菜单
Run rendered visual QA on an `agent-slides` deck using LibreOffice slide screenshots, a scored checklist, and optional auto-fixes for common issues.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Build consulting-grade presentations from a natural-language brief using a 4-phase Pre-flight -> Storyline Review -> Build -> QA workflow on top of the agent-slides CLI.
Run one experiment cycle of the demo research loop. Reads program-demo.md for the current lane and hypothesis, applies a fix, builds all benchmarks, scores them, and records results.
Guide template ingestion end to end: learn a PPTX, inspect and edit the manifest, smoke-test template-backed layouts, then hand off to create-deck.
Modify an existing agent-slides deck in place. Inspect → smallest mutation → validate. Handles layout switches, slot edits, content changes, rebinding, and batch operations.
Review an `agent-slides` deck for storytelling quality, narrative structure, message coverage, and content density without drifting into rendered visual QA.
Run the final pre-share polish pass on an `agent-slides` deck. Checks structural completeness, source lines, formatting consistency, and content completeness, then applies only safe mechanical fixes.
| name | review-deck |
| description | Run rendered visual QA on an `agent-slides` deck using LibreOffice slide screenshots, a scored checklist, and optional auto-fixes for common issues. |
Use this skill when the user asks for deck QA, visual review, design review, screenshot-based critique, or wants to know whether a PPTX deck actually looks good after render.
In this repo, prefer uv run agent-slides ... so the checked-out CLI and assets are used.
For command semantics and error handling, use uv run agent-slides contract as the canonical contract.
This skill is the rendered counterpart to agent-slides validate. validate checks structural design rules. review checks the real rendered output.
This workflow depends on:
soffice on PATHpdftoppm on PATHThe CLI command handles the rendering pipeline:
uv run agent-slides review deck.json
That command:
.pptxpdftoppmreport.md, report.json, and slide screenshots into deck.review/ by defaultRun the work in four phases:
Prefer the shipped repo commands rather than inventing alternate entry points:
uv run agent-slides reviewuv run agent-slides validateuv run agent-slides buildStart by running:
uv run agent-slides review deck.json
Read report.md, then inspect the title slide and 2-3 representative content-slide PNGs from the generated artifacts directory.
Capture three quick judgments in plain language:
Do not jump into detailed fixes until this gut read is clear.
Use the generated screenshots plus the structured report.
Use the rendered PNG as the source of truth for what the slide actually looks like.
Treat report.json as the machine baseline and the slide PNGs as the visual evidence.
Use this 38-item checklist directly during the audit:
$2.5B, not $2,487,392,104)Your output should summarize:
Use this rubric for each category:
A = 0 failures in category
A- = 1 minor failure
B+ = 1 failure
B = 2 failures
C+ = 3 failures
C = 4+ failures
D = majority of items fail
F = category completely ignored
Calculate overall grade as a weighted average:
Prefer pointing to concrete slide files such as:
[screenshot: deck.review/run/slides/slide-03.png]
If the user wants only the findings, stop here.
Only use auto-fix when the user explicitly wants fixes applied. The explicit approval path is:
uv run agent-slides review deck.json --fix
Use an iterative fix loop, not a single cleanup pass:
Fix Loop (max 3 passes):
Pass N:
-> Fix top issue
-> Re-render affected slide via LibreOffice
-> Re-evaluate against checklist
-> Pass? -> next issue
-> Still failing after 2 attempts? -> flag as unresolvable, move on
Stop early when: all categories B+ or above, or 3 passes complete
Each fix produces a before/after PNG pair as evidence
Apply fixes in this priority order:
Within that loop, prefer these common mechanical fixes when they match the top issue:
title_content slides into a follow-up slideAfter --fix, inspect the new after/ screenshots and compare the before and after grades in report.json.
Document the iterative fix loop explicitly:
Fix Loop (max 3 passes):
Pass N:
-> Fix top issue
-> Re-render affected slide
-> Re-evaluate against checklist
-> Pass? -> next issue
-> Still failing after 2 attempts? -> flag as unresolvable, move on
Stop when: all categories B+ or above, or 3 passes complete
Loop rules:
A good run of this skill leaves behind:
report.mdreport.jsonIf the render pipeline is unavailable, stop and report the missing tool exactly as the CLI surfaces it.