Chat for one-off work, a Project when it repeats, Research for depth, Cowork or Claude Code for automation

Name the job in one sentence

Write the job as one sentence with a deliverable and a reader: “a two-page comparison of a competitor’s new pricing page against ours, for the sales team, by Friday.” If you cannot name the deliverable yet, you are still exploring, and exploring is a chat.

With the sentence in hand, compare the surfaces by the job they must do. The access requirements below come from Anthropic’s own pages, each opened on September 16, 2026.

Two published guides sort the same choice by one variable. xponent21’s “Onboarding Test” asks how much documentation a job needs before it can be handed over; the Product Marketing Alliance’s decision framework sorts by task type. Neither asks what changes when the input is a deal note or a lead list, so that question comes next here.

Sort the input before you choose

Use three labels as a routing aid. Public or aggregate material, such as a competitor’s pricing page, published reviews or your approved messaging, can go in as it is. A call note or deal summary may be redactable, but removing names, company identifiers and deal figures is only the first step. Before uploading it, check for indirect or contextual identifiers and confirm that your organization’s policy, the applicable customer or recording terms, and the vendor terms permit that use. If any check fails, exclude it. Material that identifies a lead or person, or comes from an unpublished contract, stays out of every surface. A Japanese practitioner guide from uravation.com applies a three-tier classification before sending data over MCP; the labels here are our working recommendation, not proof that material is safe or authorized. Safe GTM data covers the full review.

The input label does not choose the surface. A recurring job with properly redacted and authorized inputs can still use a Project; an automatic job on public pages may fit Cowork or Claude Code. The label changes what goes in, who may access it, and which account a connection uses.

A one-off question: chat

Use plain chat when you will not need the same setup again: a first read of a competitor’s launch post, a rewrite of one paragraph, a check on whether a claim matches the wording legal approved. A chat starts from what you paste into it, so attach the source each time and ask for the answer together with the passage it rests on.

Check the answer against the source you attached, not against memory. When another chat starts by re-pasting the same context, the job has become recurring, and the next section applies.

The same context every week: a Project

A Project holds instructions and knowledge files that every chat inside it uses. Use one when the job repeats for the same product and market: weekly competitor notes, positioning reviews, answers to seller questions.

The Project setup lesson covers instructions, the reference files, connections and the first check. The check that matters for the surface choice is short: open a new chat inside the Project, ask for one fact you can verify, and confirm Claude names the file it came from. If it answers from general knowledge, the Project is not yet carrying the context and you are still in chat with extra steps.

In standard Claude chat, a Project supplies context; it does not by itself start a chat on a schedule. Cowork projects can contain scheduled tasks on paid plans. For “every Monday, read these pages and tell me what changed,” compare Cowork with Claude Code scheduling, then choose the environment whose file, tool and permission model fits the job.

Use Research for multi-step source gathering and an in-depth report

Research is a multi-step search Claude runs on its own, across the web and connected Gmail, Google Calendar and Google Docs, ending in a report with citations. Anthropic’s Research article, checked September 16, 2026, limits it to paid plans, requires web search to be on, and says a research session can use up your limits faster than an ordinary conversation.

Use it when the job is “find out what is being said about this” and you do not already hold the sources: analyst coverage of a competitor, the review sites in a market, the press pickup of a launch. The competitive research plan shows how to scope that job first.

Open every cited source you rely on and confirm that each material claim has matching support. The article describes the citations as easy to check, which is different from checked; a finding whose source you cannot open stays a lead. Verifying AI research is the fuller procedure.

A Research report is a snapshot. Run it twice on the same question a month apart and you have two reports to reconcile by hand. That is the moment to decide between a manual cadence you can keep and the next surface.

It acts on files or starts automatically: Cowork or Claude Code

Cowork scheduled tasks can run recurring work in a Cowork project on paid plans, using the files and connected tools available there. Cloud tasks can continue with the app closed. Choose Cowork when that environment and its approval settings fit the job; choose Claude Code when the work belongs in a Code host with its files, commands and MCP configuration.

Claude Code reads a folder of files, edits them, runs commands, connects to tools over MCP and can run on a schedule. The overview, checked September 16, 2026, describes it as a coding tool, lists a terminal, IDE extensions, a desktop app and the web as places it runs, and says most of them need a Claude subscription or an Anthropic Console account. A PMM reaches for it when the job has become a system: keep a folder of competitor notes current, regenerate a digest from them, deliver the result to a channel.

Setup is real work. You install something, give it a working folder, and decide which tools it may run. This page does not walk through that. MCP for marketers covers the connections, and solo competitive intelligence covers when a scheduled Claude Code job is worth building instead of doing the check by hand.

Run it once on inputs whose answer you already know, read every file it changed, and only then let it run unattended. Text it retrieves from a page or a connector is evidence to weigh, even when it reads like an instruction. If nobody on the team can read what a run changed, stop the job until someone can.

When the choice repeats, write it down

The first time you choose a surface for a job, it is a judgment. When the same choice starts to recur, record it so someone else can run the job during an absence. Record five things:

• The job in one sentence, with the deliverable and who reads it.
• Each input and its route: public, properly redacted and authorized, or off-limits.
• The surface, and the symptom that would move the job to another one.
• The check you run on the output before anyone else sees it.
• How often the job runs, and who runs it when you cannot.

That record is the job brief the next lesson builds, and the first page of the workflow handoff a teammate needs. Keep it beside the Project or the Claude Code folder it describes, and change it the day the job changes.