Roast

Roast: Devil's Advocate Analysis

Overview

This skill produces evidence-based constructive critique through 5 specialized HackerNews commenter personas: Skeptical Senior, Well-Actually Pedant, Enthusiastic Newcomer, Contrarian Provocateur, and Pragmatic Builder. The workflow spawns these personas in parallel, validates all claims against actual files and lines, and synthesizes findings into an improvement-focused report.

Key constraints baked into the workflow:

• CLAUDE.md must be read and followed before analysis begins
• Read-only mode (no Write, Edit, destructive Bash) is mandatory — enforced via read-only-ops skill invocation
• Every claim must reference specific file:line locations and be validated against actual evidence before appearing in the final report
• All 5 personas must complete before validation begins — no partial analysis
• Final report must include both validated strengths and problems, prioritized by impact
• Unvalidated claims are dismissed; unfounded critiques are shown with evidence explaining why
• Analysis must be direct and focused — no elaborate frameworks beyond the 5-persona + validation pattern
• Sarcasm and mockery are stripped during synthesis; technical accuracy and file references are preserved

Deep References

Instructions

Phase 1: ACTIVATE READ-ONLY MODE

Goal: Establish guardrails before any analysis begins.

Call the Skill tool with read-only-ops.

skill: read-only-ops

This ensures no modifications can occur during the analysis workflow.

Allowed operations:

• Read tool for file contents
• Glob tool for file patterns
• Grep tool for content search
• Bash: ls, wc, du, git status, git log, git diff

Forbidden operations:

• Write tool -- no file creation
• Edit tool -- no file modification
• Bash: rm, mv, cp, mkdir, touch, git add, git commit, git push

If read-only mode cannot be activated, stop immediately. Never proceed with unguarded analysis.

Gate: Read-only mode active. Proceed only when gate passes.

Phase 2: GATHER CONTEXT

Goal: Understand the target thoroughly before spawning critical perspectives.

Step 1: Identify target type

Step 2: Read key files

Use Read tool to examine: README.md, main documentation, key implementation files relevant to the target.

Step 3: Survey structure

Use Glob to map the landscape:

• **/*.md for documentation coverage
• Source code organization and entry points
• Configuration files and dependency declarations

Step 4: Search for patterns

Use Grep to find: specific claims to verify, usage patterns, dependency references, related test files.

Step 5: Ground verbal descriptions

If user describes a concept rather than pointing to a file, search the repo for existing implementation. Critique grounded in actual code beats critique of a strawman every time. Never analyze a verbal description without confirming the code exists.

Gate: Target identified and sufficient context gathered. Proceed only when gate passes.

Phase 3: SPAWN ROASTER AGENTS (Parallel)

Goal: Launch 5 agents in parallel, each embodying a roaster persona, analyzing the target with full evidence-gathering discipline.

Launch 5 general-purpose agents in parallel via Task tool. Load the full persona specification from the corresponding agent file into each prompt.

The 5 parallel tasks:

1. Skeptical Senior (agents/reviewer-code.md, senior lens) Focus: Sustainability, maintenance burden, long-term viability

2. Well-Actually Pedant (agents/reviewer-code.md, pedant lens) Focus: Precision, intellectual honesty, terminological accuracy

3. Enthusiastic Newcomer (agents/reviewer-perspectives.md, newcomer lens) Focus: Onboarding experience, documentation clarity, accessibility

4. Contrarian Provocateur (agents/reviewer-perspectives.md, contrarian lens) Focus: Fundamental assumptions, alternative approaches

5. Pragmatic Builder (agents/reviewer-domain.md, pragmatic-builder lens) Focus: Production readiness, operational concerns

Each agent must:

• Call the Skill tool with read-only-ops. Do this first to enforce no-modification guardrails.
• Follow their systematic 5-step review process
• Tag ALL claims as [CLAIM-N] with specific file:line references
• Provide concrete evidence for every claim — vague critiques are worthless and must be rejected during validation
• Search for actual implementation details rather than analyzing verbal descriptions

Agent prompt template: Tell each agent its persona name, role, target, and these requirements: (1) call read-only-ops first, (2) follow their systematic 5-step review process, (3) tag ALL claims as [CLAIM-N] with file:line references, (4) provide specific evidence for every claim. Load the appropriate reviewer agent file for each persona's lens.

CRITICAL: Wait for all 5 agents to complete before proceeding to Phase 4. Do not begin validation on partial results. Every persona must contribute before synthesis can happen.

Gate: All 5 agents complete with tagged claims. Proceed only when gate passes.

Phase 4: COORDINATE (Validate Claims)

Goal: Verify every [CLAIM-N] against actual evidence before including in the report.

Collect and validate every [CLAIM-N] from all 5 agents.

Step 1: Collect all claims

Extract every [CLAIM-N] tag from all 5 agent outputs. For each, track:

• Claim ID and text
• Source persona (Senior, Pedant, Newcomer, Contrarian, Builder)
• Referenced file:line location

Step 2: Validate each claim

For each [CLAIM-N], read the referenced file/line using Read tool and assign a verdict:

Critical: You must read the file and check the line. Visual inspection misses nuance. "Obviously valid" is a rationalization word. Do not accept a claim because it sounds right or all personas agree on it — consensus is not the same as correctness.

Step 3: Cross-reference

Note claims found independently by multiple agents. If 3+ personas independently identify the same issue, escalate to HIGH priority regardless of individual severity.

Step 4: Prioritize

Sort VALID and PARTIAL findings by impact:

• HIGH: Core functionality, security, or maintainability
• MEDIUM: Important improvements with moderate impact
• LOW: Minor issues or polish

Gate: All claims validated with evidence. Proceed only when gate passes.