How to Set Up an Agentic Workspace for Ecommerce

An ecommerce reporting task rarely starts with one clean prompt. The numbers may sit in three exports, the definitions in a planning document, the acceptable format in last week's deck, and the important exceptions in someone's head. Uploading those files to a chat can help once, but it does not leave behind a dependable process for the next trading meeting, campaign review, or catalogue audit.

An agentic workspace solves that continuity problem. It is a bounded environment that gives an agent the inputs, instructions, references, checks, and output locations needed to complete several connected steps, while making it clear where a person still needs to review or decide. The exact buttons differ between tools, but the operating structure does not.

1. Define the workspace around one deliverable

Start with a clear finish line, such as a Klaviyo performance review or a weekly trading report. Write down the audience, cadence, output format, and the decision the output is meant to support. If the deliverable is still vaguely described as an analysis or a set of insights, the agent has no stable definition of done.

I draw the boundary just before money, customers, or brand judgment are affected. An agent can prepare a performance diagnosis, propose catalogue corrections, or assemble a trading narrative. A person should still approve a budget change, edit live product data, or distribute the report.

2. Assemble the evidence before writing instructions

Create working copies of everything a capable colleague would need to complete the task. Include current inputs, written definitions, and one prior output that has already been accepted as good. Remove credentials, unnecessary customer data, and live production access. If the workflow can be tested with redacted or sample files first, use those.

The quality reference matters as much as the raw data because it shows the expected section order, level of explanation, and visual standard more precisely than a request to make the result polished. Treat it as a target to inspect, not proof that the agent will reproduce every detail perfectly.

3. Build a platform-neutral workspace

Use a small structure that makes the work legible to both the person and the agent:

ecommerce-workspace/
  INSTRUCTIONS.md
  README.md
  INDEX.md
  inputs/
    current/
    archive/
  references/
    accepted-output-example/
    metric-definitions.md
  plans/
  workflows/
  outputs/
  checks/

INSTRUCTIONS.md is the portable source of truth for the agent. It defines the purpose, file locations, workflow rules, permission limits, required checks, and human approval points. README.md explains the process to a person, while INDEX.md maps the available files without copying sensitive data into the index. The remaining folders separate current source data, quality references, saved plans, narrow workflow definitions, generated work, and verification results.

This structure is intentionally independent of a specific AI product. If a tool supports project-level instructions, copy the short standing rules there and keep the detailed version in INSTRUCTIONS.md. That gives the agent immediate context without making the process difficult to move later.

4. Connect the workspace in Claude Cowork or ChatGPT Work

The workspace design stays the same, but each product handles context differently.

Claude Cowork

Create a Cowork project for the recurring process, attach the local workspace folder, and add the standing instructions that should apply across the project. Anthropic's current guide says a Cowork project can bundle local folders, standing instructions, links, and project memory. Cowork can read and write local files after you grant access in the desktop app (Anthropic project guide, Anthropic Cowork guide).

Confirm the attached folder before each run, then ask Claude to read INSTRUCTIONS.md, INDEX.md, and the relevant workflow file before it proposes a plan. The instruction file cannot compensate for attaching the wrong directory or omitting a required source.

ChatGPT Work

If you already use Codex for work that is not code, you can stay there. OpenAI positions ChatGPT Work as the same core capability with an interface built for everyday tasks rather than engineering ones, which means the workspace structure applies either way.

Create a ChatGPT project when the work will continue over time, produce several outputs, or depend on the same files and sources. Add the standing rules as project instructions, then upload or connect the workspace sources. OpenAI says project instructions apply across project chats, while the Sources section holds uploaded files and connected context (OpenAI projects guide).

Switch the composer to Work for the actual task. OpenAI recommends giving ChatGPT Work a clear outcome, a small set of source materials, and an output you can review. In the desktop app, choose Work locally when the task needs files or apps on your computer. If Cloud is available, use it for work that should continue without the computer being awake and does not depend on local resources (OpenAI Work guide).

Whichever platform you use, start a separate chat for each distinct outcome. Keep the shared rules and sources in the project, but avoid letting several deliverables blur together in one long conversation.

5. Copy this workspace setup prompt

Give the agent a plain-language briefing as if you were onboarding a colleague. Include what inputs are required, what each file means, which steps are mechanical, where judgment begins, what must be checked, and who approves the result. Then ask it to turn that explanation into workspace files and a saved plan that you can inspect.

Copy and adapt this prompt in either Claude Cowork or ChatGPT Work:

Set up this folder or project as a workspace for [recurring ecommerce process].

The finished output is [deliverable] for [audience], produced [cadence].
The current source files are in inputs/current. Explain what each file contains
in INDEX.md without copying sensitive data into the index.

Use references/accepted-output-example as the quality and format reference.
Record metric definitions and non-negotiable rules in references.

Create:
1. INSTRUCTIONS.md with the workspace purpose, file locations, workflow rules,
   permission limits, required checks, and human approval points.
2. README.md with a short human-readable operating guide.
3. INDEX.md with a file map.
4. plans/first-run.md with the steps for one bounded test run.
5. outputs/ and checks/ folders, keeping generated work separate from inputs.

Before moving or changing files, show me the proposed structure. Do not connect
to live systems, send anything, publish anything, or modify production data.
Flag missing inputs and conflicting definitions instead of guessing.

Review the generated files before the first run. Correct any misunderstood definitions or guidance, and make every approval point explicit. An instruction file can describe a job, but it does not create access to a service or authority to act in it.

6. Adapt the structure to one ecommerce workflow

The workspace becomes useful when it can carry a process across several steps while leaving evidence of what happened. Split the work into stable jobs, such as preparing data, calculating metrics, diagnosing anomalies, drafting commentary, and checking the final output. Each job should name its allowed inputs, expected output, required checks, and stopping condition.

Klaviyo performance analysis

Inputs: Campaign and flow exports for the same reporting period, attributed revenue settings, audience or segment definitions, the promotional calendar, and an accepted prior analysis.

Agent job: Normalize the exports, compare performance by campaign and flow, identify material changes, and draft explanations that distinguish observed data from possible causes.

Checks: Confirm date ranges, currencies, attribution settings, metric definitions, duplicate sends, and missing rows. Reconcile any headline total back to the source export.

Output: A review pack with a metrics table, anomaly list, evidence-linked commentary, and questions that require commercial context.

Human approval: A lifecycle marketer validates the interpretation and approves any change to targeting, creative, cadence, or promotional strategy.

Shopify catalogue and merchandising QA

Inputs: Product and variant exports, collection rules, inventory status, image and URL fields, merchandising standards, and an approved exception list.

Agent job: Flag incomplete attributes, inconsistent option names, duplicate or conflicting handles, unexpected collection membership, missing images, and products whose availability conflicts with merchandising rules.

Checks: Validate row counts, unique identifiers, required fields, exception handling, and a sample of flagged products against Shopify. Keep proposed fixes separate from the source export.

Output: A prioritized QA report and a correction file that records the current value, proposed value, rule triggered, and confidence.

Human approval: An ecommerce owner approves every production change, especially updates affecting price, availability, taxonomy, URLs, or customer-facing copy.

Weekly ecommerce trading report

Inputs: Shopify sales and order exports, web analytics, paid media summaries, lifecycle performance, forecast or target files, promotion notes, stock constraints, and the previous accepted report.

Agent job: Align reporting periods, calculate agreed metrics, compare actuals with targets and prior periods, surface anomalies, and draft a concise narrative with source references.

Checks: Reconcile revenue and order totals, confirm timezone and currency, identify incomplete source feeds, apply the written metric definitions, and separate measured changes from explanatory hypotheses.

Output: A dated trading pack with the scorecard, material movements, open questions, and a decision log for the meeting.

Human approval: The trading owner confirms the narrative, decides which actions follow, and controls distribution to stakeholders.

The repeated shape is what matters: bounded inputs, a specific agent job, fixed checks where possible, a durable output, and a named person who owns the consequential decision. That operating design turns a collection of files into a reviewable process.

7. Test one bounded run before adding a schedule

Run the smallest useful part of the workflow with current or sample inputs, then compare the result with the accepted reference and trace every headline claim back to its source. Correct the saved instructions, definitions, or checks instead of quietly repairing only the output. On the next cycle, replace the inputs and run the same plan in a new chat. A process is becoming reusable when it exposes missing context consistently and produces work that can be reviewed without reconstructing how it was made.

Scheduling comes last. First decide whether the workflow depends on local files or can run with cloud-accessible sources, then confirm that the chosen tool can reach every dependency in that environment. Test the scheduled version and retain the same checks and human approval points. A recurring task should repeat a proven process, not hide an unstable one.

Key takeaway

A useful agentic workspace is a bounded operating environment rather than a longer prompt tied to one tool. Its inputs, quality standard, intermediate jobs, checks, outputs, and approval rights remain visible whether it runs in Claude Cowork, ChatGPT Work, or another capable agent. I would start with one recurring ecommerce deliverable, let the agent prepare the reviewable work, and keep consequential judgment with a named person. The setup is successful when the next run is easier to verify, not merely faster to generate.