Research report · 18 October 2026
What changed in the tagged emails instructions?
A complete selected-file comparison of 2.1.0 at v2.11.8 and 2.1.4 at v2.11.37 finds four changed areas: an outbound referral, reference-before-drafting routing, use of local Learnings, and implementation links rewritten to mutable main URLs. The bounded local-draft choice does not become a send plan.
The comparison boundary
Both tagged skills/emails/SKILL.md files were returned in full: 330 numbered lines plus a trailing newline at 2.1.0, and 342 plus a trailing newline at 2.1.4. Their Git blob SHA-1 identifiers are 6afcc8894a928653d7e7f8a0b8bbef871e08e32e and e660baf4adf7430ff90bfbdeaf9a9ad101a2d2ed, respectively. These identify selected source files, not a release ZIP or installed package. The 2.1.0 file-attributable commit and later tag-target commit 365b412152b9542499ca1439426f7fc3cc92bb65 are separate provenance points.
Six decision areas
| Area | 2.1.0 → 2.1.4 source observation | Practical boundary |
|---|---|---|
| Job and referral | The multi-email lifecycle job stays. The referral expands from cold outreach email to cold outreach and outbound sequences. | Route clarification, not new welcome-draft authority. |
Context and Learnings | Both versions request local product context and audience, trigger, relationship and goal questions. 2.1.4 additionally says to check the context file’s Learnings before drafting and apply active lessons unless instructed otherwise. | This broadens potential local and past-test data exposure. Approve a read first; mark missing lessons or results unknown. |
| Reference reads | A new top table routes tasks to matching references before drafting or audit, and says to combine flow and copy references where relevant. The three same-skill reference targets were already linked at relevant sections in 2.1.0. The Copywriting AI-tells pointer is not a new file established by this delta. | More explicit review work; neither link resolution nor package inclusion follows. |
| Deliverable, timing, metrics | The overview, per-email copy/CTA/destination, optional conditions and metrics-plan fields, generic timing guidance and seven-beat example remain substantively the same across 2.1.0 and 2.1.4. | Five emails is our proposed exercise. Timing, exits and metrics are draft questions, not configured rules or validated targets. |
| Implementation links | “For implementation” and six tool categories remain, but the registry and guide links move from relative paths to absolute github.com/…/blob/main/tools/… URLs. | Those targets are mutable and unpinned by the tag. Linked account, recipient-data and send routes remain a separate stop, not draft steps. |
| Proof and action gates | Inline AI-tell language and cross-skill pointer retain the same content; the new instructions add reads, not a proof ledger, named sign-off, consent check, ESP requirement or performance guarantee. See the 2.1.4 copy section. | Pool proof, authorization, recipient-policy and external-action gates are separate from source instructions. |
What to do with a local welcome brief
- Use the simpler first routeWrite an approved, non-personal brief in an owner-maintained template and try a reviewed prompt. A five-email Markdown outline is a review exercise, not a source requirement.
- Scope any later skill trialReview the intended package and host first. Approve exactly which local context,
Learningsand reference files may be read. Keep unsupported claims and missing performance facts visibly unresolved. - Keep action separateHave a named person check proof, terms and exact CTA URLs. Recipient-policy and ESP owners decide eligibility, consent, suppression, overlap, timing and every external email-system action.
A direct template-versus-skill comparison still needs a consenting, bounded brief, a defined rubric and named reviewers. Neither route was tested here.
Research library · Browse by job · Linked-route reading · Release-asset boundary · Tagged 2.1.0 file · Tagged 2.1.4 file