One-time purchase
MBE3D Systems Engineering Skill Pack
Your AI agent already writes code. What it does not know is the systems engineering around it: how to write a requirement that survives a design review, how to trace that requirement to the evidence that closes it out, and how to work inside MBE3Dstudio without quietly corrupting your model. Ten skills, written in the form an agent reads.
Works with Claude Code, Cursor, Codex, Copilot · No subscription · No dependencies
The problem
A general-purpose agent writes requirements that read well and fail review
Ask any assistant for requirements and what comes back is compound shall statements,
performance claims with no units, and design decisions dressed up as requirements, none of it
tracing anywhere. It reads well enough to circulate, and you find out in review. The gap is not
intelligence; it is the standards and conventions a working systems engineer carries around and a
general model has never been shown.
What’s inside
Ten skills, from a requirement to a CI gate
Five work on your own files. Five connect to MBE3Dstudio and the harness that drives it.
One of them, mbse-requirements, is also
free on its own.
Five skills · no server
Methodology: no server required
-
mbse-requirements
The five EARS patterns, permanent ID schemes, and a banned-word list. Repairs show you the diff.
-
mbse-safety-analysis
MIL-STD-882E severity, probability and RAC; FHA, FMEA and the mitigation order of precedence. An unverified mitigation earns no probability reduction.
-
mbse-traceability
Six audit queries (unverified requirements, orphans, unmitigated hazards, dead tests) and change-impact that prices the re-verification too.
-
mbse-verification
Method selection, machine-evaluable pass criteria, and an evidence schema for ML and autonomy. A retrain, dataset or ODD change invalidates the record.
-
mbse-model-format
A portable, diffable
.mbse.jsonfile and a validator that catches dangling edges, cycles and enum casing before a server sees them.
Five skills · with MBE3Dstudio
Platform operation: verified against a live server
-
mbe3d-connect
Which of the two API surfaces answers, the bodyless 401, the two separate LLM configurations, and six documented failure modes.
-
mbe3d-load
The write-path decision that determines whether your attributes survive ingest, plus a loader, an undo manifest, and a rollback table that admits what cannot be undone.
-
mbe3d-analyze
SPARQL over the real
urn:mbe:*vocabulary, so “every mass nobody confirmed” is one query. Trace, impact, competency suites, CI gates. -
mbe3d-studio
Assembly tree, scene, coverage rollups and export readiness, over a provenance model where only
statedandderivedcount as specification. -
mbe3d-harness
New in 2.1. The agent runtime beside the model: which plugins mounted and why one did not, and the append-only session log (nineteen event kinds, a typed actor on every one, replay and O(1) fork) that is the audit trail for what an agent actually did.
How it was written
Every platform behaviour was probed against a running server, including the traps
Documentation tells you what an API is meant to do. This pack records what it actually did, because every endpoint was called and the results written down, broken behaviour included, as a failure mode rather than an omission, and withdrawn again in a later version when it was fixed. The same treatment gets applied to the pack itself: version 2.1 exists mostly because a re-audit against live systems found advice in it that returned nothing. Five of the findings:
- The model lives in a named graph, so a perfectly correct SPARQL query against the default graph returns nothing at all. Measured, the same count answered 0 against a store holding 411,579 triples. Zero bindings reads as “no unverified requirements” and means “nobody asked”. Every query in the pack now names the graph, and the health check distinguishes three different causes of nothing.
-
A typed requirement cannot be linked by the route that links everything else.
/entity/linkanswers404 not foundfor an id that resolves perfectly elsewhere, so a load that looks clean can leave a model with no trace web at all. Loading the reference model wired 4 of 18 relations that way, and 16 of 18 through the route that works. The two refusals that remain are documented with their reasons. - There are two LLM configurations, one per API surface. Set only one and the agent panel looks healthy while the generation endpoints silently return nothing.
-
A request with no session header lands in the public graph, as does everything once
VECTORMBE_DEMO_SHARED_WORKSPACE=1is set, andVECTORMBE_ISOLATE_BY_SESSIONis not the switch that controls it. - A typed requirement is not an entity, so it has no delete path: bulk delete reports zero removed while it sits there. Load into a scratch deployment, or accept that a mistake stays.
None of it is in the API documentation, because none of it is intended behaviour.
Questions
Before you buy
Do I need MBE3Dstudio?
No. Five work on your own files with nothing installed. The other five need an MBE3Dstudio server, or the harness beside it, licensed separately.
Which AI assistants does it work with?
Any agent that reads Markdown instructions: Claude Code, Cursor, Codex, Copilot, Windsurf, or a plain AGENTS.md. There is no runtime to install.
Is this a substitute for a certified process?
No, and the pack says so in writing. Its guidance on MIL-STD-882E, ARP4761, DO-178C and ISO 26262 is a working summary; the governing document is always the standard as issued. It makes a competent engineer faster, it does not replace one.
What exactly do I receive?
mbe3d-se-skillpack-v2.2.0.zip, 86 KB: the single-file edition (copy one file into your skills directory), ten skill directories, two Python scripts, a 14-node reference model, the README and the license. It downloads the moment checkout completes, and that link keeps working.
Refunds?
If the pack is not what this page describes, email us within 14 days and we will refund you.
Can I put it in my company’s shared repo?
The $29 licence covers one engineer, on any number of machines, including your own client work. For a team, ask us for a site licence. Neither permits publishing it outside your organisation.
Your agent can be taught to get this right the first time
Copy one file into your agent's skills directory and everything it writes from then on follows the same conventions your reviewers already use.