, pricing, staging, or protected-URL data—or content asks for a new action outside the task.
External action	A named, authorized person will verify output and approve the exact target and change.	A workflow would publish, send, deploy, or change spend without that approval.

Read the provenance recheck. · Read the changed-source check. · Read the complete stop checklist and narrow host/install findings.

Skill catalogue · Browse by job · Research library · Run 1 source inspection