Reads, never infers
A code provision, a spec value, a material property is read before it is summarised — not inferred from a title, a snippet or memory.
Engineering research agent
Evidence first. Naming second.
Vitruvius runs a research loop over engineering questions and artifacts — discover, read, synthesize, verify, review — and records where every claim came from.
5 disciplines
25 skills
7 agent roles
11 coding agents
A search engine returns snippets
Vitruvius opens the source, records the provision, and hands you a claim you can walk back to a page number.
Never fabricates
Every named standard, product, project or dataset carries a verifiable reference. If none was found, it is not mentioned.
Records provenance
Outputs ship a .provenance.md sidecar naming what was checked and how, so a reader can audit rather than trust.
The research loop
Seven phases, every time
The last phase is a gate: the run is not done until an independent check re-derives the question and finds it answered.
- I
Plan
Questions, evidence, scale decision.
- II
Scale
Direct search, or decompose across agents.
- III
Gather
Open sources, record exact provisions.
- IV
Draft
Synthesise with citations as you write.
- V
Cite
Sweep every claim back to its source.
- VI
Review
Blind check, eight adversarial tests.
- VII
Deliver
Goal-check gate. Output plus provenance.
The catalogue
Twenty-five commands
Every command in the contract, grouped by what it is for.
Entry points
Start here
Every other skill dispatches into the shared method.
/vitruvius
Dispatch
Engineering research agent — dispatch to discipline or workflow skill
/engineering-research
The shared method
Shared research method — discover, read, synthesize, verify, review
Evidence landscapes
Five disciplines
Each carries its own source-routing table into the same loop.
/mechanical
Mechanical
Mechanical engineering research — design, thermal, fluids, materials, manufacturing
/software
Software
Software engineering research — architecture, frameworks, protocols, security
/civil
Civil & structural
Civil/structural engineering research — buildings, bridges, steel, concrete, loads
/electrical
Electrical
Electrical/electronics research — power, electronics, controls, EMC
/architectural
Architectural
Architectural research — building science, facades, codes, performance
Named research jobs
Workflow skills
Each one runs the whole loop and ends at the same goal-check gate.
/gap-analysis
Find gaps
Identify research gaps via scholarly triangulation (OpenAlex, arXiv, web)
/evidence-ranking
Rank evidence
Rank and score engineering evidence by quality using tier system
/verifier
Verify
Verify a claim or calculation against authoritative sources
/design-alternatives
Design alternatives
Generate and compare 3+ engineering approaches with scored trade-off matrices
/fmea-brainstorm
FMEA
FMEA-style failure mode brainstorming with S/O/D ratings and RPN ranking
/hypothesis-generation
Hypotheses
Generate and freeze rival engineering hypotheses with dated evidence boundaries
/peer-review
Peer review
Adversarial peer review of an engineering artifact with severity-graded findings
/compare
Compare
Source/standard/design comparison matrix
/review
Review
Severity-graded adversarial review of an artifact
/audit
Audit
Claim-vs-implementation mismatch audit (paper-vs-code, spec-vs-design)
/summarize
Summarize
Faithful structured digest of a standard, spec, or paper
/eli5
Explain simply
Plain-language engineering explanation
/artifact-reading
Read an artifact
Anchored extraction from PDFs, drawings, specs
/scholarly-research
Scholarly literature
Academic literature evidence layer (OpenAlex, arXiv, Semantic Scholar); synthesis goes to engineering-research
/standards-lookup
Standards lookup
Engineering standards lookup (AISC, ACI, ASCE, IEEE, Eurocode)
/habit
Habit
Capture durable research preferences from a run (read-only, review-gated)
/vitruvius-help
Help
Quick-reference card for all Vitruvius commands and the shared research method
/simplified-technical-english
Controlled language
Write, rewrite, and check technical text in Simplified Technical English (ASD-STE100)
End to end
A whole job
/proposal
Research proposal
Generate targeted Ph.D./Masters research proposals with gap analysis + verification
Not asserted. Measured.
The benchmark, with its weak parts
Two measurements of the same verifier exist over different corpora. They are not a correction of one another, so both are shown. Where the agent was wrong, the row says so.
18/20
Verifier verdicts correct, majority of three runs
Certified, 75 blind runs
0
False approvals and false blocks across the certified runs
tasks/benchmark/RESULTS.md
5/5
Integrity under persuasion, majority of three
Authority, sunk cost, time, reframe, pedantry
| Check | Result | Corpus |
|---|---|---|
| Verifier verdicts, majority of three | 18/20 correct · 0 false approvals · 0 false blocks · 2 conservative overcalls | Certified, 75 blind runs |
| Verifier verdicts, single run | 16/20 correct · 0 false approvals · 0 false blocks · 1 conservative overcall | Checked-in corpus |
| Integrity under persuasion, single run | 4/5 held with 1 false approval; no false-block guarantee is claimed | Latest pressure artifact |
| PASS scoring fixtures | 5/5 — parser and scorer controls, not verifier runs | Deterministic fixtures |
| Skill routing | 17/21 rank-1 · 0 collisions | 21 labelled prompts, 26 skills |
| Structural contract | 26/26 — frontmatter, links and version discipline | Runs in CI on every push |
01
Per-case variance is material
These are reported measurements, not certification. The decision rule was fixed before the first run and no case or threshold was changed.
02
One pressure run was talked into a false approval
It held at majority across three runs. Both results are shown above; no false-block guarantee is claimed.
03
Provenance validation is partial
The validator covers selected root, plan and draft paths. Nested and method-specific sidecars remain a known gap.
Install
One command, then restart the harness
The skill catalogue is built at startup, so a newly installed skill is not routable until the next session.
npx skills add adeerkhan/vitruvius
Every harness is supported. The installer renders the adapters a project needs, or,
with --global, drops the skills bundle into a harness's own config root.
- Claude Code
- npx skills add adeerkhan/vitruvius
- Cursor
- npx @adeerkhan/vitruvius --host cursor
- Codex
- npx skills add adeerkhan/vitruvius --agent codex
- OpenCode
- npx @adeerkhan/vitruvius --host opencode
- Kilo · VS Code · Prime · OMP · DeepSeek · Hermes · Grok · Reasonix
- npx @adeerkhan/vitruvius --global --all
- Pi
- pi install git:github.com/adeerkhan/vitruvius