agent-reach
Self-contained multi-platform internet search and read skill. Zero external dependencies — calls upstream tools (curl+Jina, gh,…
Image path for vision tasks (rendered SVG screenshots)
$ npx -y skills add zengwenliang416/ppt-agent --skill gemini-cli --agent claude-codeHow it fires
How this skill gets triggered: by you, by Claude, or both.
/gemini-cliContext preview
The summary Claude sees to decide when to auto-load this skill.
Image path for vision tasks (rendered SVG screenshots)
name: gemini-cli
description: |
[Trigger] When PPT workflow needs SVG slide layout and aesthetic optimization via Gemini.
[Output] Structured optimization suggestions with scores and concrete layout/style improvements.
[Skip] For content authoring or SVG generation tasks (those are handled by Claude).
[Ask] No user input needed; invoked by review-core agent.
[Resource Usage] Use references/, scripts/ (`scripts/invoke-gemini-ppt.ts`).
allowed-tools:
- Bash
- Read
- Write
arguments:
- name: role
type: string
required: true
description: "reviewer"
- name: prompt
type: string
required: true
description: "Final prompt passed to Gemini"
- name: image
type: string
required: false
description: "Image path for vision tasks (rendered SVG screenshots)"SVG slide layout and aesthetic optimizer via `scripts/invoke-gemini-ppt.ts`. Gemini's role is to **optimize layout and improve visual aesthetics** — not just check compliance. It evaluates layout balance, visual hierarchy, color harmony, typography, and readability, then proposes concrete improvements. The script automatically tries fallback models if the primary model is unavailable.
npx tsx scripts/invoke-gemini-ppt.ts \ --role "<role>" \ --prompt "<prompt>" \ [--image "<path>"] \ [--model "<model>"] \ [--output "<path>"]
| Role | Purpose | CLI Flag | | -------- | ------------------------------------ | ----------------- | | reviewer | SVG slide layout & aesthetic optimization | `--role reviewer` |
Before calling Gemini, read the SVG file content and the relevant style YAML so you can include them in the prompt. The reviewer needs the actual SVG source code to inspect element attributes (font-size, fill, opacity, coordinates).
Construct a prompt that includes:
npx tsx scripts/invoke-gemini-ppt.ts \
--role reviewer \
--prompt "$REVIEW_PROMPT" \
--output "${RUN_DIR}/reviews/gemini-raw-${SLIDE_INDEX}.md"The script tries models in order: default → gemini-2.5-pro → gemini-2.5-flash. If all fail, it exits with code 2.
Whether from Gemini or Claude fallback, write the structured report to `${run_dir}/reviews/review-{nn}.md` using the output format defined in `references/roles/reviewer.md`. The Gemini raw output (`gemini-raw-{nn}.md`) is preserved as an intermediate artifact for traceability.
---
npx tsx scripts/invoke-gemini-ppt.ts \
--role reviewer \
--output "${RUN_DIR}/reviews/gemini-raw-${N}.md" \
--prompt "
## Task
Optimize the SVG presentation slide's layout and visual aesthetics. Your primary output is **typed optimization suggestions** — concrete, actionable design improvements. Scoring is a secondary quality gate.
## Slide Content
${SVG_CONTENT}
## Style Reference
${STYLE_NAME} style with tokens: ${STYLE_TOKENS}
## Optimization Criteria
1. Layout Balance: card arrangement, visual weight distribution, whitespace usage
2. Color Harmony: palette consistency, contrast ratios, accent usage
3. Typography: hierarchy clarity, font size appropriateness, line spacing
4. Readability: text legibility at presentation resolution, information flow
5. Information Density: content-to-whitespace ratio, cognitive load
## Output Format
Follow the exact structure from references/roles/reviewer.md:
1. **Optimization Suggestions** (primary): typed suggestions using the 5-type taxonomy (attribute_change, layout_restructure, full_rethink, content_reduction, deck_coordination). Each with type, priority (1-3), description, and type-specific details.
2. **Suggestions JSON**: parseable JSON array of all suggestions for downstream automation.
3. **Quality Gate** (secondary): overall_score (1-10), pass (true/false), per-criterion weighted scores, hard rule violations.
"---
The dual-model approach provides value through independent aesthetic perspective. **Aesthetic optimization requires cross-model review** — Claude cannot meaningfully optimize its own visual design choices (production data: 12 slides all scored 8.2-8.8 with identical notes during self-review).
When Gemini is unavailable (exit code 2), review-core performs **technical validation only**:
1. Run hard-rule checks from `references/roles/reviewer.md` Quality Standards table:
2. Report violations as pass/fail — **no aesthetic scores, no optimization suggestions**. 3. Mark output header as: `**Reviewer**: Claude technical validation (Gemini unavailable) — a
多智能体 PPT 幻灯片生成工作流,支持跨平台运行。 LLM 生成 + Gemini 审查,输出 SVG 1280×720 Bento Grid 布局的演示幻灯片。
Self-contained multi-platform internet search and read skill. Zero external dependencies — calls upstream tools (curl+Jina, gh,…