SKILL.md

Drop this into your agent's skill/prompt directory so it knows how to use TuringTap effectively.

Download SKILL.md

Worked patterns

Case 1 — mid-flow handoff

1. goto(url) / act(...) until you hit a wall (2FA, captcha, consent).
2. ask_human(prompt) — non-blocking; keep polling get_traffic().
3. When human_active flips false (or you see the traffic you needed), dismiss_human(handoff_id, "Thanks — got it.").
4. Resume act(...).

Case 2 — freeform demo

1. ask_human("Show me how you book a class", open_url).
2. User demonstrates on their phone or desktop; hits Done.
3. analyze(host) to summarize the endpoints.

Credential hygiene (for the AI agent)

• get_traffic() returns raw, unfiltered traffic — any credentials it contains enter your context and your model provider's logs. There is no redaction: partial redaction invites false confidence, so TuringTap does not promise it.
• Never echo credential values into the chat. Refer to a credential by host and kind ("the session cookie for gym.com"), never by value.
• Prefer analyze() (endpoint shapes only) and narrow host filters when you don't need raw exchanges.
• End the interaction by reminding the user to rotate any credentials they consider sensitive that transited the session: their AI provider's logs may retain the conversation.