huashu-excel - Huashu's open-source AI data analysis skill - AiBoss

What is huashu-excel?

Huashu-excel is an open-source AI data analysis skill developed by Huashu. Through an eight-step process—including health check, cleaning, alignment, analysis, reconciliation, delivery, image verification, and quality control—it addresses the core issue of preventing errors in AI data analysis. Huashu-excel uses openpyxl to read raw cells to identify structural pitfalls, then uses in-table totals for free verification and reconciliation, and a separate agent recalculates and corrects errors. The tool supports HTML/Excel/Word report output, depends solely on openpyxl, and is open-source under the MIT license.

The main functions of huashu-excel

Data health checkUse openpyxl to read raw cells and identify structural issues such as misaligned headers, merged cells, and mixed subtotal rows before creating a dynamic pandas table.
Data cleaningThe dirty tables are transformed into normalized analysis tables, and an auditable and replayable cleaning script is generated, recording how many rows are affected at each step.
Demand AlignmentFirst, understand the characteristics of the data and external benchmarks, then confirm the actual needs and analytical criteria with the users to avoid conducting a complete analysis with incorrect questions.
Data AnalysisScan for common data traps such as Simpson's paradox, ghost grouping, small cardinality, and time breakpoints, and push the findings to "so what?"
Digital reconciliationUse the cleaned details to check the "Total" row in the table, and use free validation to identify manual entry errors and discrepancies in the data.
Report deliveryOutput can be in HTML (inline SVG can be opened offline), native Excel (charts are linked with data), or a six-page Word narrative report.
Chart verificationCheck the SVG rendering for out-of-bounds, overlapping, occlusion, and dual-axis misleading to ensure the chart is honest and colorblind-friendly.
Independent quality controlWe assigned another agent who wasn't involved in the creation process to recalculate the original data, specifically to dismantle and verify the reliability of the conclusions.

How to use huashu-excel

Install Skill:run npx skills add alchaincyf/huashu-excel Alternatively, you can manually clone it to the agent's skills directory.
Select pathClaude Code ~/.claude/skills/Codex / Kimi Code ~/.agents/skills/At the project level, it can be moved with the warehouse.
Natural Language CallSpeak plainly, such as "Help me analyze this sales report," "These two numbers don't match," or "Help me clean the report."
On-demand trigger processAsk for a number and look it up directly; get a report and go through all eight steps; form cleaning only requires physical examination and cleaning, and comes with a replayable script.
Scripts can be used independentlyRun directly without an agent profile_table.py(Physical examination)verify_numbers.py(Account reconciliation)verify_visual.py(Verification of images), etc.

The core advantages of huashu-excel

Medical examination first, then calculationBy using openpyxl to read the original cell structure, potential problems such as merged cells and mixed subtotal rows can be detected before the pandas table is generated, thus avoiding misjudgments due to information corruption.
Total line checksumUse the cleaned details to check the "Total" row in the table, turn the true value calculated by the original table author into a free checksum, and identify manual input errors.
Independent quality control dismantlingWe assigned another agent who wasn't involved in the creation process to recalculate the original data, specifically to verify the reliability of the conclusions and solve the problem of "internal data being completely correct but the overall data being wrong for a whole fiscal year".
An error will occur if the calculation is wrong.By employing multiple gates such as row number conservation, total sum conservation, and cross-reconciliation, AI data analysis is transformed from "silent error" to "no delivery if it doesn't match".
Chart honestyBased on Cleveland & McGill perceptual precision step selection visual coding, pie charts are rejected if there are more than 3 categories, bar axes are forced to start from 0, and dashed lines are drawn at the end of incomplete categories.
Cross-Agent GeneralIt is not bound to a specific platform; it can be installed with Claude Code, Cursor, Codex, OpenClaw, and Hermes. Pure standard library scripts can also run independently without an agent.
Minimalist offline availabilityThe core dependency is only openpyxl, requiring no pandas, plotting libraries, LibreOffice, or internet connection. HTML reports can be drawn inline with SVG and opened directly offline.

The project address for huashu-excel

GitHub repository：https://github.com/alchaincyf/huashu-excel

Comparison of huashu-excel with similar competing products

Comparison Dimensions	huashu-excel	Julius AI
Product Form	The open-source Agent Skill (MIT) can be installed on any platform such as Claude Code, Cursor, and Codex.	Closed-source SaaS product, web-based chat interface, requires account registration.
Data reading	First, use openpyxl to read the original cell data for analysis, retaining information about merged cells, formatting, and original data types.	The uploaded file is parsed directly, resulting in merged cells being expanded and formatting information being lost.
Error detection	The eight-step process incorporates multiple gates; if it doesn't match, delivery will not proceed, and errors will be proactively reported.	Based on the output of the dialog stream, calculation errors are usually silently resolved, requiring the user to make their own judgment.
Reconciliation mechanism	Use the cleaning details to check the "Total" row in the table, and use the true value calculated by the original table author as the free checksum.	Without built-in cross-validation, the total rows are processed as ordinary data rows.
Quality control mechanism	A separate agent was assigned to recalculate the data from the original data, specifically to dismantle and verify the conclusions.	Single dialogue path, no independent review role
Chart Guidelines	Encoding is selected according to Cleveland & McGill perceptual precision steps, with the bar axis forced from 0, and pie charts rejected for classes exceeding 3.	Charts are generated based on instructions, but there are no built-in anti-misleading rules, which can easily lead to truncated axes and dual axes.
Report Output	Self-contained HTML (inline SVG, can be opened offline), native Excel (charts move with data), and a six-page Word document format.	Embedded charts in web pages + export static images/PDFs, offline forwarding may result in format loss.
Operating environment	It only depends on openpyxl, can run completely offline or on an intranet, and the script can be executed independently without the agent.	Internet access is required; data must be uploaded to a third-party server.
Customizability	Open source and modifiable, SKILL.md itself is a methodology document, and the process can be adjusted according to business needs.	It operates as a black box, meaning users cannot modify the underlying analysis process.
Cost Model	Completely free, no subscription fee, no usage limits	The free version has limited availability; the Pro version starts at $20/month.

Application scenarios of huashu-excel

Daily operations morning meeting debriefingEvery day, the system retrieves the previous day's operational data from the backend, automatically completes the process of data check, cleaning, analysis, and reconciliation, and submits a data fluctuation attribution report with a statement of definition before the morning meeting.
In-depth review of sales performanceIt can automatically identify structural problems such as filling gaps in new products and losing money on old products when a million-line transaction detail is lost, and output an HTML report that can be directly submitted to the sales VP.
Financial budget execution reviewAnalyze municipal or corporate budget statements and identify the pitfall of "internal accuracy but overall error for a fiscal year" by reconciling with external benchmarks, and provide a committee-level explanation.
HR Pay Fairness DiagnosisThe study processed tens of thousands of salary questionnaires, controlled for industry, education level, length of service, and city, and calculated the gender gap, clearly marking the analytical boundaries of "what it can prove and what it cannot prove".
Data cleaning and audit handoverClean the dirty tables and come with a replayable and auditable pandas cleaning script. It records how many rows are affected by each step, making it easy to trace responsibility or reuse later.