Menu
Multi-expert architecture review — boundaries, integration patterns, failure modes, evolvability
Instalar com Codex ou Claude Copie este prompt, cole no Codex, Claude ou outro assistente e deixe que ele revise a página da skill e instale para você.
Baseado na classificação ocupacional SOC
Multi-expert AI/ML specification review with scoring gate — model architecture, evaluation rigor, safety/alignment, and production readiness for AI systems and LLM apps
Multi-expert business analysis with advisory recommendations (no scoring gate)
Multi-expert intelligence pipeline review — discovery quality, ingestion resilience, scoring validity, platform compliance, taxonomy coherence, cost efficiency
Multi-expert mobile/native app specification review with scoring gate — Android, iOS, Swift, SwiftUI
Multi-expert personal-development review with scoring gate — learnability, adoption, human-centeredness, and capability impact for talent/learning/AI-augmentation designs
Multi-expert research and discovery panel with source evaluation, API feasibility, collection strategy, and intelligence gap analysis. Use when evaluating data sources, APIs, scraping strategies, OSINT collection plans, or any research/discovery effort that needs expert validation before implementation.
| name | sh-architecture-panel |
| description | Multi-expert architecture review — boundaries, integration patterns, failure modes, evolvability |
/sh:architecture-panel [content|@file] [--mode discussion|critique|socratic|debate] [--focus boundaries|integration|reliability|evolution] [--experts "name1,name2"] [--iterations N] [--verbose]
--verbose): Full expert deliberations, cross-expert dialogue, reasoning traces, and detailed per-expert analysis before scores and findings.Silent mode still performs full internal analysis — quality is preserved, only the output is compressed.
/Users/jcords-macmini/projects/20_agentflow/experts/PANEL_PROTOCOL.md and apply its Grounding and Refute Stage sections. This is load-bearing — findings that are not grounded per the protocol, or that do not survive the refute stage, MUST NOT be reported./Users/jcords-macmini/projects/20_agentflow/experts/panels/architecture-panel.yaml for panel definition, focus areas, auto-select rules, and scoring config (absolute path — relative paths fail when CWD is outside agentflow)./Users/jcords-macmini/projects/20_agentflow/experts/individuals/ for each selected expert.auto-select keywords — add matching experts up to max-experts: 6 cap.--focus area or use default-experts from panel YAML. --experts override replaces defaults entirely.Experts are defined as individual markdown files in /Users/jcords-macmini/projects/20_agentflow/experts/individuals/. Each file contains structured frontmatter with:
The panel YAML (/Users/jcords-macmini/projects/20_agentflow/experts/panels/architecture-panel.yaml) defines:
--mode discussion)Collaborative architecture refinement. Experts build on each other's observations sequentially — boundary critique flows into integration concerns flow into failure-mode review.
--mode critique)Systematic review with severity-classified findings (CRITICAL / MAJOR / MINOR). Each finding includes expert attribution, the architectural assumption it challenges, a concrete recommendation, and the cost of leaving it unfixed.
--mode socratic)Foundational questioning — "what's the seam?", "where does this fail?", "how does this change look in two years?". No direct answers; forces the author to defend the design.
--mode debate)Two-camp adversarial format. Experts split into pro-design and skeptic positions, exchange three rounds, and the panel runner synthesizes the disagreement into explicit trade-off statements.
4 dimensions, each scored 0-10:
| Dimension | Description |
|---|---|
| Boundary clarity | Are module/service boundaries explicit, with one-way dependency arrows and unambiguous contracts at each seam |
| Integration correctness | Are integration patterns appropriate; idempotency, ordering, and delivery guarantees handled explicitly |
| Failure-mode coverage | Are failure modes named and addressed; timeouts, retries, circuit-breakers, fallbacks specified where needed |
| Evolvability | Can the design absorb foreseeable change without rewrites; versioning and migration paths exist |
Pass threshold: overall score >= 7.0
Architecture review document containing:
AUTO-FIX, NOT SYNTHESIS-ONLY — this panel produces the analysis AND then applies fixes for every finding (high, medium, and low) automatically, per 00_Governance/CLAUDE.md §8 Panel Auto-Fix Policy. It never asks which findings to apply and never presents a menu: it fixes everything, then reports what changed. A below-gate score means fix the findings and re-run, not stop and ask.
Fix ALL findings automatically — high, medium, and low severity. Do not ask which findings to fix. Do not present a menu or an a/b/c choice. Fix everything, then report what was changed. This overrides any softer "synthesis only" framing. Per 00_Governance/CLAUDE.md §8.
The quality gate (quality_gate.run_stage3_panel) consumes this panel via
claude -p and reads the score from stdout, not the exit code. Therefore
the final line of your output MUST be exactly one of:
PANEL-VERDICT: <overall_score>
PANEL-VERDICT: FAIL: <machine_reason>
<overall_score> is the numeric overall score (0–10, one decimal — e.g. 8.3).
The gate applies the pass threshold (default 7.0); do NOT pre-apply it — just
report the score you computed.FAIL: <machine_reason> form only when no score could be produced
(structural failure) — snake_case naming the first blocker (e.g.
no_content, panel_config_missing, experts_unavailable).panel_no_verdict (inconclusive).