Browse task

Skill fnord123/hermes-skills/browse-task

Skills for use with Hermes that work with local models

Install

npx -y skills add fnord123/hermes-skills --skill browse-task

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

Carry out a multi-step task on the web in a real browser — open pages, navigate menus, apply filters, click through results, and report back what it found or did. PREFER THIS SKILL when the user wants something *done on a website* that takes several steps: checking live price or availability by working through a site's own filters, looking something up that needs navigating menus or forms, or operating a web app on their behalf. It is read-only by default; a task that must sign in, submit, buy, book, post, or send requires explicit user approval first. Do NOT use it for a single web search or reading one page — the ordinary web tools handle those better. Activate on any of: "browse to", "go to X and …", "on the website …", "check the site for …", "look up … on <site>", "find the … and tell me", "work through …", "operate …", "fill out …", "navigate …".

The file declares its own license as MIT. That is the author’s claim about this one file, and it is not the same thing as the license GitHub reports for the repository, which is listed with the other numbers below.

SKILL.md

5.3 KB, as published. Nobody here has run it

browse-task — do a multi-step web task in a browser

When to use

The user wants a result that requires operating a website over several steps: apply filters and read the outcome, page through listings, follow a flow across screens, use an interactive web app.
The answer isn't on one page — it takes navigating menus, forms, or results.

When NOT to use

A single web search or reading one page's text — those are faster and cheaper with the ordinary web search / fetch tools.
A task that would change something (sign in, submit a form, buy, book, post, send) unless the user has explicitly approved that exact task — then run it with --confirm.
Anything that isn't on the public web (a desktop app, a local file).

The tool

One script at ${HERMES_SKILL_DIR}/scripts/browse_task.py, invoked as python3 <path> [args]. It prints ONE JSON object on stdout.

Command	Purpose
`browse_task.py --task "<what to find/do>" [--start-url <url>] [--max-steps N]`	Carries out a read-only web task (look up / read / compare) and returns the answer. The agent is told not to change anything.
`browse_task.py --task "<what to do>" --confirm [...]`	Carries out a task that must act on a site (sign in, submit, buy, book, post, send). Use only after the user approved this exact task.

Defaults: opens a search engine first; up to 25 steps. If the task is about a specific site, pass --start-url for that site (e.g. --start-url https://www.rei.com/) so the agent starts on it and handles it best — don't rely on it finding the site from a search. Raise --max-steps for longer flows.

Turning the user's words into calls

User said	Call
"check REI for the price of the X jacket in medium"	`browse_task.py --task "Find the price and availability of the <X> jacket in size medium" --start-url https://www.rei.com/`
"what times is the Ferry Building open on Sunday"	`browse_task.py --task "Find the Ferry Building Marketplace hours for Sunday"`
"find the cheapest nonstop SFO→JFK on Aug 1 and tell me"	`browse_task.py --task "Find the cheapest nonstop flight from SFO to JFK on 2026-08-01 and report the airline, time, and price" --max-steps 40`
"book that flight" (after you showed it and they approved)	`browse_task.py --task "Book the 9am United nonstop SFO→JFK on 2026-08-01" --confirm`

Output shape

Done → {"ok": true, "status": "complete", "answer": "…", "steps": 12, "task": "…", "acted": false}
The agent needs a decision from the user → {"ok": true, "status": "needs_input", "question": "…", "task": "…"} — relay the question to the user and, once answered, call again with the answer folded into --task.
Didn't finish → {"ok": false, "status": "max_rounds" | "timed_out" | "aborted", "error": "… partial finding: …"}.

Relay the answer to the user in plain language. Mention steps only if useful. For needs_input, ask the user the question rather than guessing.

Common flows

Read-only lookup

browse_task.py --task "On <site>, find <thing> and report <fields>"
→ relay answer

A task that acts (needs approval)

# 1. Describe the exact action and get the user's explicit go-ahead.
# 2. Only then:
browse_task.py --task "<the approved action>" --confirm

It asks a question mid-task

{"status": "needs_input", "question": "Which delivery date do you want?"}
→ ask the user, then re-run with the choice included in --task

When the tool reports an error

"not configured" / "not installed" → the skill's one-time setup isn't done; point the user to README.md. Don't try to browse another way.
status: max_rounds / timed_out → the task was too broad or the site too slow; report the partial finding and suggest narrowing it. Don't silently retry.

Always ask the user for guidance when there is an error; do not proactively try to resolve errors yourself.

Empty or inconclusive results

If answer is empty or says the information wasn't found, tell the user plainly that the site didn't yield it — don't invent a value or assume the task succeeded. If the task needed to act but ran read-only (no --confirm), the answer will describe what action would be required; relay that and ask whether to proceed.