Architecture learnings from competitive research (fixtures, deliverables, authoring conventions, README ergonomics)

Architecture learnings from competitive research (2026-07)

Actionable architecture/packaging patterns from the 10-library competitive review (archive: ~/.claude/skills/deep-research/references/research-archive/2026-07-24-claude-skill-libraries-competitive-learnings.md). The biggest single one — pressure-test / skip-resistance evals — is captured on #479. This issue tracks the rest so they're not lost. Grouped by effort.

Ties to existing issues (fold in)

• Canonical fixture company (Vellotti's TaskFlow pattern) — one fictional brand (product-marketing.md + facts + constraints) reused across demos, tutorials, and the eval framework's fixtures. Converges with #479 / #480 / #484 (Magister brand fixtures). Build once, reuse everywhere.
• Deliverables layer (Zubair's suite) — a branded-PDF report generator (tools registry) + a proposal skill to make our audit/analysis skills agency-sellable. Anchors on existing audit/audit-marketing, sales-enablement, client-intake.

Quick wins (README / repo ergonomics)

• README decision table — "Skills vs Prompts vs Projects vs MCP vs Subagents" (from Travis's awesome-list) to orient newcomers.
• Security / trust note — "these skills can invoke CLI tools + MCP; review before installing." We have none; the top awesome-list treats it as table stakes.
• Single-skill install UX — document npx skills add coreyhaines31/marketingskills --skill <name> + keep an ALL_SKILLS.md index (Bajaj).
• Get listed in travisvn/awesome-claude-skills (Community → Collections) — free, high-visibility backlink. (Distribution, not architecture — do it.)

Skill-authoring conventions (evaluate, then encode in skillify / a meta-skill)

• "standalone mode" section in skill authoring (Feld) — skills degrade gracefully without an MCP connection. Novel portability lever.
• Approval gates + "edit resets approval" (Feld) — any skill that sends/publishes (emails, social, typefully-adjacent) requires explicit sign-off, and editing the draft re-triggers it.
• Voice-profile + recent-message tone calibration (Feld) — load brand voice + pull recent messages to the recipient before drafting. Applies to brand-voice, emails, cold-email.
• Meta-skill / router — formalize skillify into a "skill that writes marketing skills" (consistent contributions) + a skill-navigator-style router (Bajaj, obra).
• Named-artifact convention (Zubair) — skills emit predictable output files (MARKETING-AUDIT.md, etc.) so multi-skill workflows chain.
• Enforced workflow spine (obra) — chain product-marketing → positioning → content-strategy → copywriting → launch as explicit hand-offs, not just "see X."

product-marketing context options (evaluate)

• Granular 3-file context split (King/PMA: about-me / voice-and-style / working-rules) as an option alongside our single product-marketing.md.
• HANDOVER.md session-continuity doc (King) — complements our context-doc version history.

Moat reminder

Our differentiators the competitors lack: tools registry + open eval framework + curation. Lean into those; don't chase raw skill count.

From the 2026-07-24 competitive-learnings research brief.