System Architecture Harness
Evidence-backed architecture decisions, trade-offs, failure paths, and proof.
Free
Works with the AI Agents you already use
Claude Code
Cursor
Codex CLI
Muse
Dots+22 moreClaude CodeCursorCodex CLIMuseDots
OpenClaw
Manus
Grok
Grok Bot+18 agentsSystem Architecture Harness
Example session with this skill installed
Think through a production order and inventory platform. We have six backend engineers, a peak of 2,000 create-order requests per second, finite stock with a no-oversell requirement, prepaid payments, and a one-region launch. Our budget ceiling is $18,000/month; target recovery is 60-minute RTO and 5-minute RPO. Compare a modular monolith with microservices. Recommend a launch shape, state the core invariants and trade-offs, identify failure and security risks, and give launch evidence gates plus unresolved questions. Treat the supplied targets as assumptions until measured.
- Read your context and instructions
- Compiled the system architecture harness
Illustrative result based on the skill's bundled fictional order-platform example. The workload and recovery targets are assumptions, not measurements.
Summary
Start with a region-local modular monolith backed by PostgreSQL as the order and inventory authority. Commit order acceptance, stock reservation, status history, idempotency, and an outbox record in one transaction. Keep notifications and analytics asynchronous. Do not adopt microservices or active-active writes until measured constraints justify their coordination and operating cost.
Decision and trade-offs
| Decision | Why it fits | Cost or reversal trigger |
|---|---|---|
| Modular monolith with explicit module ownership | Six engineers can operate one deployable; a single transactional authority protects order and stock invariants. | Modules cannot scale independently. Revisit if a domain needs distinct ownership, SLOs, or sustained resource headroom. |
| Transactional outbox for side effects | Avoids a database/broker dual write and keeps provider latency off checkout. | Consumers must deduplicate, tolerate lag, and support replay and reconciliation. |
| One primary region with tested cross-region recovery | Keeps write correctness simpler for launch. | Regional recovery is not proven until restore and failover drills meet the 60-minute RTO and 5-minute RPO. |
Invariants and failure behavior
- One tenant/idempotency key produces at most one semantic order.
- Inventory reservation and order acceptance commit atomically; no accepted order can oversell stock.
- If a payment call times out, its result is unknown until provider lookup or reconciliation resolves it. Never blindly retry an ambiguous charge.
- Bound queue depth, retry count, concurrency, and deadlines. Deduplicate events and monitor consumer lag.
- Enforce tenant authorization at data access, not only in the UI or gateway.
Evidence gates before launch
Confirm payment timeout, idempotency, lookup, and refund semantics. Load-test the stated workload and contention profile. Exercise duplicate requests, outbox replay, provider failure, and reconciliation. Restore a backup and measure RTO/RPO. Verify tenant isolation and rehearse rollback. Keep launch blocked if a critical gate fails.
Open questions
The order-size distribution, inventory hot-key rate, measured capacity, payment sandbox behavior, recovery duration, and actual monthly cost are still unknown. Assign owners and collect evidence before approval.
Connects securely to your tools. The creator never sees your data.
About this skill
Think Through Production Systems turns an idea, existing codebase, incident, or migration into architecture that can be built, operated, reviewed, and evolved.
Work in four modes: Think for a decision record, Review for prioritized findings, Change for bounded work with rollback, and Verify for observed checks and residual risk. The workflow covers requirements and workload, invariants and ownership, interfaces and failure behavior, scale and recovery, security and privacy, operations, migration, economics, and validation.
Every material decision needs a motivating requirement, invariant, risk, or constraint; a realistic alternative; the trade-off and consequence; and a validation path. The skill challenges unbounded retries, unsafe dual writes, unsupported scale claims, weak authorization boundaries, and recovery claims without evidence.
This package includes 24 reference files, 6 papers, 3 fictional worked examples, and 10 templates for architecture specs, ADRs, SLOs, threat models, risk registers, failure analysis, API/event contracts, and evidence maps. It is one skill from ArcForge's 18-skill library. ArcForge's site reports 147 reference papers and 81 behavioral evaluation cases across the library.
Built for Claude Code and Codex. The package is Markdown instructions and reference material; it needs no API key or background service. It does not guarantee correctness or replace security testing, legal review, or domain experts.
Source: d4rkNinja/arcforge v0.4.6 (MIT).
How to use
Skills on Agensi run inside your AI agent. Copy the prompt, paste it into your agent, and it connects to Agensi and installs this skill for you. If a payment is needed, your agent shows it to you first.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill passes our trust checks below.
Security scanned
Passed our 8-point scan before listing
Fresh listing
Recently published to Agensi
Free forever
No account required to browse
Trust & safety
Security scanned
Verified clean 13 days ago
- Free to download with an account