Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Implement Flexport reference architecture for supply chain integrations with best-practice project layout, service boundaries, and data flow. Trigger: "flexport architecture", "flexport project structure", "flexport system design".
.claude/skills/jeremylongshore-flexport-reference-architecture/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-01 | ✗→✓ | ▲ Improved | -9% | 0% |
| case-02 | ✗→✓ | ▲ Improved | 34% | 0% |
| case-08 | ✗→✓ | ▲ Improved | 59% | 0% |
| case-11 | ✗→✓ | ▲ Improved | 37% | 0% |
| case-13 | ✗→✓ | ▲ Improved | 40% | 0% |
Use three explicit provider planes: REST v3 for resource operations, MCP Streamable HTTP JSON-RPC for permissioned assistant tools, and signed webhooks for notifications. Join them only through durable application policy and reconciliation.
Show api.flexport.com, mcp.flexport.com/mcp, and public HTTPS webhook ingress as distinct external nodes.
Use endpoint-scoped OAuth clients per REST workload, authenticated MCP sessions with documented tool roles, and a separate webhook secret.
Put business approval and policy before booking/trade mutations; keep payload handling in minimized data services.
Persist operation keys, provider references, event dedupe state, cursor/link checkpoints, and redacted outcomes.
Connect events and uncertain outcomes to authorized REST/resource reads. Never make webhook delivery the sole source of truth.
Model token-cache failure, permission denial, additive schema change, duplicate/missing event, ambiguous mutation, and rollback.
REST calls authenticate with a cached OAuth 2.0 client-credentials Bearer token using audience https://api.flexport.com, or an explicitly accepted broad API key. Use distinct credentials per workload and never log credentials or tokens. MCP calls use the authenticated connection to https://mcp.flexport.com/mcp and remain subject to each tool's documented account permissions.
Use Read and Grep for discovery and evidence. Use Write or Edit only for the approved artifact, code, configuration, test, or receipt described by this workflow; do not make an unapproved Flexport-side change.
Return a machine-reviewable receipt in this shape; adapt the operation values, but never place credentials or provider payloads in it:
yamlsurface: rest-v3 operation: shipment-read decision: approved outcome: verified evidence: release_sha: recorded-out-of-band provider_reference: redacted rollback_owner: logistics-platform
A booking assistant queries rates through MCP, sends an exact candidate to an approval service, and records one operation key before booking. Signed events enqueue notifications, while a reconciler reads provider state to close gaps.
| Failure | Response | | --- | --- | | REST and MCP contracts collapsed | Split transport, auth/session, errors, and pagination before implementation. | | Webhook directly mutates core state | Insert verified durable enqueue and reconciliation. | | Approval follows booking | Move it before the provider mutation. | | Architecture logs payloads | Replace them with redacted operation receipts. |
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→pass | 32,617 | 24,664 | -24% | 1 | 1 | 0% | 4,893 | 4,475 | -9% | 0 | 0 | — |
case-02 | fail→pass | 16,531 | 15,720 | -5% | 1 | 1 | 0% | 2,995 | 4,028 | +34% | 0 | 0 | — |
case-03 | fail→fail | 25,756 | 27,469 | +7% | 1 | 1 | 0% | 4,105 | 6,172 | +50% | 0 | 0 | — |
case-04 | pass→pass | 24,933 | 18,002 | -28% | 1 | 1 | 0% | 3,264 | 3,806 | +17% | 0 | 0 | — |
case-05 | fail→fail | 11,517 | 11,310 | -2% | 1 | 1 | 0% | 1,928 | 2,847 | +48% | 0 | 0 | — |
case-06 | pass→pass | 16,180 | 20,105 | +24% | 1 | 1 | 0% | 2,553 | 3,397 | +33% | 0 | 0 | — |
case-07 | pass→pass | 11,774 | 14,391 | +22% | 1 | 1 | 0% | 2,017 | 3,480 | +73% | 0 | 0 | — |
case-08 | fail→pass | 12,932 | 18,226 | +41% | 1 | 1 | 0% | 2,078 | 3,309 | +59% | 0 | 0 | — |
case-09 | fail→fail | 20,705 | 21,753 | +5% | 1 | 1 | 0% | 2,579 | 3,658 | +42% | 0 | 0 | — |
case-10 | pass→pass | 20,504 | 13,620 | -34% | 1 | 1 | 0% | 2,377 | 3,109 | +31% | 0 | 0 | — |
case-11 | fail→pass | 12,274 | 17,668 | +44% | 1 | 1 | 0% | 2,231 | 3,046 | +37% | 0 | 0 | — |
case-12 | fail→fail | 18,463 | 17,355 | -6% | 1 | 1 | 0% | 2,460 | 3,291 | +34% | 0 | 0 | — |
case-13 | fail→pass | 17,784 | 16,717 | -6% | 1 | 1 | 0% | 2,294 | 3,214 | +40% | 0 | 0 | — |
case-14 | fail→pass | 13,171 | 12,213 | -7% | 1 | 1 | 0% | 2,427 | 3,172 | +31% | 0 | 0 | — |
case-15 | pass→pass | 15,497 | 15,964 | +3% | 1 | 1 | 0% | 1,932 | 2,830 | +46% | 0 | 0 | — |
case-16 | pass→fail | 9,424 | 8,009 | -15% | 1 | 1 | 0% | 1,741 | 1,454 | -16% | 0 | 0 | — |
case-17 | fail→pass | 18,969 | 11,329 | -40% | 1 | 1 | 0% | 2,137 | 2,956 | +38% | 0 | 0 | — |
case-18 | fail→fail | 16,244 | 9,591 | -41% | 1 | 1 | 0% | 2,272 | 2,777 | +22% | 0 | 0 | — |
case-19 | fail→pass | 16,040 | 15,844 | -1% | 1 | 1 | 0% | 2,047 | 3,178 | +55% | 0 | 0 | — |
case-20 | fail→fail | 20,471 | 20,342 | -1% | 1 | 1 | 0% | 3,446 | 4,580 | +33% | 0 | 0 | — |
case-21 | fail→fail | 19,222 | 23,003 | +20% | 1 | 1 | 0% | 4,533 | 5,249 | +16% | 0 | 0 | — |
case-22 | fail→fail | 28,432 | 30,297 | +7% | 1 | 1 | 0% | 4,946 | 6,878 | +39% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted. The headline lift of +32 percentage points is the difference between those two pass rates over the 22 comparable cases. 1 case got worse with the skill loaded, and it is included in that figure.
The publisher has shipped newer versions since this run, so these numbers describe v1, not the version currently listed.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.