Debate

Debate

Adversarial evaluation of the current state of a feature, system, or codebase area. Five agents argue from different angles, a moderator synthesizes a verdict. The output is a readiness report -- not a plan, not code.

Use this when you don't know if what you have is good enough.

Invocation

/debate [topic] [--scope "focus areas"] [--constraints "known constraints"]

Examples:

• /debate "API security posture"
• /debate "onboarding experience" --scope "auth, setup, first use"
• /debate "launch readiness" --constraints "feature X is gated behind flag"

Pipeline

Phase 0: Constraints Gathering (Main + User)

Before spawning agents, main MUST:

1. Ask the user for known constraints:

• What's in scope for this review?
• Any decisions already made that agents should respect?
• Known limitations or trade-offs to evaluate against?

1. Check memory for relevant project constraints
2. Compile a constraints block:

CONSTRAINTS (non-negotiable):
- [constraint 1]
- [constraint 2]
DO NOT propose solutions that violate these constraints.

This prevents the #1 failure mode: agents recommending something already ruled out.

Phase 1: Debate (Parallel Agents)

Spawn 5 agents. The 4 debaters run in parallel. Moderator waits for all 4.

| Agent | Role | Method | |-------|------|--------| | arch-critic | Code auditor -- finds gaps, bugs, half-baked features | Reads source code, tests, architecture docs | | comp-critic | Competitive benchmarker -- holds us to industry bar | Web search for competitors, best practices | | code-defender | Argues readiness from actual code evidence | Reads source code, test coverage, recent commits | | industry-defender | Argues competitive positioning from market context | Web search for market context, user expectations | | moderator | Synthesizes all 4 reports into a verdict | Receives reports via SendMessage, writes report |

Agent prompt template for debaters:

You are [ROLE] in an adversarial debate about: [TOPIC]
Scope: [SCOPE]

CONSTRAINTS:
[constraints block]

Your job: [ATTACK gaps / DEFEND readiness] with EVIDENCE.
- Read the actual code, don't guess
- Cite specific files, functions, line numbers
- Compare against [competitors / best practices]
- Be brutal -- no hedging, no "it depends"

Send your report (max 50 lines) to the moderator agent when done.
Format:
## [Your Role] Report
### Key Findings (numbered, max 10)
### Verdict: [READY / NOT READY / CONDITIONAL]
### Evidence: [file:line citations]

Moderator prompt:

You are the moderator. Wait for all 4 debate reports.

Once you have all 4, write the readiness report to:
docs/brainstorms/{TOPIC}_READINESS_REPORT.md

Report format:
# [Topic] -- Readiness Report
> Generated: [date] | Debate agents: 4 | Moderator synthesis

## Verdict: [READY / NOT READY / CONDITIONAL]
[2-3 sentence summary]

## Consensus Points (both sides agree)
- [point]

## Contested Points (sides disagree)
| Point | Critics Say | Defenders Say | Moderator Assessment |
|-------|------------|---------------|---------------------|

## Critical Issues (must fix before launch)
1. [issue] -- [evidence] -- [severity]

## Improvement Opportunities (nice to have)
1. [opportunity] -- [effort estimate]

## Raw Reports
[Append all 4 reports for reference]

Phase 2: Handoff (Main)

Once the moderator writes the report:

1. Read the verdict line only (1 line, not the full report)
2. Report to user:

``` Debate complete. Verdict: [VERDICT] Report: docs/brainstorms/{TOPIC}READINESSREPORT.md

To act on findings: /plan-features --doc docs/brainstorms/{TOPIC}READINESSREPORT.md ```

1. Clean up the team

That's it. Debate produces a report. Planning and execution are separate skills.

Main Context Discipline

• Phase 0: Conversation only. No tools except memory check.
• Phase 1: Only Agent/TeamCreate/SendMessage. Main does NOT read code, grep, or search.
• Phase 2: Read one line from the report (the verdict). Summarize to user.

Main's total context growth: ~30 lines. The entire debate happens outside main's context.

Anti-Patterns

1. Main reading the full report. The report can be 200+ lines. Main reads the verdict line, tells the user where the file is. The user or /plan-features reads the rest.
2. Debaters being polite. The whole point is adversarial tension. Critics should be harsh. Defenders should push back hard. Moderate positions produce useless reports.
3. Skipping Phase 0. Without constraints, agents will recommend prohibited libraries, impossible timelines, or solutions that violate business rules.
4. Bundling planning/execution. This skill produces a REPORT. If you want to act on it, run /plan-features. Separation keeps each skill focused and composable.
5. Moderator taking sides. The moderator synthesizes, doesn't advocate. If critics found 8 issues and defenders addressed 3, the moderator says "5 unresolved issues remain" -- not "defenders made a strong case."

Source & license

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

• Author: justnau1020
• Source: justnau1020/claude-os
• License: Apache-2.0

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