name	hn-interaction-safety
description	Apply safety and confirmation rules to HN voting, comments, and reply drafts whenever an interaction could affect the user's public account.

HN Interaction Safety

Use When

The user asks to vote, submit a comment, reply, edit, delete, flag, or otherwise interact with Hacker News.
The agent prepares or polishes a reply draft that could later be posted.
A request is ambiguous about whether the user wants a draft or a public action.

Do Not Use When

The task only reads, summarizes, searches, clusters, or analyzes public HN content.
The task reviews recommendation quality without changing HN or product state.
The user asks for general writing advice unrelated to an HN interaction.

Safety Rules

Treat every vote, comment, reply, edit, delete, flag, and submission as an external side effect.
Never perform an HN side effect from inferred intent, prior approval, or a request to analyze content.
A reply draft remains a draft. Clearly label it and do not imply it was posted.
Before any side effect, show the exact target and final action or text, then obtain explicit confirmation for that specific action.
Confirmation must be current and unambiguous. Material edits after confirmation require confirmation again.
Do not request, expose, or repeat HN credentials, cookies, session tokens, or another user's context.
Respect rate limits and application-owned idempotency. Report uncertain outcomes; never retry a possibly completed action blindly.
Refuse impersonation, coordinated manipulation, deceptive voting, harassment, spam, or evasion of HN controls.

Draft Rules

Preserve the user's position and disclose substantive changes.
Remove unsupported factual claims or mark them as uncertainty.
Keep disagreement directed at ideas and evidence, not people.
Do not fabricate personal experience, affiliation, or consensus.

Output Contract

For read-only or draft work, return mode: draft and the proposed text or analysis.
For a pending side effect, return mode: confirmation_required, the exact target, and the exact proposed action.
Only after explicit confirmation may an application-owned typed mutation tool run.
If confirmation or authorization is missing, return mode: blocked with the missing requirement.

HN Interaction Safety

Use When

The user asks to vote, submit a comment, reply, edit, delete, flag, or otherwise interact with Hacker News.

The agent prepares or polishes a reply draft that could later be posted.

A request is ambiguous about whether the user wants a draft or a public action.

Do Not Use When

The task only reads, summarizes, searches, clusters, or analyzes public HN content.

The task reviews recommendation quality without changing HN or product state.

The user asks for general writing advice unrelated to an HN interaction.

Safety Rules

Treat every vote, comment, reply, edit, delete, flag, and submission as an external side effect.

Never perform an HN side effect from inferred intent, prior approval, or a request to analyze content.

A reply draft remains a draft. Clearly label it and do not imply it was posted.

Before any side effect, show the exact target and final action or text, then obtain explicit confirmation for that specific action.

Confirmation must be current and unambiguous. Material edits after confirmation require confirmation again.

Do not request, expose, or repeat HN credentials, cookies, session tokens, or another user's context.

Respect rate limits and application-owned idempotency. Report uncertain outcomes; never retry a possibly completed action blindly.

Refuse impersonation, coordinated manipulation, deceptive voting, harassment, spam, or evasion of HN controls.

Draft Rules

Preserve the user's position and disclose substantive changes.

Remove unsupported factual claims or mark them as uncertainty.

Keep disagreement directed at ideas and evidence, not people.

Do not fabricate personal experience, affiliation, or consensus.

Output Contract

For read-only or draft work, return mode: draft and the proposed text or analysis.

For a pending side effect, return mode: confirmation_required, the exact target, and the exact proposed action.

Only after explicit confirmation may an application-owned typed mutation tool run.

If confirmation or authorization is missing, return mode: blocked with the missing requirement.

hn-interaction-safety