I've evaluated 500+ agent skills on a six-dimension rubric — trigger quality, structure, workflow design, content, engineering, and security. These are the ten I'd actually install, each with the one mistake people make when using it.
The skill ecosystem is a gold rush right now, and most of what's being panned is fool's gold. A skill is just a folder with a SKILL.md that teaches your agent a job — but the gap between a great one and a lazy prompt-wrapper is the gap between "wow, it actually works" and "why did it just delete my formatting."
After running 500+ skills through a structured evaluation, here are the ten that earned a permanent spot, with the trap most people fall into with each.
1. xlsx (official) — 9.8/10
Reads and writes Excel workbooks: formulas, formatting, charts.
Why it's the best: It verifies its own output. After every write, it recalculates and checks the result before claiming success. Most skills just say "done" and hope.
What people get wrong: Asking it to "fix the numbers" without opening the file afterward. Its verification catches formula errors, not intent errors — check that the right thing got computed.
2. docx (official) — 9.7/10
Word documents: tracked changes, comments, find-replace, styled output, TOCs.
Why: A task-routing table sends each job type to a proven recipe, and a render-to-image check catches layout breakage that text-only verification misses.
What people get wrong: Trusting tracked changes to be complete. Accept them in Word's review pane, not by reading the diff in chat — the visual check is the whole point.
3. pptx (official) — 9.7/10
PowerPoint creation and editing, template-aware.
Why: The engineering is exceptional — it knows the corruption patterns naive XML editing causes and routes around every one. Like xlsx, it verifies after building.
What people get wrong: Feeding it a company template with 40 master slides and expecting magic. Give it your simplest working template, not the one your design team made in 2019.
4. archify — 9.4/10
Generates architecture diagrams and flowcharts as self-contained HTML you can export.
Why: The engineering discipline is a masterclass — acceptance gates, subagent dispatch per diagram, state you can audit.
What people get wrong: Treating the output as a spec. It's a diagram — great for communicating a design, wrong as the source of truth for one.
5. baoyu-slide-deck — 9.3/10
Generates slide decks as styled images — 17 style presets, consistent visual identity across slides.
Why: A genuinely different philosophy from pptx: instead of wrestling PowerPoint XML, it renders each slide as a designed image. For "make this look good fast," it wins.
What people get wrong: Not realizing the output isn't editable PowerPoint. If your boss needs to tweak slide 12 after the meeting, use pptx instead. Pick your use case first.
6. prompt-master — 9.3/10
Turns a vague request into a production-quality prompt for any AI tool, via a 9-dimension intent extraction.
Why: Best trigger design I've evaluated — it knows when NOT to fire, which almost no skill gets right.
What people get wrong: Using it to write prompts for the skill itself. Meta-prompting a skill that will prompt an agent that will... just write the prompt yourself.
7. rag-architect — 9.3/10
Designs RAG architectures: chunking strategy, retrieval pipeline, evaluation methodology.
Why: Every implementation step ends with an assertion checkpoint. You can follow its work like a test suite.
What people get wrong: Expecting it to build the thing. It's an architect, not a contractor — it designs the system and tells you how to evaluate it. The building is still yours.
8. codex-ppt — 9.2/10
Builds presentation decks as unified image sequences, with per-slide subagents enforcing a style contract.
Why: A 30-slide deck keeps visual consistency where single-context generation drifts. The approval gates mean nothing ships without your sign-off.
What people get wrong: Skipping the configuration step. It needs an image-generation backend — spend the five minutes on setup or the output will disappoint.
9. humanizer — 8.8/10
Rewrites AI-sounding text so it reads naturally.
Why: It's built on a catalog of specific AI tells with before/after examples, not vague "make it casual" instructions. Fact-preservation checks stop it from rewriting your claims.
What people get wrong: Running it on text that's already fine. Humanizers are for first-draft slop — running it twice makes prose weird and defensive, not "more human."
10. pdf (official) — 8.6/10
PDF reading, merging, splitting, form-filling, extraction.
Why: Official skill with a gated form-filling workflow that verifies field mappings before writing anything.
What people get wrong: Expecting it to beat PDF itself. Scanned documents and odd encodings still fall over — that's the file format's fault, and 8.6 is the problem domain talking, not the skill.
What the other 490 skills taught me
Three patterns hold across everything I evaluated:
- Verification is the rarest feature. The official skills verify their own output; almost nothing else does. When a skill says "done," ask what evidence it has.
- The description IS the product. Half of all skills have trigger descriptions that either never fire or fire constantly. A skill that doesn't trigger when you need it is dead weight in your context directory.
- Popularity and safety are uncorrelated. The worst thing I found — credential-harvesting scripts behind an innocent description — sat in a repo with a five-figure star count and zero issues filed. Read the scripts before you install, or use something that reads them for you.
Every skill above was scored on the same public rubric, and each one's full scorecard — with written rationale per dimension — is at skill123.me. No black-box ratings.
What's in your top ten? Genuinely curious what I'm missing — the good stuff hides in small repos.
Top comments (0)