ORBiS TREE · ENGLISH DOCUMENT
Products →ChatGPT Work
A task environment for creating deliverables with available files and tools, explained through clear goals, access boundaries and reviewable outputs.
한국어
English
This English page is a curated translation layer linked to the Korean source article. Community changes are currently made on the Korean source, where the full revision history and anonymous edit trail are preserved.
Overview
ChatGPT Work carries out multi-step tasks using available files and tools. OpenAI describes it as an environment for delegating work with a clear outcome, including creating deliverables and recurring tasks.[1] Work is a product environment; GPT-6 Astra is a model. Model choice and feature availability can vary by account, client and workspace.[2]
Specify the outcome
A useful request identifies the deliverable, source material, scope and acceptance criteria. For a sales report, specify the period, units and the columns that should be compared. Decide what evidence must accompany an explanation and which actions require review. These are ORBiS Tree's practical questions, not a measured performance guarantee.
Tools and access
A model's ability to describe a task does not give it access to a file or server. Work depends on enabled tools, connected services and permitted resources.[1] Check which source was actually read and whether the requested output was actually saved. Distinguish a proposed action from a completed action with observable evidence.
Worked example: reviewing a quarterly report
This example is hypothetical. A polished report compares a quantity column as though it were revenue. Review the original column labels, units and aggregation period before accepting the narrative. If the source does not establish why sales changed, describe the explanation as a hypothesis. Recheck totals against the original rows and verify that the deliverable opens correctly.
Continuing a task
Keep important decisions and source versions in the work product or task record. Do not assume every detail of an earlier conversation is always available. AI Memory explains stored state and retrieval; AI Agent Oversight explains how to place meaningful review points around consequential actions.
Write a reviewable task contract
ChatGPT Work's official guide ↗ describes multi-step work with a clear deliverable.[1] Specify inputs, scope and acceptance checks alongside the requested output. The following is an ORBiS Tree prompt format, not a separate built-in contract feature.
- Inputs: sales-q1.csv and sales-q2.csv, including date, product, quantity and amount.
- Outputs: a product comparison, calculation evidence, unresolved items and a report file.
- Scope: compare the two quarters without mixing quantity and amount.
- Completion: reconcile totals, explain missing or duplicate rows, and open the file and links.
- Execution: review the draft; external sharing uses confirmed recipients and content.
Model choice concerns reasoning capabilities; available tools determine executable access. The model guide ↗ should not be read as granting file or account permissions.[2]
Worked example: checking quarterly totals
These figures are hypothetical. In Q1, product A has quantity ten and amount 100; B has quantity five and amount 200. Totals are 15 and 300. In Q2, A has quantity 12 and amount 120; B has quantity five and amount 250. Totals are 17 and 370. Amount increases by 70, or approximately 23.3%; quantity increases by two, or approximately 13.3%. Swapping these rates makes the report incorrect.
B's quantity is unchanged while its amount increases. The data alone does not establish whether pricing, discounts or exchange rates caused the difference. Separate observed changes from causal hypotheses. This is a review example, not actual sales data.
Connect intermediate outputs and completion
Keep input names, versions and column meanings with calculations and result files. A later correction can then identify the exact sources and outputs to revise. Do not accept completion solely from a message: verify that files exist, open successfully and contain the expected calculated values.
Task state preserves decisions and source versions; provenance identifies where numbers came from. When further evidence is required, consider retrieval and source organization before producing a stronger claim.
Place review points around consequences
Drafting a report and publishing or sending it have different effects. Preserve recoverable states for file changes and verify authorized scope and confirmed content for external actions. Repeated approval prompts do not establish accurate calculations. Execution oversight and unsupported explanations help separate review of content from review of actions.
SOURCES
Sources
KNOWLEDGE LINKS
Continue from here
BACKLINKS