Changelog

Unreleased — 0.3.0: guidance MCP server and the promptingcraft identifier

• Renamed the skill and plugin identifier from promptcraft to promptingcraft (skills/promptingcraft/, $promptingcraft, /promptingcraft); MCP record kind promptingcraft_methodology and ledger fields promptingcraft_* (ledger schema_version 2). Migration: docs/migration-0.3.0.md. Historical evidence keeps the old identifier.
• PromptingCraft.com guidance MCP server (mcp/, Cloudflare Worker at https://mcp.promptingcraft.com/mcp): search_guidance, get_guidance, check_source_freshness; read-only, fixed-category inputs, no logging.
• Combined plugin: root mcp.json, review test cases (5 positive, 3 negative) and release notes; the standalone skill keeps working without the server.
• Fallback: when the guidance tools are absent or fail, the skill says so and labels replacement evidence by how it was obtained (evals/mcp/cases.json).
• No tag or release created.

Unreleased — EPIC 8: public launch and marketplace readiness

• Cursor as a target agent: references/cursor.md (Auto treated as an unknown model), cursor.com docs classified agent-product (cannot back model prompting advice), install_skill.py --host cursor, and native Cursor CLI activation results (evals/results/cursor-trigger-1/: automatic 0/30 with Auto, named 2/2).
• Listing metadata: capabilities (6), keywords (11), new starter prompts, long description naming Cursor, the published evaluation and the absence of a prompt backend.
• .cursor-plugin/marketplace.json for the single root plugin; docs/cursor-submission.md.
• Published A/B evidence: grader verdicts and prompt, run log and runner (scripts/evaluate_ab.py); trigger eval host failures corrected to 5 of 65 attempts.
• Worked examples: each lists its contract; new Cursor investigation, prompt review, no-change and official-guidance examples (authored illustrations, not eval results).
• Package stays 0.2.0 (unreleased); no tag or release created.

Unreleased — Agent Skills standard audit

• skills-ref validate passes for the source skill and both ZIPs; specification, progressive disclosure and resource references comply (docs/agent-skills-audit.md).
• Activation eval per agentskills.io: 20 queries, fixed train/validation split, 3 runs each, counted only from actual SKILL.md loads (scripts/evaluate_triggers.py, evals/trigger/queries.json, evals/results/trigger-1/). Description unchanged.
• Native reviewer-case outputs and a blinded with/without comparison (evals/results/review-native/, evals/results/ab-1/).
• agentskills.io added as a separately classified official standards source (classification: standard, role standard), never prompting guidance.
• No SKILL.md instruction change; daily freshness schedule stays inactive.

Unreleased — 0.2.0 candidate: OpenAI Plugins Directory readiness

• Listing metadata for a Skills-only submission: category Developer Tools, short description "Source-backed agent prompts", a clearer long description with explicit limits and independence, and three starter prompts (no @mentions).
• The validator enforces directory rules offline: category list, at most three unique one-line starter prompts of 128 characters, no screenshots, and no apps, hooks or MCP settings; package.py checks the plugin ZIP against the skills-only archive rules.
• Reviewer-reproducible cases in evals/review/cases.json (6 positive, 4 negative; P7, a Cursor agent task, was added later for the directory submission) and the field-by-field submission guide docs/openai-submission.md.
• No skill instruction change, no onboarding skill, no MCP server, no release or submission.

Unreleased — 0.2.0 candidate: PromptingCraft.com brand and plugin readiness

• Public brand is PromptingCraft.com (https://promptingcraft.com). The skill and plugin identifier stays promptcraft for compatibility with existing installations, so invocations, archive names and the generated dist/promptcraft-plugin are unchanged.
• ChatGPT listing metadata: display name, 30-character short description, honest long description, website, support, privacy and terms URLs, brand colors, and light/dark icons and logos packaged under assets/. The validator checks text limits, https URLs, package-relative square SVG/PNG assets, brand-color contrast and skill/manifest display consistency.
• Canonical repository URL becomes https://gitlab.com/revanalex/promptingcraft-skill.
• Minor version: new metadata and packaged assets, no identifier or behavior change. No directory submission, MCP server, tag or release.

Unreleased — contract fidelity and native validation

• Preserve expressly immutable text verbatim; do not infer whole-file freezes.
• Preserve discussion/implementation mode, unresolved choices and planning freedom.
• Reject invented API exceptions and blanket overrides of host/project authority.
• Allow no-change when the original task is suitable; keep necessary corrections.
• Optional checker adds exact-literal, local action and no-change identity checks.
• Frozen A/B/C source experiment, stricter artifact acceptance, observed tool/source metrics and honest timeout collection; no effectiveness claim from source quality.
• Native packaged Codex evidence and evaluation results are recorded separately.
• Canonical v0.1.0 release, tag and public assets remain unchanged.

Unreleased — official knowledge and contract quality

• Curated, attributed official recommendations with applicability and limits.
• Five requirement classes and bidirectional final-clause provenance review.
• Optional offline annotation-assisted contract checker; semantic review required.
• All-case baseline/candidate evaluation, four realistic public scenarios and a bounded isolated downstream protocol; results reported separately.
• Preserve portable packaging while omitting Python bytecode caches.
• No new release/tag/version or hosted API dependency.

0.1.0 — 2026-10-12

• One portable prompt-authoring skill for Claude Code and Codex handoffs.
• Contract extraction, applicable guidance, bounded critique and drift check.
• Original provider summaries, sources, rubric and worked examples.
• Read-only source freshness comparison and index candidate discovery.
• Deterministic fixture checks, isolated installation and reproducible archives.
• Synthetic behavior evaluation protocol, with results reported separately.
• GitLab canonical CI/review route; explicit GitHub mirror, no Actions.

Developed by the dondego.com team.