Safety

Picaro.AI prompts share a common set of safety rails to prevent operational mistakes and runaway execution. This page documents the underlying specs behind the “no changes until you approve” / “auto-clamped” language that appears on individual prompt pages.

3 design principles

1. Nothing is written to the ad account without approval. The AI returns a proposal and only applies it after the user replies with “execute”. A human review step exists until Phase 4 (full automation).
2. Magnitudes and counts are auto-clamped. Bids, budgets, and exclusion counts all have ceilings. Even if the prompt specifies a value outside the safe range, it’s automatically clamped to prevent runaway changes.
3. Every proposal has an expiration (TTL). Proposals can’t be approved against stale prior conditions.

Approval flow

All write-style prompts (bid optimization, budget adjustment, negative keywords, etc.) work in two stages:

1. Run the optimization prompt
   → The AI generates candidates (the "proposal") and returns them
   → Nothing has been written to Amazon Ads yet

2. The user replies "execute"
   → The proposal ID and approval scope are applied
   → The write reaches Amazon Ads

Partial approval is supported: by specifying proposal-item IDs, you can apply “5 out of 12” rather than all-or-nothing. See Set automation rules Prompt 2 for details.

Proposal expiration (TTL)

Proposals have a 24-hour TTL. extend_proposal can stretch this to a maximum of 48 hours. Past the deadline, the proposal must be regenerated.

Proposal type	TTL	Extendable?
Bid optimization	24 hours	up to 48 hours
Budget adjustment	24 hours	up to 48 hours
Negative keywords	24 hours	up to 48 hours
Placement multiplier	24 hours	up to 48 hours
Impression-share optimization	24 hours	up to 48 hours

Reason: Amazon-side metrics shift over time. Approving a proposal built on stale assumptions produces unintended outcomes.

Auto-clamping (limits on magnitude and count)

Even if a prompt requests an extreme value, the safety rails auto-clamp into a safe range.

Bid change limits

Optimization prompt	Increase ceiling	Decrease ceiling
ACoS-based reduction	—	-30%
All-target unified optimization	+20%	-30%
Amazon recommended-bid following	+20%	-30%
IS% / CVR bid optimization	+30%	-20%

Minimum-bid guard: any calculation result below 1 unit (¥1 / $0.01) is clamped to the floor.

Daily-budget change limits

Prompt	Increase ceiling	Decrease ceiling	Minimum daily budget
Daily-budget optimization	+20%	-20%	¥500 / $5 (default)

The default for max_increase_ratio is 500%, but the daily-budget prompt clamps to +20% as a safer default.

Negative-keyword count

A single execution caps negative-keyword registrations at 50 entries. Prevents accidentally excluding bulk lists.

Placement multipliers

Multipliers are auto-clamped to the 0–900% range.

Rollback

Bid and budget changes can be reversed from history. The pre-change values are logged. To roll back, either tell Picaro “roll back” or open a Support / Feedback ticket.

Manual setting and rollback of placement multipliers is available directly via Adjust bids Prompt 7.

Read-only prompts (no writes)

The following categories are read-only — no writes to the ad account occur at all.

• View KPI dashboards
• N-gram sales decomposition
• SQP market search funnel
• Create reports (HTML output only; no ad-setting changes)
• Support / Feedback Prompt 2 (review past feedback)

These are safe to run repeatedly.

Auto-masking of sensitive info

When generating reports or sending feedback, the following are automatically masked:

• Access tokens or API-key-like strings
• Trading-partner names / account information (per report rule)
• Personal names (configurable via custom mask settings)

For details, see Create reports Prompt 4 “Save and reuse report rules”.

Audit log

Changes to KPI targets, budget caps, automation rules, and event logs all land in the audit log. The log captures who changed what, when, and on which account.

Related

• Automation phases (Phase 1–4) — the roadmap for deepening automation while keeping the safety rails
• Troubleshooting — known issues and workarounds
• How Picaro differs from other tools — the design philosophy behind the “propose → approve → execute” three-stage structure