How this skill is triggered — by the user, by Claude, or both
Slash command
/aa-jl:developing-julia-packageThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
Notes on developing Julia packages.
Notes on developing Julia packages.
Assume the package is named MyPkg. Substitute your actual package name wherever MyPkg appears.
Use src/MyPkg.jl as the package entry point. Keep the module declaration, exports, and include list there; put substantial implementation in focused files under src/.
module MyPkg
export fit_model, FitResult, GaussianModel
include("models.jl")
include("fit.jl")
include("preprocess.jl")
end
Split files to improve readability, not to recreate Python-style class or submodule hierarchies. Prefer one public module unless there is a real user-facing namespace boundary.
Guidelines below.
exportsexport helpers that are only used internally just so tests can reach them. Prefer importing explicitly in tests:# test/runtests.jl
using Test
using MyPkg: <internal-only helper>
Test public behavior through the public API first. Import internals only when the helper has meaningful behavior that is hard to exercise through the public path:
using Test
using MyPkg
using MyPkg: initial_guess
@testset "fit_model" begin
result = fit_model(x, y; model = GaussianModel())
@test result isa FitResult
end
@testset "initial_guess" begin
@test initial_guess(x, y, GaussianModel()) isa NamedTuple
end
Prefer splitting behavior across methods instead of a large if/elseif chain on isa, unlike typical Python style.
# Do not write if else end
function f(x)
if x isa Integer
return 2x
else
return x
end
end
Instead, use multiple dispatch:
f(x) = x # generic implementation
f(x::Integer) = 2x # specialized implementation for x::Integer
For package APIs, make the dispatch object explicit and keep symbol options as a thin compatibility layer if needed:
abstract type AbstractModel end
struct GaussianModel <: AbstractModel
baseline::Bool
end
GaussianModel(; baseline = true) = GaussianModel(baseline)
fit_model(x, y; model::AbstractModel = GaussianModel()) =
fit_model(model, x, y)
function fit_model(model::GaussianModel, x, y)
guess = initial_guess(x, y, model)
# gaussian-specific implementation
end
Use a marker or configuration struct for the model choice, and use a separate result struct for fitted values. Do not mutate a model type into a mixed "algorithm plus fitted state" object unless that is clearly the public contract.
struct FitResult{P,T}
params::P
residuals::Vector{T}
converged::Bool
end
This keeps GaussianModel() as the method-selection/configuration value and FitResult as the returned fitted state.
.JuliaFormatter.toml at the repository root (or rely on defaults) so everyone applies the same rules.From the package root:
using JuliaFormatter
format(".") # formats src/, test/, etc. under the current directory
Run formatting before merging substantive edits; wire the same command into CI or pre-commit hooks if the team wants enforcement.
@benchmark / @btime from BenchmarkTools.jl rather than guessing.ArgumentError, DomainError, or other appropriate exceptions; messages should tell the caller what to fix.@doc when you attach documentation programmatically.Guides creation, editing, and verification of skills for AI coding agents using test-driven development with subagent scenarios. Use when authoring or debugging skills.
npx claudepluginhub atelierarith/atelier-arith-julia-development-skills --plugin aa-jl