Best Office Claude Skills I use in 2026: less prompting, more workflow
I know this sub is mostly coding, but I think office workflows are an underrated use case for Claude Code-style agents.
A lot of “office AI” is just one huge prompt wrapped in a UI. This repo is more interesting because it treats office tasks as reusable skills with files, intermediate states, and recovery paths:
https://github.com/OpenSenseNova/SenseNova-Skills
Not affiliated. Just thought it was worth starring.
01. Image & Visualization
sn-infographic turns a text requirement into an infographic workflow: analyze content, choose structure, pick layout/style, generate the visual.
sn-image-imitate is the more technical one. It extracts a layout blueprint from a reference image, locks that structure, rewrites the content, generates, then runs VLM review + retry.
That is a much better abstraction than “please copy this style”.
02. Presentations (PPT)
The PPT system has a clean split:
sn-ppt-entry: parses the task and materialssn-ppt-standard: editable business PPTXsn-ppt-creative: full-bleed visual presentation pages
I like that it distinguishes editable decks from visual decks. Most generators confuse those two and produce something that looks okay but is annoying to maintain.
03. Data Analysis (DA)
The Excel workflow is where the repo feels practical.
sn-da-excel-workflow routes based on row count:
<10k: direct read10k-100k: Parquet cache100k+: streaming large-file analysis
The large-file path uses openpyxl read_only + iter_rows, chunked Parquet, type downcasting, and vectorized operations.
That is the kind of implementation detail you only add after you have seen spreadsheet workflows break in real life.
04. Deep Research
sn-deep-research is file-oriented:
request.md -> plan.json -> sub_reports/*.md -> synthesis.md -> report.md
This is probably the best design choice in the pack.
The agent does not just search and summarize. It plans, researches by dimension, stores sub-reports, synthesizes, then writes the final report. If the run dies, the task can resume from the existing artifacts.
That is how long-running agent workflows should work.
05. Search
Search is split by evidence type:
- academic
- code / technical
- English social
- Chinese social
- general web
This matters because “search” is not one operation. Different sources have different reliability, freshness, and failure modes.
Why I care
The pattern here is useful beyond this repo.
Office work has a lot of implicit SOPs: analyze this file, make charts, write the report, convert it into slides, add visuals, revise. Encoding those SOPs as skills is much more robust than hoping one prompt does everything.
This is the kind of repo I star before I need it, because the day I need it is probably 20 minutes before a meeting.