BlankVisuals© Validator — evidence-backed validation for Cursor skills, agents, and agentic repos (open source, MIT)

Showcase Built for Cursor
, ,
1

Hey everyone,

I kept shipping :cursor: Cursor agents that looked correct on inspection and broke at runtime. Wrong tool path in one place, right in another. Missing terminal condition. Scope broader than needed. Structural issues that don’t show up until the agent is actually running.

So I built a validation framework to catch that layer before deployment.

What it does :backhand_index_pointing_down:

BlankVisuals© Validator (FW-001) is a Cursor Agent Skill that validates any agent, skill chain, MCP orchestration, or agentic repo before deployment. It doesn’t guess — every rule requires verified evidence
(command output, file citation, or tool result) or returns INCONCLUSIVE.

There’s no soft pass. On critical rules, INCONCLUSIVE blocks the result the same way FAIL does.

It covers:

  • Structure (entrypoints, skill frontmatter, canonical docs)
  • Topology (terminal conditions, role boundaries, human gates)
  • Contracts (inter-step I/O, schema validation)
  • Security (secrets, hook scope, tool allowlists)
  • Observability (logging, output paths, error surfacing)
  • Runtime (verify commands, MCP smoke tests)

How to use it

Install as a personal skill:

git clone https://github.com/cengicadis/blankvisuals-validator \
  ~/.cursor/skills/blankvisuals-validator

Then in any Cursor chat:

“Validate this repo against FW-001”
“Validate this MCP orchestration against FW-001”
“/validate this orchestration workflow”
“Run BlankVisuals validator — precise & verify”

The agent stamps framework meta at the start, runs parallel checks where safe, sequential where order matters, and delivers a validation canvas + chat summary at the end.

:pushpin: Real example — doc-cleaner agent validated

Ran the validator on another agent I built. Results in under :hourglass_done: 3 minutes:

  • Target: doc-cleaner v1.0.0
  • Framework type: skill-chain + heuristic-workflow
  • Rules evaluated: 21
  • Result: PASS :white_check_mark: — 18 pass · :cross_mark: 0 fail · :warning: 3 minor warnings

Two real findings:

  • CON-01: Tool path mismatch — agent definition calls tools/doc-cleaner/chunker.py but actual path is tools/chunker.py. clean.sh had the right path; direct Cursor invocation would have failed silently.

  • RUN-01: No VALIDATION.md — missing documented verify command.

Neither was obvious from a manual review. Both fixed in under :hourglass_done: 5 minutes.

What you get

  • Per-rule YAML evidence blocks (command output / file citation / tool result)
  • PASS / FAIL / INCONCLUSIVE per rule — no invented results
  • Validation canvas in canvases/ directory
  • Chat summary with critical failure IDs

Repo:

MIT, open source. Rule catalog is in reference.md — each rule has a check, pass condition, and fail condition. Custom rules via VALIDATION.md in your repo.

Would love feedback on the rule catalog — especially if anyone has edge cases from their own skill chains, subagent graphs, or MCP orchestrations.