| name | observe-page |
| description | Observe one page or a list of pages in the browser and write an observation note — what data is there and which durable handles reach it. |
Observe a page
Turns a page into a note another agent can build a skill from without reopening the
browser. Read references/durability-doctrine.md first; it defines the handle ladder
this skill applies.
Inputs
One or more URLs, and what the user wants from each page. With no URL list, take the
page list from notes/brief.md.
Steps
For each page, in sequence — never fan out parallel loads against a live site:
- Load it on the chosen route and confirm identity: final URL after redirects,
rel=canonical, title, and any identifier echoed back. Note any mismatch — a
numeric id whose slug is decorative will serve a different item at 200.
- Read the page as a person. What is this page for, what does it answer, what is
the unit of the thing on it (a product, a variant, a family, an order line)?
- Find the data at the highest rung available. Check, in order: JSON-LD and
microdata,
og:/meta tags, embedded state (__NEXT_DATA__, dataLayer,
hydration payloads), then accessible names, labels and heading-anchored structure.
Record the rung, not just the result.
- Write the durable description of each data point — how a person would tell a
colleague to find it. Record any CSS selector as a dated hint beneath it, never as
the definition.
- Record the assertion that proves the right element was found (the model cell
matches the requested model; the heading names the requested order).
- Record what is not on the page. Stock, freight, delivery date, tax treatment —
whatever a caller will assume is there and isn't.
- Note the traps: repeated blocks,
rowspan/colspan, symbols printed on the
first row only, values that differ when signed in, anything locale-dependent.
- Record interactions as element descriptions, never as positions. If a skill
will have to click something, note what the element is called and how a person
would identify it — never a coordinate, an offset, or "the third button".
- Classify every line against
references/publication-policy.md as you write it.
Account-bound values and defence observations go to the private workspace.
Output
notes/<surface>.md in the target repo, from templates/observation-note.md, with
today's date. Batch write-ups can be delegated to the observation-recorder agent —
hand it the captured text; it does not drive the browser.
