What’s inside

One SKILL.md (the playbook) + one starter template. A preview:

Browser Task Agent

When to use

• Filling out long web forms: applications, registrations, government portals, supplier onboarding
• Multi-step web jobs: checking a price across five sites, booking something with no API, updating listings
• Reading a page and acting on it: "find the cheapest flight for these dates and hold it," "check if this item is back in stock"
• Repetitive browser chores you do weekly: pulling a report from a dashboard, copying data between two sites
• Anything where you want proof it was done right — every run ends with screenshots

Workflow

1. Task brief (input: your words → output: task spec)
   State the task plainly: what (fill the supplier form), where (URL), and the done-criteria (submitted + confirmation number, or data copied into the sheet). Include the inputs the task needs (your details for the form, the dates for the search) and the hard stops (never submit a payment, never agree to terms, stop and ask if the site asks for something unexpected). A task without done-criteria can't be verified — "check the prices" needs "report the three cheapest with links."
2. Step plan (input: task spec → output: plan with checkpoints)
   Break the task into steps, each with a checkpoint: after step 2, verify the search results loaded; after step 4, verify the form shows your name correctly. Checkpoints are where the agent stops, looks, and confirms before continuing — they're what keep a 15-step task from going off the rails at step 3. Mark which steps are irreversible (submit, purchase, delete) — those always stop for your approval.
3. Playbook + evidence (input: step plan → output: executed playbook with screenshots)
   The agent runs the plan: navigate, read, click, fill — following the checkpoints. Every checkpoint gets a screenshot. At the end, you get the playbook back filled in: each step marked done / skipped / failed, the screenshots as evidence, and the result (confirmation number, the data, the report). If a step failed, the playbook says exactly which one and what the page showed — enough for you or the agent to retry just that step.

Output spec

• Task spec: what / where (URL) / done-criteria / inputs / hard stops — one page
• Step plan: numbered steps, each with its checkpoint (what to verify) and reversibility marked
• Executed playbook: step / status (done/skipped/failed) / screenshot / notes
• Final result: the confirmation number, the collected data, or the report — whatever the done-criteria demanded
• Any failed step must record what the page actually showed, not just "it didn't work"

Good vs. bad examples

Template preview

Browser Task Pack

Task spec

• What: […]
• Where (URL): […]
• Done-criteria: […]
• Inputs: […]
• Hard stops: [never submit payment / never agree to terms / stop and ask if …]

Step plan

Executed playbook

Result

• […]
• Confirmation / reference: […]

Buy the full skill pack — $4.99 Pro — all 471 skills