12 · Web research and competitor intel

At a glance

Solves	Finding current information, competitor research, industry round-ups
Needs	The web-search Skill; agent-browser for complex pages
Takes	5–15 minutes per round
Risk	Low for files — but high for information reliability

The situation

Your boss asks: what have our competitors shipped in the last six months, and which of it should we follow?

You open five sites, spend forty minutes, and produce a list you are not sure is complete.

Preparation

Install two Skills:

• web-search — when a question involves "recent", "latest" or "now", model memory is not enough. This has WorkBuddy search first, organise, then answer
• agent-browser — lets it enter a page and perform browser-level operations: open, scroll, click, expand, screenshot, read. For pages whose content only appears after interaction, this changes what is reachable at all

They do different jobs

web-search solves "can it be found". agent-browser solves "can it be opened and fully read".

Use the first where search results give you what you need; the second where content requires scrolling, expanding or a session.

The prompt

Basic research

Search for significant developments in AI office tooling over the past month
and summarise the key points.

Requirements:
- Source link and publication date on every item
- Only items with concrete information: product releases, feature updates,
  funding, disclosed data, policy changes
- No opinion pieces, commentary or advertorial
- At most 10, ordered by importance
- One line on "who did what", then one on "why it matters"

Delete items where you cannot find a reliable source — do not keep them.

Structured competitor research

Research competitors A, B and C (site links below).

For each, establish:
1. Main features shipped in the last 6 months, in date order
2. Pricing (tiers, prices, the core differences between them)
3. Publicly disclosed user scale or usage data
4. The differentiators they emphasise publicly

Data requirements:
- Source link and retrieval date on every item
- Distinguish "officially disclosed" from "third-party estimate / press report"
- Leave fields you cannot find blank and mark "no public information found" —
  do not fill with industry rules of thumb

Output as CSV first: vendor / category / content / source type / source link / date.
I will decide about analysis after reviewing the data.

Pages that need interaction

Open this page, read all the content that only appears after expanding and
scrolling, produce a structured summary, and screenshot the key pages for the record.

Specifically:
- On the pricing page, expand every "see details" and "compare all features"
- Expand every FAQ accordion
- Save screenshots to a screenshots directory in the workspace

Round two: analysis

Only after you have the data and have checked it:

The data is right. Write an analysis from that CSV:

1. What direction do all three share, and what is each betting on uniquely
2. What do the pricing differences tell us
3. Name three things they do that we do not, that are worth following
4. Name anything they do that appears to be working badly, that we should skip

Cite the specific data behind each judgement.
Mark anything purely speculative as "speculation, not supported by data".

Why it is written that way

Absolute time ranges. "The past month", "the last 6 months", "so far this year". "Recent" to a model can mean any length.

Demanding sources. Necessary for verification and citation, and the only way to judge whether something is real. A conclusion with no link is one you cannot check, and cannot defend when challenged in a presentation.

Distinguishing official disclosure from third-party estimate. The layer most often flattened in research. "5 million users" from a financial report and from someone's guess in an article differ by an order of magnitude in reliability.

"Delete items with no reliable source" plus "leave fields blank". Together these lift the pressure to fill everything in. Without that permission, blanks get filled with plausible fiction.

"No opinion pieces, commentary or advertorial." The largest noise source in research work — plentiful, loudly headlined, low in information density.

CSV before analysis. Same logic as the spreadsheet playbook: when the data is unreliable, analysis just packages the unreliability more attractively.

Acceptance checklist

• [ ] Open five source links at random and confirm they work and match the summary
• [ ] Publication dates fall inside the range you asked for
• [ ] Official disclosure and third-party estimates really are distinguished
• [ ] "No public information found" fields are genuinely absent, not laziness
• [ ] Every judgement in the analysis maps to specific data
• [ ] Speculation is labelled

Checking links is mandatory

There is no substitute for this step.

In AI-generated research, fabricated source links look exactly like real ones — plausible domain, plausible path structure, plausible title. Opening them is the only way to tell.

One dead link in five means the whole report needs re-verifying.

Variations

Ongoing tracking

Turn a one-off into a habit:

Create an automation for Mondays at 09:00:

Search for last week's public developments from A, B and C, categorise them as
product / commercial / people / other, attach source links, and append to
D:/competitors/updates.md, one section per week, newest at the top.

If a week has nothing worth recording, write "nothing this week" rather than
padding it out.

See Automations on a schedule.

Deep-reading one page

Read this page in full (including anything requiring scrolling or expanding), then:

1. Summarise the core argument in three sentences
2. List every verifiable factual claim (numbers, dates, quotations)
3. Identify which points are the author's opinion and which are reported from others
4. Note any obvious information gaps or logical jumps

Do not assess the quality — structural breakdown only.

Cross-checking one claim

For the claim "company X had revenue of 1.2bn in 2026", find every source you
can and compare what they say.

Where sources disagree on the number, list them all with their sources.
Do not decide which is correct on my behalf.

That last line matters — have it surface the disagreement rather than digesting it away.

Store it in the knowledge base

Save this research report to the Library under "Team space / Competitors",
named competitor-research-2026Q1.md.

Store the CSV data there too, so it can be referenced directly later.

See The Library.

Fact verification: one hard rule

A research report cannot go out unchecked

AI-generated research has three typical distortions:

1. Fabricated sources — links that do not open, or open onto something unrelated
2. Misattribution — a real number attached to the wrong company or period
3. Stale information — old data from model memory presented as current

All three share a feature: they read flawlessly.

Before it goes anywhere, verify every number, every date, every direct quotation and every source link.

An easy way to make that cheaper is to have it flag its own risk:

Add a "needs human verification" section at the end listing every number, date
and attribution you are not fully confident about, and where you got each.

FAQ

Everything it found is old.

Pin the time range in the prompt and require publication dates. When something is outside the range, say so:

Items 3, 5 and 7 fall outside the range — delete them and search again.

Pages that need a login.

agent-browser can perform browser-level operations, but be careful about credentials — never put account details in a prompt.

Safer: sign in yourself and copy the content out, or screenshot and paste it.

Can it crawl a whole site?

Technically it can do crawling work, but consider three things: the target's robots.txt and terms of service, the load your crawl puts on their servers, and whether your intended use of the data is lawful.

Reading publicly accessible pages for competitor research is fine; bulk crawling needs a compliance answer first.

How do I stop it treating search results as instructions?

A real risk — page content can contain text designed to influence model behaviour. Two defences:

1. Do research in Ask mode, where it cannot perform file operations
2. Stay on default permissions so dangerous actions stop for confirmation

If it suddenly wants to do something you did not ask for mid-research — write a file, visit an address — cancel, then ask it why.

Are malicious sites blocked?

WorkBuddy carries a built-in malicious-domain list and checks target domains before network access or page fetching inside the sandbox, refusing on a match. The check runs locally and the list updates daily.

Note the boundary: it does not cover network access inside plugins, or search services.

End of the playbooks

Twelve done, and you should now have several processes you can reuse. Next:

• Freeze them → Turn a routine into a Skill
• Hand over the recurring parts → Automations on a schedule
• Look up a prompt → Prompt library