ZCode Work Mode: Coding AI Automates Office Audits, Financial Analysis & Doc Review

Scenario 1: Cross-Checking Project Reports Against Raw Data

The author simulated a messy pre-meeting folder with four files: a draft report that overstated progress, and three raw source tables (requirements, test results, user feedback). The task given to ZCode: verify the draft against the raw tables, flag mismatches, and generate an Excel summary, a 6-slide PPT, and a Word memo.

ZCode produced all three deliverables and uncovered three major issues:

Exposed fake completions : The draft claimed "image validation" and "permission isolation" were fully tested, but the test-results table showed two unresolved bugs for each.

Caught deferred items : The draft listed "attachment preview" as delivered this cycle, but the requirements table confirmed it was postponed to the next version.

Enforced acceptance criteria : Five features passed tests but lacked sign-off records; ZCode marked them "pending confirmation" in the PPT instead of counting them as accepted.

The final PPT showed the true completion rate (9 of 12 items, 75%) and placed the risks on the first slide. Notably, after generating the files, ZCode invoked its built-in viewer to self-check each slide, detected a color-label error on slide 5 and a table-text truncation, and automatically regenerated a corrected version.

Scenario 2: Automated Financial Analysis of Xiaomi 2023–2025 Annual Reports

Using ZCode's plugin marketplace (which includes "Equity Research" and "Macro Strategy" plugins with Wind/10Jing data connectors), the author pointed ZCode at an empty folder and asked it to analyze Xiaomi's operating performance across three full fiscal years, producing an Excel comparison and a 5-slide PPT.

ZCode first tried Xiaomi's official site but hit anti-scraping protection. It then switched to the HKEXnews disclosure platform and automatically downloaded the official PDFs: 2023 (347 pages), 2024 (379 pages), 2025 (415 pages) — 1,141 pages total — into the local directory.

During cross-verification, a financial plugin's AI summary stated "Xiaomi phone revenue +7.5% in 2023." ZCode checked the freshly downloaded original filings, found the figure did not match, and rejected the plugin's number, using only the primary-source data.

Within 9 minutes, the folder contained a 6-sheet Excel workbook, a 5-slide PPT, and a cell-level source log. The Excel summary sheet aligned three years of revenue, multiple profit metrics, and operating cash flow, with both original disclosed numbers and year-over-year calculated results side by side, plus trend charts. The PPT distilled key metrics, revenue and segment trends, profit and cash flow, three core findings, and a methodology/sources page. The author spot-checked against the original PDFs and confirmed 100% data accuracy.

Scenario 3: Cross-Document Audit of 17 Long Articles via /workflow

The author tested ZCode's dynamic workflow ( /workflow) on the JavaGuide open-source knowledge base's concurrency section: 17 long articles, over 10,000 lines, with potential contradictions and citation errors. The workflow writes an execution plan (SOP), dispatches multiple AI sub-teams to review in parallel, and finally assigns a "reviewer" agent to verify findings against the original text.

Key findings from the generated Excel maintenance list and Word report:

Caught a false citation : The author had written "This strategy is commonly used by Meituan" with a link at line 198 of one article. ZCode followed the link, read the Meituan source, and reported: "Verified against Meituan original — that document never mentions this strategy."

Detected copy-paste artifacts : Found two articles contradicting each other on the same technical detail, and identified a stray period ( "两种模式：。") duplicated in article 1 and article 3.

Suppressed false positives on intentional counter-examples : Initial screening flagged 100+ suspect passages, but the reviewer agent re-read context and recognized 20 were deliberate teaching counter-examples, removing them from the final 81 genuine issues. ZCode also left a reusable Python verification script.