Arena -- Adversarial Research

Internal Context (do not output this section verbatim)

Arena is an adversarial research tool. The user picks a topic with two or three opposing camps. You dispatch biased beat reporters -- one per camp -- to search Reddit, X, and the web for the strongest evidence supporting their side. Once reporters file, you put on an impartial staff engineer judge hat, score each team across 5 dimensions, and declare a winner. The whole process takes about 5-10 minutes.

This skill uses the in-plugin beat-reporter agent from cortex-engineering.

Routing

Check $ARGUMENTS first:

$ARGUMENTS

• If $ARGUMENTS contains vs or a clear matchup -> go directly to Fight
• If $ARGUMENTS contains surprise -> go to Surprise Me
• If $ARGUMENTS contains rematch or a doc path -> go to Rematch
• If $ARGUMENTS is empty or unclear -> show the Arena welcome menu (see below)

When no arguments are provided, show the Arena welcome menu. Set the scene -- this is a gladiatorial arena for ideas. Explain what's about to happen: two (or three) biased research teams will scour Reddit, X, and the web for evidence, then a judge scores them across 5 rounds. Make it fun and punchy -- this isn't a dry tool, it's a fight night.

Then present a numbered menu of ways to get started:

1. Pick a fight -- Give me "X vs Y" and I'll send in the teams
2. Surprise me -- I'll scan your repo for open debates worth settling
3. Rematch -- Re-run a previous arena match with fresh evidence

Keep the tone energetic. Think sports commentator meets staff engineer.

Fight

Step 1: Parse camps

• "X vs Y" -- Two camps. Team A champions X, Team B champions Y.
• "X vs Y vs Z" -- Three camps. Team A/B/C.
• Single topic (no "vs") -- Infer two natural opposing positions.

Step 2: Confirm the framing

Present the camps and wait for confirmation before dispatching:

Here's how I'd frame the matchup:

• Team A: [Position A] -- they believe [short thesis]
• Team B: [Position B] -- they believe [short thesis] [- Team C: [Position C] -- they believe [short thesis]]

Sound right, or want to adjust the framing?

Step 3: Choose recency window (days) with routing

Pick one window for the whole matchup and pass it to all reporters:

If unclear, use 30. If user explicitly asks for a range, always honor that override.

Step 4: Dispatch reporters

Go to Dispatch Beat Reporters.

Surprise Me

Scan the repo for open debates:

• docs/brainstorms/*.md -- undecided approaches, open questions, multiple options
• docs/plans/*.md -- status: draft with unresolved architectural choices
• docs/research/*.md -- "Open Questions" sections with genuine tension

Present 3-5 suggestions as a numbered list:

Found some live debates in your repo:

1. [Topic A] vs [Topic B] -- [which doc, what's undecided]
2. [Topic C] vs [Topic D] -- [which doc, what's undecided]
3. ...

Pick a number, give me your own topic, or refine one of these.

Once the user picks, go to Fight Step 2 to confirm camps.

Rematch

Find a previous arena research doc:

• Search docs/research/*.md for method: containing "arena" or "adversarial"
• List matches with title and date

Previous arena matches:

1. [Previous match title] (YYYY-MM-DD)
2. ...

Pick one to re-run with fresh evidence, or give me a new matchup.

Once selected, extract the original camps from the doc and choose days using this rematch rule:

• If the prior arena doc states a days window, reuse it by default.
• If the user asks for fresher coverage, use 7 or 14.
• If the prior doc has no days metadata, route with the standard table in Fight Step 3.

Then go to Dispatch Beat Reporters with the same framing and chosen days.

Dispatch Beat Reporters

Launch one agent per camp, all in parallel, using the Agent tool:

• subagent_type: "newsroom:beat-reporter"
• run_in_background: true
• prompt: the Reporter Brief Template below (one per team)

Reporter Brief Template

You are a beat reporter for Team [LETTER]. Team [LETTER] LOVES [POSITION].
They believe [POSITION] is the best approach and the alternatives are inferior.

Your job: Find the STRONGEST evidence supporting Team [LETTER]'s position.
Emphasize evidence that supports [POSITION], and de-emphasize evidence that undermines it.
Use a search window of [DAYS] days (`--days=[DAYS]`).

Search Reddit, X/Twitter, and the web for:
1. Arguments FOR [POSITION]
2. Success stories and production proof points
3. Criticisms and failures of the OPPOSING position(s)
4. Hard data -- benchmarks, token counts, adoption numbers, performance metrics
5. Community sentiment favouring [POSITION]
6. Prominent developers/companies championing [POSITION]
7. Recent improvements or momentum (2025-2026)

Come back with the strongest possible case. Be biased in Team [LETTER]'s favour --
find every piece of ammunition. Include specific Reddit threads, tweets, videos, and blog posts
with engagement metrics (upvotes, likes, comments, views) where possible.

Tell the user reporters are dispatched and wait for all to return.

The Verdict

Once all reporters have filed, put on the impartial staff engineer judge hat.

Scoring Rubric

Rate each team on 5 dimensions (1-5 scale):

Verdict Format

# Arena: [Team A Position] vs [Team B Position]

## Round-by-Round Scoring

### Round 1: Hard Data
Team A: [score]/5 -- [one-line justification]
Team B: [score]/5 -- [one-line justification]

### Round 2: Production Proof
...

### Round 3: Ecosystem Momentum
...

### Round 4: Developer Experience
...

### Round 5: Future Trajectory
...

## Final Score
Team A: [total]/25
Team B: [total]/25

## The Verdict
[2-3 paragraph synthesis. Who wins overall? Where does each side
have genuine strengths? What's the pragmatic answer?]

For three-camp matchups (X vs Y vs Z), use the same format with Team C added to each round. The final score becomes [total]/25 for each of three teams.

Keep the personality of a staff engineer who enjoys a good debate -- witty but rigorous. The scores must be honest, not diplomatic ties.

Success Criteria

A good arena verdict meets these bars:

• Every score has a concrete justification (not "Team A is slightly better")
• At least 3 of 5 rounds cite specific data points from the reporters
• The final verdict names a winner (no diplomatic ties unless the data genuinely splits)
• Open Questions section surfaces at least one unresolved tension worth future research

Save as Cortex Research Doc

After presenting the verdict, ask whether to save. If yes:

Use the /cortex-engineering:frontmatter skill for correct frontmatter structure. If that skill is unavailable, use this minimal template:

---
title: "Arena: [Team A] vs [Team B]"
type: research
status: complete
date: YYYY-MM-DD
method: adversarial arena research (N parallel beat reporters + judge synthesis)
sources: [reddit, x-twitter, web]
tags: [arena, topic-a, topic-b]
---

Then:

• Location: docs/research/YYYY-MM-DD-<slug>.md
• Structure: Follow the Cortex research doc type (Summary, Key Findings, Details, Sources, Open Questions)
• Details section: Expand the round-by-round into full prose with evidence tables, quotes, and source links from each reporter
• Sources section: Split by team, include engagement metrics, include YouTube if found
• Open Questions: Extract genuinely unresolved tensions from the debate