Phantom

• 

  Phantom Preview

• 

  Action Guide

• 

  Evidence Page

• 

  Settings Page

Inspiration

Moderation decisions often happen under pressure. A post or comment may clearly need attention, but the downstream effects are not always obvious: removing one item can break an active discussion, orphan reply chains, confuse users, or escalate a situation that could have been handled with a softer intervention.

Phantom was inspired by that gap between "can I remove this?" and "what will this removal do to the community conversation?" The goal was to give moderators a quick, practical preview before they act, without replacing moderator judgment or automating sensitive decisions.

What it does

Phantom adds moderator menu actions for posts and comments. When a moderator selects Phantom Preview, the app creates a private workflow-style Devvit custom post that analyzes the selected target and shows:

• an influence / impact score
• direct reply count
• downstream reply-chain impact
• target activity in the thread
• crosspost exposure
• account-age context
• score drivers explaining why the risk is high or low
• evidence snippets from the surrounding discussion
• recommended safer actions such as monitoring, adding a mod comment, locking a branch, or escalating to manual review

Phantom also includes subreddit-level settings so each community can tune how strict the scoring is. Moderators can choose a scoring preset, hide or show the numeric score, adjust individual score weights, and enable automatic pinning of newly created Phantom preview posts.

The app records lightweight aggregate decision stats in Devvit KV Store so moderators can see how often previews were opened, how often actions were cancelled, and how often high-risk actions were reconsidered.

How we built it

Phantom was built as a Devvit app using TypeScript and Reddit's Devvit public API.

The app uses:

• Devvit menu actions for post, comment, and subreddit moderator workflows
• Devvit custom posts for the interactive Phantom preview UI
• Reddit API access to inspect posts, comments, authors, crosspost counts, and subreddit context
• Devvit settings for per-subreddit configuration
• Devvit KV Store for aggregate event logging and usage statistics
• TypeScript scoring logic to calculate the blast radius of a moderation action

The core scoring engine looks at multiple signals rather than one simple metric. It considers direct replies, reply chains, target participation, crosspost exposure, and account-age context, then turns those signals into a readable impact summary and recommendation set.

Challenges we ran into

The biggest challenge was designing something useful without pretending to be a fully automated moderation judge. Moderation is contextual, so Phantom needed to explain risk and alternatives instead of simply saying "remove" or "do not remove."

Another challenge was making the preview fast and readable inside a Reddit moderator workflow. The app has to gather enough context to be useful, but still present it in a compact interface that moderators can scan quickly.

We also had to make the scoring configurable. Different communities have different norms: a high-volume discussion subreddit may want to preserve reply chains more aggressively, while a stricter support or safety-focused community may care more about repeated target behavior.

Finally, we added automatic pinning for Phantom preview posts so active moderation reviews can stay visible to the mod team when the setting is enabled.

Accomplishments that we're proud of

We are proud that Phantom focuses on decision support rather than automation for its own sake. It helps moderators see the likely community impact before acting, while leaving the final call with the human moderator.

We are also proud of the explainability. The app does not just show a score; it shows the drivers behind the score, evidence from the thread, and practical next steps.

The subreddit settings are another important accomplishment because they make Phantom adaptable. Communities can tune the scoring model and presentation to match their moderation style.

What we learned

We learned that moderation tooling is most useful when it fits naturally into existing workflows. A separate dashboard would be easy to ignore, but a post/comment menu action appears exactly when a moderator is already making a decision.

We also learned that "impact" is not a single number. Reply structure, user participation, crossposts, and account context all matter in different ways. Building Phantom forced us to think carefully about how to turn those signals into something moderators can actually use.

What's next for Phantom

Next, we would like to add richer team workflows around previews, such as shared moderator notes, review status, and follow-up reminders.

We would also like to improve the scoring model with more community-specific calibration, better historical trend analysis, and clearer explanations for edge cases.

Longer term, Phantom could become a broader moderation risk-preview layer for Reddit communities: not a replacement for moderators, but a tool that helps them act with more context and fewer unintended consequences.

Built With

• devvit
• devvit-custom-posts
• devvit-kv-store
• devvit-menu-actions
• node.js
• reddit
• reddit-developer-platform
• settings
• tsx
• typescript