30 Writing a skill for analysis
Let’s turn the spec sheet from Section 3 into something reusable. The spec sheet answered “how do I analyse topa vs. coral cover, this one time.” A skill answers “how do I analyse any count-response GLM on this kind of survey data, every time.”
30.1 From spec sheet to skill
The spec sheet’s modelling section was specific to one prompt:
We will use generalized linear models… Topa abundance is probably over-dispersed, so we will need to use a negative binomial family… MASS::glm.nb(pres.topa ~ CB_cover*soft_cover, data = fish_coral_cover_sites)
A skill generalises the procedure, not the specific variable names:
Description: Use when fitting a GLM to a count response (abundance, counts, occurrences) that may be overdispersed.
Instructions: 1. Check the response variable’s mean vs. variance — if variance notably exceeds the mean, use MASS::glm.nb() instead of glm(family = poisson). 2. Start from a full model with all hypothesised predictors and interactions, then simplify using likelihood ratio tests (anova(m1, m2, test = "Chisq")), dropping non-significant interactions first. 3. After selecting a final model, check the dispersion parameter and residual diagnostics before reporting results — never report a model that hasn’t been checked. 4. Don’t library(MASS) globally; call MASS::glm.nb() directly to avoid namespace conflicts.
Notice this skill doesn’t mention pres.topa or CB_cover anywhere — that’s the difference between a spec sheet (a project’s context) and a skill (a repeatable procedure). You still attach the project-specific data and variable names each time; the skill just supplies the modelling judgment calls so you don’t have to restate them.
Save the skill above (or your own version) and run it against fish-coral-cover-sites.csv in two separate agent sessions, asking it to fit pres.topa ~ CB_cover. Compare the two runs — does the skill produce the same family and same simplification choices both times? Compare this to the same-prompt-twice test from Section 2 and Section 3 — did packaging the instructions as a skill reduce the variation you saw there?