Kami · Paper

Evidence shows significant effort toward least privilege and data-flow transparency: SKILL.md explicitly states the update check is read-only, at most once per day, fails silently; MCP server docs emphasize rendering only trusted local HTML; materials check requires user-provided assets, presenting gaps as a table rather than guessing; brand profile fills gaps silently. No red-line risks such as credential theft, exfiltration, or destructive defaults found. Deductions: the skill actively runs an external script (check-update.sh), although declared no data sent, unverifiable statically; MCP server allows rendering local HTML but does not explicitly restrict file access scope; dependencies on external fonts (e.g., TsangerJinKai02) from CDN involve network requests; no explicit rollback or undo procedure; publisher identity unverified, but only used as maintenance attribution, not a deduction by itself. Score 15: main risks visible but permissions, confirmation, rollback incomplete.

Static review observes high self-consistency in instructions: SKILL.md provides clear steps, template selection tables, verification commands (e.g., build.py --check-*), AGENTS.md documents known traps and CI gotchas, showing engineering rigor. But nothing executed, so key paths (e.g., render.py, build.py --verify) unconfirmed. Dependencies like WeasyPrint, Pygments, Marp CLI are optional, but no full fallback mechanism for installation failure (except font recovery script). Error handling: SKILL.md says external command failures should silently continue, but no clear, diagnostic error message templates. Static limit allows max 10: happy path plausible but tests, edge cases, error handling thin.

Use cases clear: eight document types and landing pages, audience and scenarios clear. SKILL.md provides detailed trigger terms in Chinese, English, Japanese, enabling precise triggering, avoiding false triggers. Environment fit: explicit support for Chinese, English, Japanese, Korean, with CJK font fallbacks; but font dependency on commercial TsangerJinKai02 and system fonts may affect Chinese rendering if user cannot install, and no mainland-network optimization (e.g., CDN accessibility) addressed. Capability boundaries well-declared: no hosted service, not a theme engine, progressive disclosure. Deductions: non-fit scenarios (e.g., simple plain-text formatting) not explicitly listed, may cause over-triggering; environment fit for mainland network with external CDN font downloads not risk-assessed. Score 13: clear scenario but some boundary/environment evidence thin.

Good information architecture: SKILL.md provides progressive disclosure, loading needed specs per document type; install/dependency notes detailed in README (plugin install, ZIP download); parameter and naming stability with script consistency checks; examples (assets/demos) and FAQ-like notes present. License explicit MIT, but no version or changelog in supplied files (VERSION exists but not shown); maintenance responsibility clear (one maintainer, contact in contact.md), but update path (e.g., PR workflow) not fully explicit. Deductions: version/changelog missing in provided files, likely in docs/release.md but not statically verified; known limitations like commercial font usage not emphasized in SKILL.md. Score 12: readable and usable but hidden assumptions or incomplete governance.

Static review cannot verify actual outputs, but SKILL.md describes output formats (HTML, PDF, PPTX) and verification commands with high plausibility. Value proposition clear: constraint system reduces inconsistency, claimed more efficient than manual or alternatives. But no actual generated documents (e.g., assets/demos PDFs) in reviewed material, so direct usability unconfirmed. Static limit allows max 7: completes core task but output completeness, direct usability, comparative-benefit evidence limited. Deduction: no execution evidence to confirm task completion.

SKILL.md and AGENTS.md reference CI workflows (.github/workflows/check.yml and release.yml) and test suite (scripts/tests/test_build.py), but these files not included in provided material, so cannot verify their existence or content. Provides reproducible verification commands (e.g., build.py --check), but no third-party execution evidence (e.g., CI logs). Thus, key claims (e.g., CI existence) lack cross-corroboration. Static limit allows max 5: auditable primary material or limited reproduction but thin coverage. Deduction: no third-party evidence, tests/CI not shown in files.