Changelog

All notable changes to the product-view skill are documented here.

This project follows Keep a Changelog and Semantic Versioning.

[0.1.3] — 2026-06-26

Changed

• Tightened the auto-activation clause in the description: it now scopes to user-facing feature explanations and explicitly excludes internal components (a service, a queue, an algorithm), which stay in code-language. A triggering eval (3 independent judges per query) found the broad "most software explanations" rationale occasionally misfired on internal-component questions like "summarize what this microservice does" (fired 1/3). After the change, that and the other boundary splits (rate limiter, recommendation algorithm) resolved to consensus "no" with no regression on genuine user-facing triggers.

[0.1.2] — 2026-06-26

Added

• New SKILL.md guidance for explain and reframe mode: ground before you translate. When the feature or bug actually exists, read the real artifact (code, ticket, trace) before describing it, so the product-language description is true rather than a plausible-sounding reconstruction from the feature's name. Surfaced by an eval run where the skill produced a generic flow instead of grounding in the actual feature — a confident-but-wrong journey is a fluency trap that's harder to catch than a leaked code-word.

[0.1.1] — 2026-05-07

Added

• New SKILL.md section: Stepping out of the view unprompted. Even with view on, the skill now permits Claude to surface a brief code-side aside when a critical technical detail would change the user's decision (misleading product view, infeasibility, security/privacy issue). High bar: the default is to stay in the lens; the aside is formatted as clearly out-of-lens and returns to product-view immediately.
• New EXAMPLES.md entry (#6): a planning conversation where the skill steps out of the lens for one paragraph to flag that the user was about to promise "real-time sync" to a customer when the implementation only supports eventual sync.

Changed

• Marketplace source now uses the explicit { "source": "github", "repo": "..." } object form instead of the "." shorthand, for compatibility with older Claude Code versions.

[0.1.0] — 2026-05-07

Initial release.

Added

• skills/product-view/SKILL.md — the skill itself, with auto-activation triggers, manual view on / view off toggle, three sub-modes (explain, reframe, plan), and explicit do-not-use-for clauses.
• references/strip-list.md — the strip discipline: code-language to remove, product-language to keep, judgement-call edge cases, substitution cheat-sheet, and audience-specific vocabulary.
• references/mode-selection.md — heuristics for picking explain / reframe / plan when intent is ambiguous, with worked examples.
• references/format-selection.md — narrative vs structured vs journey, with an upgrade/downgrade rule and anti-patterns.
• EXAMPLES.md — six before/after pairs covering explanation, reframing, planning, mid-conversation drop-in, a non-customer audience (a game player), and a counter-example showing correct non-activation on a regex performance question.
• Plugin manifest at .claude-plugin/plugin.json for marketplace install.
• GitHub issue templates for view-leak reports and activation misfires.
• Eight test cases under tests/cases/.