Product

research:product — use the product, then show it working

You were asked to research a product. The failure mode this skill exists to kill is the report that read about the product — a page of accurate prose, one screenshot of an idle landing page, and a feature list nobody watched run. That doc proves nothing. A product is verified by driving it: you sign in, you click through the real journeys, you capture each step, you record the headline flow, and you put what the marketing said next to what the product did. Humans read visually — so the output leads with the product working, not with paragraphs.

Product / question: $ARGUMENTS.

The two bars you may never skip

1. Every feature you report is SHOWN being used — a screenshot of it mid-flow, or a clip. A feature you only read in the docs is a claim, labeled as such, never stated as observed. If you couldn't drive it (paywall, version gate), say so — don't launder a doc line into a demonstrated fact.
2. The journey is captured, not summarized. A step you took but didn't screenshot didn't happen, as far as the report is concerned. One idle-landing-page screenshot is the exact defect this skill replaces.

1. Frame — the product, the journeys, the visual identity

• Name the target(s) and the specific question (a product is explored for something: "how does Devin support background agents", not "look at Devin"). --compare a,b,c explores a set on the same journeys for a fair matrix; default is one product, deep.
• List the real user journeys you'll walk — the 3–6 flows a real user actually does (onboard → first result; the core loop; the differentiated feature; the pricing/limits wall). These become the report's spine and each gets its own diagram + strip.
• Grab the visual identity up front. Fetch each product's favicon/logo and embed it as a base64 data: URI so every product card and table row is tagged with its mark (self-contained — no CDN; see the artifacts/design no-external-asset rule):

  curl -sL "https://www.google.com/s2/favicons?domain=<domain>&sz=128" -o /tmp/<slug>.png   # or the site's /favicon.ico, or the logo from its press/brand page
  base64 -w0 /tmp/<slug>.png     # → data:image/png;base64,… inline in the artifact
  

• Pick --depth: quick (one product, top journey + claims), standard (default: all journeys, clip on the headline flow), deep (every journey clipped, pricing/limits probed, edge cases).

2. Get the public picture first — compose research:research

Before you drive anything, know what to verify. Run the multi-engine sweep (research:research) to pull the claimed surface and the sentiment, in parallel:

• Positioning + claimed features + pricing — the landing page, docs, changelog (Codex/web).
• What real users say — X/Twitter, Reddit, HN: complaints, "it doesn't actually do X", launch reactions (Grok — its privileged X data is the point).
• The wedge — what it claims only it does. This is the list you'll try to break in §4.

Stage these as the claim column for §5. A claim with no source is dropped, not carried.

3. Get in — install or sign up (driven, signed in, real)

You are not limited to reading the landing page. Get inside the product:

• Credential exists in agents secrets → sign in and drive the real account.
• No credential → default to a free/trial signup only if the owner's run policy allows it (this creates a real third-party account — an outward-facing act; honor --signup ask|auto|creds-only, default creds-only: sign in if we have creds, else drive the public surface + docs and flag the login wall rather than silently creating accounts). Never fake being inside.
• CLI/installable product → install it for real on a fleet box and run it end-to-end (agents run / a worker), quoting actual output — the eve exploration did this part right.
• Drive with agents browser (web) or agents computer (native desktop, element mode), signed in. If a step needs a secret, inject it from agents secrets — don't stop.

4. Drive every journey — screenshot each step, record the headline flow

This is the step the weak report skipped. For each journey from §1:

• Walk it end-to-end as a real user would, and screenshot every meaningful step — the empty state, the input, the mid-action, the result. Not one capture: a strip that reads as a filmstrip of the flow. Zoom / crop to the decisive interaction so a reader sees what happened, not a full-window thumbnail.
• Record a short screen clip of the 1–2 headline journeys (--record on, default for standard/deep). Capture it with the always-available surface — agents browser's screen capture, or ffmpeg on the recorded frames — and trim to 10–20s; that clip of the core loop is worth more than any paragraph. If the extras create plugin is installed, polish it with create:edit/animator (trim, caption, speed) — optional, never required: with no extras, ship the raw trimmed clip or an animated GIF from the step frames. --record off skips it.
• Feed real, representative input — a real repo, a real document, a real prompt of the shape the product is for. Toy "hello world" input hides exactly the behavior you're evaluating.
• Capture the limits — hit the pricing/quota/error wall on purpose and screenshot it; that's where claims and reality diverge most.

Save every capture into the run's artifact dir; name them by journey + step so §6 can assemble the strips deterministically.

5. Claims vs reality — the differentiated finding

Merge §2's claim column with §4's captures into the table that makes this report worth reading:

• ✓ As claimed — you drove it and it did the thing (cite the capture).
• ⚠ Partial / caveated — it does it, but with a limit the marketing omits.
• ✗ Overstated / couldn't verify — it didn't, or was walled. Say which; never guess a ✓.

The gap between the landing page and the driven product is the research finding.

6. Synthesize — one highly-visual artifact

Author a kind: report artifact and render it with the artifacts skill. It must carry, per product:

• A favicon/logo-tagged product card (identity at a glance — §1).
• A per-journey flow diagram — an inline-SVG of the steps/states you walked (distinct from any architecture diagram; this is the user's path). Use the artifacts diagram conventions.
• The screenshot strip for each journey, and the embedded clip for the headline flow.
• The claims-vs-reality table from §5.
• For --compare: one comparison matrix of the products across the same journeys, each row favicon-tagged, plus a "reach for which when" split.
• Honest gaps — what you couldn't drive and why (version gate, paywall), named, not buried.

A product exploration that ships as mostly prose with one screenshot is incomplete — send it back through §4.

7. Deliver where the owner will see it

• Inspect headlessly first — render, screenshot the rendered page, view it; both themes, no overflow, images resolve. (See it before you call it done.)
• Show it on the owner's interactive box in one reused tab:

  ROOT_DATE=$(date +%F); scp "$ARTIFACT.html" <interactive-box>:/tmp/
  agents browser navigate --device <interactive-box> --url file:///tmp/<file>.html
  

• Promote the durable output to the project's artifacts home (.agents/artifacts/<YYYY-MM-DD>/product-<slug>.{md,html} by default) with the captures beside it; share it if it's meant to leave the machine.
• Close with an honest line: what you drove and verified vs. what you only read.

Anti-patterns (each is the weak report this skill replaces)

• One idle screenshot. A landing page or "Ready" status is not the product working. Capture the flow, not the front door.
• Describing instead of driving. "It supports X" with no capture is a claim, not a finding. Open it, do X, screenshot X.
• No clip, no journey diagram. The headline flow gets a clip; every journey gets a flow diagram. Text-only is the failure mode.
• Features claimed, never shown. If you didn't drive it, it's in the claim column marked unverified — never asserted as observed.
• Toy input. Real, representative data or a same-shape sample; toy input hides the failures.
• Silently skipping the login wall. No creds and policy is creds-only? Flag it and do the public surface — don't fake being inside, don't silently create accounts.
• Leaving it in /tmp or as prose in scrollback. Durable, visual, on the owner's screen.

Compose map

• Public intel (claims + sentiment) → research:research (agents run codex|grok|antigravity).
• Get in + drive the surface → the browser skill (web) / computer skill (native) + secrets.
• Real-surface discipline (signed in, real inputs, capture every step) → the demo skill.
• Clip the headline flow → agents browser screen capture / ffmpeg (always available); polish with create:edit/animator only if the extras create plugin is installed.
• Render the visual artifact + diagrams → the artifacts skill; publish → share.
• A wall only the owner can clear (paid signup, an enterprise gate) → agents feed post "<ask>" --blocked.