← All posts

How to build a competitor activity workflow in Claude Cowork

2026-09-10

Most organisations monitor their competitors in bursts. A tender is lost, a rival announces a partnership, or a board member asks a question nobody can answer, and someone spends a fortnight assembling a picture of the market. The deck gets circulated. Then it goes quiet until the next surprise.

The alternative is a workflow that does a little of the work on a schedule and hands you a short brief you can read in two minutes. Claude Cowork can run this kind of recurring work, so building it is a matter of describing the job once rather than writing and maintaining code.

This guide walks through it step by step. Allow an afternoon for the first version.

Before you start

You will need Cowork, which is available on paid plans (Pro, Max, Team and Enterprise) on desktop, and in beta on web and mobile. Scheduled tasks are available on all of those paid plans.

Two things worth knowing up front. Cowork uses more of your usage allocation than ordinary chatting, so a daily task has a real cost attached. And if you are on a Team or Enterprise plan, an administrator controls whether Cowork is available to you at all, so check before you invest time in a build.

Step 1: Define what counts as competitor activity

Do this before you open Cowork. In this workflow the prompt is the entire specification, so vague inputs produce a vague brief every single morning.

Write down three things.

The organisations you are tracking. Name them. For most teams this is five to fifteen, not fifty. Include direct competitors, adjacent players who could move into your space, and any organisation that regularly appears alongside you on shortlists.

The signals that matter. A competitor updating their website footer is not a signal. A contract award in your core market, a senior appointment, a new service line, or hiring that points to a new capability all are. Decide which categories of change would genuinely alter a decision you might make.

Who reads the brief, and what they do with it. If nobody can name a decision it informs, you are building a reading habit rather than a working one.

Keep this list short. Every category you add dilutes the brief.

Step 2: Create a project to hold the workflow

Cowork projects are persistent workspaces with their own files, instructions and memory, and memory is scoped to the project so what Claude learns in one does not leak into another. That makes a project the right home for recurring work.

Find Projects in the left navigation panel, click the "+" button, and choose to start from scratch. Name it something obvious, like "Competitor monitoring".

Put your Step 1 decisions into the project instructions: the organisations, the signal categories, and any house rules about tone or format. Anything you would otherwise repeat in every prompt belongs here.

On Team and Enterprise plans you can share a project with colleagues, giving them view or edit access, which is worth doing if more than one person will maintain the workflow.

Step 3: List your sources explicitly

Left to its own devices, Claude will search broadly and return a slightly different mix each day. That makes briefs hard to compare week to week. Naming your sources fixes this.

For organisations working with or alongside Australian government, the most useful sources tend to be:

For listed companies, add their ASX announcements.

Use public sources only, and do not point the workflow at anything behind a login or a paywall. There is more on why in Step 8.

Step 4: Run the check manually, once

This is the step people skip, and it decides whether the workflow ends up being useful.

Open a task in your new project and run the job by hand. Ask for exactly the brief you want, covering the last week rather than the last day so you have enough material to judge.

Read what comes back properly. Your first attempt will almost certainly return too much, miss the source you cared about most, or summarise at the wrong altitude. That is normal and it is much cheaper to discover now than after a fortnight of scheduled runs.

If the task needs to visit websites, Claude can work through a browser built into the desktop app, so there is nothing to install.

Step 5: Tighten the prompt

Rerun and adjust until a single run produces something you would genuinely read. Six instructions do most of the work.

Name the sources. Specific organisations, specific portals.

Give an exact date range. "The past 24 hours" beats "recent". Ambiguity here is why briefs start repeating themselves.

Require a source link on every item. This is the most effective control against fabricated detail, because it makes verification a single click.

Tell Claude to flag uncertainty. Instruct it to write "unclear from source" rather than inferring. Without this, an ambiguous tender notice becomes a confident summary.

Cap the length. Ten items maximum. A brief that is always readable in two minutes gets read; one that is occasionally thirty items long gets ignored permanently, including on the days it matters.

Say what a quiet day looks like. Instruct it to write "nothing significant today" rather than padding. A quiet brief is information, and it confirms the workflow is still running.

A worked example:

Prepare a competitor activity brief covering the past 24 hours.

Organisations to track:
[your 5-15 organisations]

Sources to check:
- Each organisation's news/insights page and careers page
- AusTender contract notices naming any tracked organisation
- NSW eTendering, Buying for Victoria and QTenders award notices
- Public announcements and media coverage naming any tracked organisation

Report only these categories:
- Contract or tender awards
- New service lines, partnerships or acquisitions
- Senior appointments or departures
- Published research or submissions in our practice areas
- Hiring that signals a new capability

For each item, give:
- A headline
- A two-sentence summary
- The organisation, the date, and a direct source link

Rules:
- Maximum 10 items, ranked by significance
- If a detail is ambiguous in the source, write "unclear from source"
  rather than inferring it
- Do not repeat items covered in the previous three briefs
- If nothing significant occurred, write "Nothing significant today"
  and stop
- Save the brief as a document titled "Competitor brief [date]"

Step 6: Turn it into a scheduled task

Once a manual run produces good output, schedule that version. You can type /schedule inside the task you have just been working in, which converts the process you have already proven, or click "Scheduled" in the left sidebar to create one from there.

Setting a task up manually asks for a name, the prompt, an approval mode, the frequency, and optionally a model and a working folder. Three choices deserve thought.

Cadence. The options are hourly, daily, weekly, weekdays or manual. Weekdays suits most competitor briefs, since weekend runs mostly repeat Friday.

Approval mode. Manual approval pauses for you before actions. Automatic approval lets Claude work without interruption but reviews each action for safety first and blocks anything unsafe, at the cost of more usage. Skipping approvals means nothing checks its actions. For a task reading external websites unattended, keep meaningful oversight.

Where it runs. Scheduled tasks run in the cloud, so they fire on their cadence even when your computer is asleep and the desktop app is closed. They work with your connectors and the files saved to your Claude account, and cannot be tied to a folder on your computer. A task that genuinely needs local files or applications will only run locally, which means your machine has to be available.

Run it once on demand before relying on the schedule.

Step 7: Decide how the brief reaches people

Claude can email the brief to you directly. Scheduled tasks have access to the same capabilities as any other Cowork task, including your connectors, so connecting Gmail or Outlook lets a run finish by sending the brief rather than just saving it. You manage connectors from the "+" menu in the message box or under Customize > Connectors. Enable only what this workflow actually needs.

Add the delivery instruction to the end of the prompt:

- Email the finished brief to [your address] with the subject
  "Competitor brief [date]"

Two things to know before you rely on it.

On Team and Enterprise plans, there is an admin setting in the way. The organisation setting "Allow 'Always allow' for connector tools" controls whether members can skip per-task approval for write-capable connector tools, and it is off by default. While it is off, "Allow for all tasks" is greyed out in the approval dialog and previously saved always-allow preferences are not honoured, so members approve those tools task by task. Sending email is a write action, so an unattended scheduled run will sit waiting for an approval that nobody is there to give. If you want the brief to send itself, that setting needs to be turned on by an owner, which is a conversation to have deliberately rather than a box to tick.

Anthropic's own guidance counsels caution here. The advice on scheduled tasks is to avoid automating consequential actions, and specifically not to schedule tasks that send messages on your behalf. Step 8 explains why that is well founded for this workflow in particular.

So treat delivery as a ladder and climb it as you build confidence:

  1. Leave it in Cowork. The task saves a document each run and you read it from the Scheduled page. No outbound action at all. This is the right default while you are still tuning the prompt.
  2. Write to a shared location. A document in a folder or drive the team already checks. More convenient, still no messaging.
  3. Email it to yourself. Once the output has been stable for a fortnight, add the connector and send to your own address.
  4. Email a wider list. Only after the brief has proven itself, and never to a client-facing list from an unattended task.

The brief being somewhere predictable matters more than it landing in an inbox, so there is no urgency about reaching step four.

Step 8: Set your guardrails

Prompt injection deserves attention here. Anthropic identifies web content as a primary vector for prompt injection, where instructions hidden in a page Claude reads hijack the task. A competitor workflow reads external websites by definition. The risk becomes material when Claude can both read untrusted content and take consequential actions, so keeping the brief read-only, or keeping a person between the brief and anything that leaves your organisation, breaks that combination. Watch the outputs for scope creep too: if the brief starts covering sites you never named, stop and investigate.

Respect terms of service. You remain responsible for actions Claude takes on your behalf, including respecting third-party website terms and any restrictions on automated access. For a firm bidding on government work, this is a probity question as much as a legal one.

Keep confidential information out. Anything arriving through a client relationship, a former employee or a tender process does not belong in this workflow. It watches what is published. Make that boundary explicit to everyone with access.

A person still reads it. AI summarisation is a drafting aid, not an authority. If an item will inform a bid decision or a board conversation, someone opens the source first.

Check your organisation's AI use policy before this becomes business as usual.

Step 9: Review and maintain

Agent-based workflows decay quietly. This one will not throw an error. It will just start returning slightly less useful output, and you may not notice for a month.

Read past runs fortnightly. The Scheduled page shows upcoming and past runs. Open a few end to end and look for drift: repetition, vagueness, sources that have stopped appearing.

Review the source list quarterly. Add the organisation that has started appearing on your shortlists. Remove the one that has exited.

Pause what you are not using. You can pause, resume, edit or delete any scheduled task. If nobody has acted on anything for three months, either the signals were wrong or the brief is reaching the wrong person. Both are fixable, but not while the task quietly keeps running.

Common mistakes

Tracking too many organisations. Fifteen is a working list. Fifty is a filter problem you will never fix.

Scheduling before testing. A prompt that has never produced a good manual run will not produce a good scheduled one.

Automating the send too early. Convenience is not worth an unattended task emailing clients.

Never reading the runs. An unmonitored workflow is not a monitoring workflow.

Where to start

Build the smallest useful version: three competitors, two sources each, run by hand for a week. Then schedule it, keep it read-only, and add only the sophistication the first fortnight tells you is missing.

A simple brief that runs every weekday for a year will teach you more about your market than an elaborate one that quietly stopped being read in March.

Sources

← Back to all posts