Fast, low-cost coding agents get Nim’s ownership hooks, C interop and test setup wrong, and review and happy-path tests miss it. Verified Nim Skills drop into Claude Code, Codex, Cursor and other agents. Every testable claim in them is proven by a compiler-run test.
verified on Nim 2.3.1 · Nimony 0.6.3
proc `=wasMoved`(w: var Writer) =
w.data = nil # destroy is now a no-op...
# pos and capacity stay stale
var b = move(a)
a.writeByte 'z' # reusing a moved-from value is legalSIGSEGV: Illegal storage access. (Attempt to read from nil?)The next legal write goes through nil. The advice behind it, “=wasMoved only needs to make =destroy a no-op”, passes review and the destroy test. The verified skill resets every field, and a test reproduces the crash and proves the fix.
All three compile, look idiomatic and pass a quick test. Only running the code caught them.
A =copy guard that compares addresses misses aliasing, so the copy reads a buffer it just freed. The demo and the stress test still pass. Without the skill, the agent’s container failed this memory-safety check. With the skill, it passed.
=dup hooks that run on garbageCopy into freshly allocated raw storage and the compiler injects a =sink that treats uninitialized memory as a live value. Our verified skill had the right pragma for the wrong reason. A compiler-run test now pins the real rule.
Add a test file and the suite stays green, because the new test never runs. Without the skill, the agent’s runner listed the tests by hand. With the skill, the runner finds every test file.
The pattern holds across the catalog. On 12 of 17 skills, the agent without the skill failed a check that the skill-guided agent passed. Other failures include C-library builds that work on only one machine and fuzz harnesses missing required flags.
Four hard Nim tasks, each run five times with the skill and five times without: a bug report with a reproducer, a C-to-Nim port, cross-module error handling and a manually managed container. Every run was blind and sandboxed. Deterministic checks scored it, and the plan was fixed before the first run.
Every task solved in most runs.
Three of the four tasks never solved.
Four times cheaper per solved task. The skill-guided agent reads more per attempt and finishes far more often.
A larger model without skills passed 0 of 6 on two of the tasks. The audit is what you pay for: on three earlier tasks, the unverified drafts of the same skills passed 1 of 3 each, and the verified skills passed 3 of 3. Full results.
Skills that models wrote for themselves gave no average gain in SkillsBench. Curated skills added about 16 points. Our own careful drafts failed the compiler audit in over a hundred places.
Docs describe the hooks. They don’t warn your agent that its =wasMoved leaves stale fields, or that its address guard misses aliasing. Keep the MCP for lookup and add the skills for safety.
A frontier model costs more on every request and still has not seen the compiler build you installed last week. Verified guidance costs a flat monthly fee.
The test suite runs again on new Nim and Nimony builds, so compiler drift reaches you as a fix in your next update. The same tests catch compiler bugs: verifying our Nimony error-handling skill exposed seven, all fixed upstream within two days of our reports.
Debug builds raised. -d:release returned 42 as if nothing happened. Only a test in both modes catches that.
Errors truncated to “no error”, a handler that segfaults, a handler that catches the wrong type, and invalid generated C.
Planning for Nim 3? The catalog already covers Nimony error handling and compiler plugins.
The public skills stay free for non-commercial use. They are separate community work, not a free tier.
| Verified Nim Skills | Public skills | |
|---|---|---|
| Claims tested against the current compilers | All testable claims | Not systematically |
| Audit corrections applied and retested | Yes | Some synced, no guarantee |
| Nimony: error handling, plugins | Yes | Plugins and native release only |
| Re-verified on new compiler builds | While subscribed | Best effort |
| Use in commercial work | Internal use | Non-commercial only |
Catalog areas: ownership hooks, error handling, API design, C bindings and wrappers, testing, fuzzing, debugging, defect analysis, bug reporting, porting, performance, code organization, style, doc comments, Atlas, and Nimony error handling and plugins.
Works with Claude Code, Codex, Cursor, GitHub Copilot, OpenCode, Windsurf and more.
Check out with Polar. Invoices carry your company name and VAT ID.
Accept the read-only invite to the private catalog.
The open-source installer puts the skills where your agent looks, and nyx update pulls each re-verification. Setup guide.
One developer.
$10 / month, billed yearly · or $15 month to month
Founding members: $96 / year until 30 November 2026, kept while subscribed.
Two or more developers, one invoice.
$10 / seat / month, billed yearly · or $15 month to month
Minimum 2 seats. Seats are set up by email.
Agent-written Nim that crashes on an unusual path costs a review cycle when you are lucky and an incident when you are not. The skills target exactly those paths. We build and test them on low-cost models, so the agent you already pay for gets them right.
Each testable claim has a program that the real compiler builds and runs, and the suite runs again on new compiler builds. When a test exposes a wrong statement, including in our own earlier versions, we fix it and ship it in your next update.
Yes, and keep the MCP. Docs tell the agent what exists. The skills tell it what is safe on today’s compiler, what to avoid and how to isolate a failure.
Add a two-line pointer to AGENTS.md or CLAUDE.md and it will load the matching skill. Copy the pointer.
Anything NYX installs into: Claude Code, Codex, Cursor, OpenCode, GitHub Copilot, Crush, Pi, Cline, Windsurf, Roo Code, Continue and others. Full list.
Repository access ends at the end of the paid period and updates stop. You keep the versions you received under the license. NYX deletes nothing.
No refunds once repository access has been granted: the files are yours to keep. EU and UK consumers can withdraw for a full refund before access is granted. Statutory rights for faulty or undelivered content always apply. See the refund terms.
Give your agent Nim guidance the compiler has already checked.