How to use the Lark Sheets Editor agent skill (lark-sheets)

What is an agent skill, and what does lark-sheets add?

An agent skill is a folder of instructions and reference files that an AI agent loads when a task matches it. The lark-sheets skill is one of these. Its main file is SKILL.md, and it points to a set of reference files that cover specific operations such as charts, pivot tables, filters, formula verification and history rollback.

Once installed, the agent knows how to work with Feishu/Lark spreadsheets through lark-cli. The source says it can read and write cells, formulas and styles, sort and filter rows, build charts, pivot tables and conditional formatting, search and replace text, check formulas, and review or roll back version history. It can also import and export local Excel and CSV files.

Who should use it, and who should not

The skill fits teams that keep budgets, reports or trackers in Feishu/Lark Sheets and want an agent to update them. It also fits analysts who need charts, pivot tables or formula checks built in an existing Lark spreadsheet, and users who want an agent to translate or classify text columns with native AI formulas.

It is not the right tool for Feishu multi-dimensional tables (Base/bitable); the skill says to use lark-base for those. It is also a poor fit for sensitive spreadsheets where you do not want an agent writing without review, because the skill can revert history and run batch updates. The skill cannot work offline, so you need a Feishu/Lark account and the lark-cli setup.

• Use it for: Feishu/Lark Sheets that teams update regularly, charts and pivot tables inside an existing sheet, formula checks and migrating Excel formulas.
• Do not use it for: Base/bitable tables (use lark-base), finding spreadsheet files by name (use lark-drive), or sensitive sheets that need human review before any write.

Install the lark-sheets skill

Download the skill ZIP from this page and unzip it into one of two places. For a personal setup, use ~/.claude/skills/. For a single project, use the .claude/skills/ folder inside that project.

If you use the Claude apps instead of a file system, open Settings, go to Capabilities, then Skills, and upload the ZIP there.

The skill also depends on two other pieces: the lark-cli command-line tool and the lark-shared skill, which handles authentication and permissions. Install both before you start. You can check the command reference with the command below.

• Required: lark-cli command-line tool
• Required: lark-shared skill (authentication and permissions)
• Required: a Feishu/Lark account with access to the target spreadsheet

How do you trigger the skill with a natural request?

You do not need to name the skill. Describe the spreadsheet task in plain language and mention the sheet or its link. The skill's routing is based on the task and on the URL path pattern and token, so a link to a Feishu sheet is the clearest signal.

The skill also says it should be used when you give a doubao.com /sheets/ URL or token, instead of falling back to WebFetch. Routing depends on the URL path pattern and token, not the domain.

If you want to find a spreadsheet by name first, ask for that explicitly. The skill tells the agent to use lark-drive for keyword search across cloud storage.

• Update a budget: "In this Feishu sheet, fill the Q3 column with formulas that sum the monthly rows."
• Build a chart: "Make a line chart of weekly sales from the Data sheet in this link."
• Check formulas: "Verify the formulas I just added to rows 2 to 200 and tell me if any fail."
• Roll back: "List the edit history of this sheet and revert to yesterday's version."

Examples of what the agent does with a sheet

For data work, the agent reads with +csv-get for plain values or +cells-get when it needs formulas, styles or comments. It writes with +csv-put for plain text, +table-put for typed values such as numbers and real dates, and +cells-set for formulas. The skill's rules say to read the data back after writing and to keep existing content unchanged unless you asked for a change.

For analysis, it can create pivot tables with +pivot-create, filters with +filter-create, conditional formatting with +cond-format-create, and basic charts with +chart-create-basic. For text columns, the skill prefers native AI formulas such as =AI(prompt, range) for translation, classification or summarising, filled row by row.

• Sorting uses +range-sort so whole rows move together.
• Deleting rows uses +dim-delete, not clearing cells.
• Charts must be real chart objects that +chart-list returns; the skill forbids pasting matplotlib images instead.

Safety: scripts, network and high-risk operations

The skill runs scripts and uses the network, because it calls lark-cli against Feishu/Lark on your behalf. Authentication and permissions are handled through the lark-shared skill. Only give the agent access to sheets you are willing to have changed.

Some operations are flagged high-risk in the skill files. These include +batch-update, which bundles multiple writes, +cells-batch-clear for clearing several ranges, and +history-revert, which rolls a sheet back to an earlier version. The revert runs asynchronously, so check its status with +history-revert-status. The batch update requires the --yes flag.

• Keep a backup or rely on history before large batch changes.
• Ask the agent to read back and report results, then review the output.
• For sensitive sheets, review every write before it is final.

Tips for reliable results

Give the agent exact sheet names, the range (for example A1:B2), and the sheet ID or name when the workbook has several sheets. The skill's --range flag takes only the A1 form, and sheet targeting is passed separately with --sheet-id or --sheet-name.

For formulas, ask the agent to verify them. The skill runs +formula-verify with --exit-on-error on the written range and expects every segment to report success. For AI formulas, it uses +formula-verify --ai-only, because those calculate asynchronously.

• Ask the agent to list the workbook first with +workbook-info when the file has many sheets.
• Name the header row and the expected row count so the agent can check the result.
• Ask for changeset review with +changeset-get after a large AI-driven edit; the version gap is capped at 20 revisions per request.