Six Sigma in R — AI Skill

Version 1.0.0 — released 2026-04-17 (see VERSION)

A reusable skill that teaches Claude Code (and compatible Codex-style agents) how to produce correct, runnable R code for typical Six Sigma / statistical process control tasks.

When this skill is installed, asking your coding agent something like "plot a P chart from this CSV" or "is this process capable?" yields an R script (or a reusable function) that uses the right formulas and constants, not approximations.

Release notes

1.0.0 (2026-04-17)

Initial release. Skill covers the full code-based Six Sigma body of knowledge listed below, plus a method-selection layer (by DMAIC phase, by problem type, by data shape) that recommends which technique to reach for when the user describes a practical problem instead of naming a specific test or chart.

Evaluation summary (Claude Opus 4.7):

• On mainstream Six Sigma tasks, the skill and a bare model produce equally correct R (30/30 = 30/30 across 5 prompts).
• On harder tasks (correct SPC constants for n=5, full Nelson rules, non- normal capability, within-vs-overall sigma for subgrouped Cp/Cpk), same result (33/33 = 33/33 across 5 prompts).
• On method-selection + implementation prompts the skill wins by +2 on 36 assertions (36/36 vs 34/36). The differentiators: correct analysis-order sequencing (stability before capability) and explicit scope transparency on out-of-scope topics (MSA / Gage R&R).

Full benchmarks, outputs, and side-by-side viewers live under eval-workspace/iteration-{1,2,3}/.

What the skill covers

Behaviour conventions baked in

• Base R by default for plotting; ggplot2 only on request.
• Script vs function and file vs inline plots — the skill asks the user once when it's ambiguous, then sticks to the choice.
• Proper control-chart constants are used (A2, D3, D4, d2, A3, B3, B4, c4); LCLs are clamped to 0 where the statistic is a count or rate.
• Cp/Cpk (within-subgroup σ, short-term) and Pp/Ppk (overall σ, long-term) are treated as distinct — not conflated.
• Hypothesis-test scripts compare p.value to alpha and print a verdict, not just the raw test object.
• Every reference sheet explains why a choice is made, so edge cases stay judgeable.

Repository layout

.
├── README.md                 # this file
├── INSTALL.md                # install for Claude Code or Codex, global or per-project
└── skills/
    └── six-sigma-r/
        ├── SKILL.md          # the skill entry point
        └── references/
            ├── method_by_dmaic.md         # "we're in the Analyze phase, what tools?"
            ├── method_by_problem.md       # "customers are complaining about X"
            ├── method_by_data_shape.md    # "I have this kind of data, what fits?"
            ├── descriptive_stats.md
            ├── charts_visualization.md
            ├── control_charts.md
            ├── capability_sigma_level.md
            ├── distributions.md
            ├── correlation_regression.md
            ├── hypothesis_tests.md
            ├── anova.md
            └── sample_size_power.md

The three method_by_*.md files are selection aids — they take a practical framing (a DMAIC phase, a business problem, a data shape) and recommend which technique to use, forwarding to one of the code-level references. The skill uses them when the user hasn't named a specific technique.

Prerequisites on the user's side

The skill produces R source code — running it needs a local R installation:

1. Install R: https://cran.r-project.org/
2. Add R's bin/ folder to PATH so Rscript is available from the shell.
3. For ggplot2-based requests, install the package once: install.packages("ggplot2").
4. For skewness / kurtosis without hand-coded formulas: install.packages("e1071").
5. For ANOVA power (pwr.anova.test): install.packages("pwr").

The skill does not assume any package beyond base R; extra packages are only brought in when the user's request requires them.

Installation

See INSTALL.md for step-by-step instructions to install the skill globally (available in every project) or locally (scoped to a single project), for both Claude Code and Codex.

License

The code and skill content in this repository are distributed under the Apache 2.0.