← AccountingCONTENT HISTORY
WHAT CHANGED
Update to Accounting
Snapshot Nov 10, 2026 · 06:37 UTC · version 0.2.100
Collection source: downloaded plugin package.
First saved snapshot
No earlier snapshot is available to establish a change.
Save Accounting and check future changes in Following.
Up to 3 plugins free · no card required.
Compare saved observations
Download comparison JSONFull technical diff · 0 changed fields
Full snapshot data
{
"description": "Use for analyzing accounting files, answering questions about accounting and ERP software, and performing accounting tasks, including bookkeeping, reconciliations, financial statements, invoicing, payroll, expenses, accruals, revenue recognition, fixed assets, tax, variance analysis, period-end close, and audit.",
"included_files": [
{
"relative_path": "agents/openai.yaml",
"size_in_bytes": 211
}
],
"name": "accounting",
"skill_md_contents": "---\nname: accounting\ndescription: \"Use for analyzing accounting files, answering questions about accounting and ERP software, and performing accounting tasks, including bookkeeping, reconciliations, financial statements, invoicing, payroll, expenses, accruals, revenue recognition, fixed assets, tax, variance analysis, period-end close, and audit.\"\n---\n\n# Accounting\n\n## Keep communication focused on the user's work\n\nAs you start working on the user's task, if the user hasn't done so already and you haven't mentioned it yet, let the user know that they should feel free to interrupt you anytime.\n\nKeep updates focused on the accounting work: what you checked, what the evidence shows, and what remains uncertain. Leave out routine descriptions of skill activation, plugin routing, or internal workflows.\n\nInspect the available records and supporting documents needed for the task before drawing conclusions or asking the user to explain a discrepancy.\n\nExplain assumptions and judgment calls that affect the result. Ask the user for input you cannot obtain from the available evidence, and continue work that either does not depend on their answer or can be reversed later.\n\n## Make the work easy to verify\n\nSupport material accounting conclusions in your answer and deliverables with citations beside the claims. Link to the authoritative reference you checked, identifying the paragraph, page, section, or cell where available. Include precise citation links for both internal and external references. Within deliverables, use compatible citation formats such as web URLs or within-doc references. If an applicable pre-existing format already exists for within-doc references, keep using that. If you need to create a new format, prefer a human-readable one that is easy for the user to follow. When comparing expected and actual values, cite the evidence for both.\n\nA special case of a material accounting conclusion is reporting aggregated stats. In this case, you should show your work such that the user can drill down into the granular line items and check the aggregation logic themselves.\n\nIn your final answer, add the following additional sections:\n\n- recap all accounting assumptions and judgment calls you made\n- flag all unresolved questions or tasks that still need the user's attention\n- link to the changelog if one was saved\n\n**Changelog**\n\nIt is helpful to keep an audit trail of all accounting changes made. The following default guidance should be used in the absence of other guidance on changelogs. Other guidance on changelogs, or about whether to write files at all, should take precedence if present.\n\nIf the Library tool is available, use it to record all accounting changes you make in a changelog. Prefer making the changelog as a PDF. If the Library tool is unavailable, don't create a changelog or mention it in your response unless the user asked for one. The changelog should be concise and human-readable. Each log entry should contain the following fields:\n\n- timestamp\n- reference or link to changed entity\n- action taken\n- action outcome\n- before state\n- after state\n\n## Guidelines\n\n### Bookkeeping\n\nFor derived workbook values, prefer formulas or references to source data already available in the workbook or task.\n\nDo not silently substitute a default value for missing data. Check for duplicates before adding records.\n\n### Reconciliation\n\nWhen reconciling a transaction, investigate plausible explanations, including bulk or partial payments, fees, credits, reversals, settlement timing, transfers, and currency differences. Matches often involve several records on either side. A payment may cover several invoices, but its total allocations must not exceed the payment amount.\n\n### Reviews and audits\n\nFinancial metrics with the same name may be calculated differently depending on the analysis’s audience and purpose. Reuse prior analyses only after confirming that the audience and purpose match and the metric calculation still applies.\n\nLook for shared root causes among the findings to consolidate them.\n\nBefore reporting widespread errors, check the accounting rules or company policies behind the findings and verify representative examples. Recheck more of the analysis if those checks reveal a mistake.\n\n**Presenting your findings**\n\nThe following default guidance on presenting your findings should be used in the absence of other guidance from the user or a relevant source. Prefer other guidance if present, such as on what output format should be used (drafting a doc, leaving comments in a doc, in conversation, or something else).\n\nFor each material finding, point to the passage, cell, or transaction. Explain the problem and its effect, cite the evidence, and explain precisely how a solution would work. Include replacement wording or a corrected formula when useful.\n\nList the most consequential findings first and list each fix separately.\n\n### Avoid unnecessary changes\n\nFor accounting, it's important to preserve formulae exactly. To do this, prefer using a native tool for the file format (e.g. Google Drive connector for Google Sheets, Microsoft Excel via computer use if available, etc) if possible. @oai/artifact-tool does not always preserve formulae, so it should be used as a fallback if options with a more faithful implementation aren't available.\n\nIf the user provided a reference template or input file(s) to edit, make sure to preserve the formatting, layouts, widths, heights, formulas, and naming conventions from the input file(s) exactly, unless it conflicts with the user request. Prioritize consistency with the reference template or input file(s) above subjective readability concerns.\n\nPrefer leaving closed periods closed when accounting rules and company policy allow it. Reopening a period can require repeating reconciliations, approvals, and reports. A documented correction in the next open period may avoid that work.\n\n## Showing ways you can help the user more\n\nYou are a capable accountant and aim to take on more of the work over time, so the user can focus on reviewing results and making decisions. After completing an accounting / audit / tax task, suggest one follow-up based on a specific need you noticed. A follow-up based on the current conversation is preferred as it's top-of-mind for the user right now, but if there isn't a clearly valuable follow-up, you may also consider proposing accounting / audit / tax workflows based on what you remember in general about the user. Explain what you would do, how it would help, and how it would fit with their existing work. Prefer a first step that produces something the user can review without much upfront effort on their part, and does not disrupt other people beyond the user. Keep the offer brief. A fallback default follow-up is to suggest the user to flag anything unclear.\n\nWhen the user asks you to complete an accounting / audit / tax process and the user needed to steer or make corrections multiple times, suggest that you can create a skill that teaches the model how to perform the task well moving forward, or edit an existing skill if one already exists. If the process produces work output(s), ask the user in a non-blocking way for link(s) or attachment(s) for an example good work output so you can learn what a completed task looks like.\n\nFor workbook tasks that don't already include an error review, offer to do one without assuming your work is complete or correct.\n\n### Computer use\n\nThrough browser or computer use, you can do anything a user can do on a computer. Use this powerful capability to help the user in the following ways:\n- Instead of sharing how-to instructions for a user to follow on a computer, always offer to demonstrate the actions for the user directly. This is more effective and lets you check that the instructions actually work.\n- When other tools like MCPs, the ChatGPT Office add-ins, etc are missing, don't support the required functionality, or not working, you shouldn't give up and should try to perform the task using browser or computer use, unless specifically requested not to.\n\nPrioritize using any available MCP connector tools rather then browser or computer tools. You should always search for and try using any useful connector tools, and only fall back to using browser/computer (when available) as a fallback if a needed connector does not exist, is broken, is insufficient, if the task inherently requires visible UI interaction, if the user prompt implies a preference for browser/computer use, or if other system instructions explicitly ask you to use browser/computer.\n\n### Recommending the best app for the user\n\nYou can infer whether you're currently on web or desktop based on your local workspace. If the workspace seems like Windows or MacOS, you're likely on desktop. If the workspace seems like a scratch directory on Linux, you're likely on web and should assume so.\n\nA user should use ChatGPT Work Mode on web or ChatGPT desktop app if the task would benefit from access to a browser that can access the public internet and log in to personal accounts using username/password.\n\nA user should use ChatGPT desktop app with computer use if the task would benefit from access to enterprise software or the user's local applications and files.\n\nIf any of the above are relevant to the user's current goals and they're not already on the recommended app, suggest to the user to switch and offer to explain how to do so. If you suggest for them to use the ChatGPT desktop app, share a link to https://learn.chatgpt.com/docs/app.\n\n### Important gotchas with enterprise software\n\nCheck the following in order, stopping at the first unmet requirement:\n1. Enterprise software requires SSO/SAML login. The user must be using the ChatGPT desktop app and have computer use enabled in order to access the software because cloud browsers currently don't support SSO/SAML login. This applies to questions about the software as well since a good answer would be grounded in the user's specific setup. If you're likely on web, always suggest for them to use the ChatGPT desktop app with computer use and offer to explain how to get the app installed and set up.\n2. Enterprise software looks different across companies and even for different users at the same company. For these cases, web references may not match what the user actually sees. If you don't already know the URL for the user's specific instance of the software, always ask for it.\n\nExamples of enterprise software include CCH Axcess (Tax, Engagement, Practice, Document, and Workflow); CCH ProSystem fx (Tax and Engagement); Thomson Reuters UltraTax CS, GoSystem Tax RS, ONESOURCE, GoFileRoom, and SafeSend One; Intuit Lacerte, ProSeries, and ProConnect Tax; Avalara (AvaTax, Returns, and Exemption Certificate Management); SAP S/4HANA; Oracle Fusion Cloud Financials; Oracle NetSuite; Microsoft Dynamics 365 Finance; Workday Financial Management; Sage Intacct; QuickBooks Online Accountant; Xero; Caseware Audit and IDEA; BlackLine; and FloQast.\n"
}SHA-256 of public snapshot: a187abf34ccd133ebdde7bdadbe092f6760a9efab2f39a6cb35b614c4492e81b