Skywatch Labeling Standards

Labelling Standards

Every mcp__skywatch-mcp__ozone_label action MUST include a comment that meets the standards below. These comments are the permanent moderation record — a future reviewer seeing this label must understand exactly why it was applied without re-investigating.

Comment Format

Every label comment follows this structure:

[Label] applied — [one-line policy basis]

Evidence:
- at://did:plc:.../app.bsky.feed.post/[rkey] — "[verbatim text]" — [editorial note]

Summary Line (Required)

The first line is always:

[Label] applied — [one-line policy basis]

• [Label] — the exact label value being applied (e.g., troll, spam, maga-trump)
• [one-line policy basis] — why this label applies, referencing the policy or behavioural pattern (e.g., "coordinated inauthentic behaviour per IO Policy", "dehumanising language toward immigrants — one-strike")

Evidence Section (Required)

After the summary line, an Evidence: block with cited posts.

Each citation line:

- at://did:plc:.../app.bsky.feed.post/[rkey] — "[verbatim text or excerpt]" — [editorial note]

Components:

| Component | Required | Notes | |-----------|----------|-------| | AT-URI | Always | Full at:// URI to the specific post | | Verbatim text | When quoting text | Exact text from the post in quotes. Required for text posts. Not applicable for image-only posts — describe the content instead | | Editorial note | Always | Brief explanation of why this citation is relevant to the label (e.g., "TDS framing per maga-trump precedent", "denies trans identity", "bot-like posting cadence") |

Citation Requirements by Context

Post-Level Labels (Single Post)

Minimum 1 citation — the specific post being labelled.

troll applied — quote-post trolling targeting original poster

Evidence:
- at://did:plc:abc123/app.bsky.feed.post/xyz789 — "imagine thinking this is a good take lmao ratio" — mocking tone directed at OP, no substantive engagement

Include thread context (parent post AT-URI) when the label depends on the reply relationship.

Account-Level Labels (Pattern-Based)

Minimum 2 citations — enough to establish a behavioural pattern, not a one-off.

spam applied — repetitive promotional posting across multiple threads

Evidence:
- at://did:plc:abc123/app.bsky.feed.post/aaa111 — "Check out [link] for amazing deals!" — promotional post in unrelated thread
- at://did:plc:abc123/app.bsky.feed.post/bbb222 — "You won't believe these prices [link]" — same promotional pattern, different thread
- at://did:plc:abc123/app.bsky.feed.post/ccc333 — "Best deals online [link]" — third instance confirming systematic behaviour

Batch Operations

Batch labelling (multiple subjects, same label) still requires per-subject evidence comments. The batch context (e.g., "sentiment analysis batch", "coordinated network takedown") goes in the summary line's policy basis, NOT as the entire comment.

Wrong:

Quote post trolling - sentiment analysis batch

Right:

troll applied — quote-post trolling identified in sentiment analysis batch

Evidence:
- at://did:plc:abc123/app.bsky.feed.post/xyz789 — "lol imagine being this wrong" — dismissive mockery with no substantive engagement

Data Sourcing for Evidence

Building an evidence comment requires the actual text of the posts you're citing. Follow this priority order — each step is cheaper and faster than the next:

1. ClickHouse (always try first) — query osprey_execution_results for the DID's posts. Cheapest source, includes rule match context, covers ~2 months of data. If you already have ClickHouse results from a prior step (triage, investigation, batch analysis), use them — don't re-query.

1. Slingshot (for specific AT-URIs) — if you have the AT-URI of the post you're labelling but don't have its text (e.g., it fell outside the ClickHouse window, or you're in a context without ClickHouse access), hydrate it via Slingshot's getRecordByUri endpoint. Fast, cached, no auth needed.

1. PDS direct (fallback) — use PDSX mcp__pdsx__get_record or mcp__pdsx__list_records when neither ClickHouse nor Slingshot can provide what you need. This includes: accounts that pre-date the ClickHouse indexing window, taken-down accounts where Slingshot may not have the record cached, or when you need to browse a collection (listRecords is not available via Slingshot).

Never label without reading the content first. If you can't retrieve the post text through any of these sources, you can't build a valid evidence comment, and the label should not be applied.

What Fails Validation

The mcp__skywatch-mcp__ozone_label PreToolUse hook will reject label actions when the comment:

• Is missing entirely
• Has no summary line matching [label] applied —
• Has no Evidence: section
• Has zero AT-URI citations (at:// links)
• For account-level labels (subject is a DID, not an AT-URI): has fewer than 2 AT-URI citations

When rejected, the agent receives an error message directing it to load this skill and rebuild the comment.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: skywatch-bsky
• Source: skywatch-bsky/skywatch-agent-skills
• License: MIT
• Homepage: https://bsky.app/profile/skywatch.blue

Install and usage instructions live in the source repository linked above.