Matt Pocuck had an interesting suggestion that I believe might be inspired from an episode of the office.
The whole idea is that end a Claude Code prompt with this:
Be extremely concise. Sacrifice grammar for the sake of concision.
So I couldn't help but try some prompts with this idea. Here are some results.
Normal prompt
explain huberregression from scikit-learn.
Huber regression in scikit-learn is a **robust linear regression** method that behaves like ordinary least squares for small residuals but becomes linear (less sensitive) for large residuals. It uses the Huber loss:
- If |r| ≤ ε: loss = 0.5 * r² (quadratic, like least squares)
- If |r| > ε: loss = ε * (|r| − 0.5 * ε) (linear, like absolute loss)
Here `r = y − ŷ` and ε is the threshold that controls when the loss switches from quadratic to linear.
Why it’s useful:
- Reduces the influence of outliers compared to plain least squares.
- Still differentiable everywhere, which makes optimization stable.
Key points in scikit-learn:
- Class: `sklearn.linear_model.HuberRegressor`
- Important params: `epsilon` (switch point), `alpha` (L2 regularization), `max_iter`, `tol`
- It **scales features internally**; coefficients are returned in original scale.
- If `epsilon` is large, it behaves more like standard regression; smaller values increase robustness but can down-weight legitimate high-residual points.
When to use:
- You suspect outliers or heavy-tailed noise.
- You want a linear model that’s more robust than `LinearRegression` but smoother than `RANSAC`/`TheilSen`.
Prompt + consise
explain huberregression from scikit-learn. Be extremely concise.
Huber regression (sklearn.linear_model.HuberRegressor) is a robust
linear model: it uses squared loss for small residuals and linear
loss for large ones, reducing outlier influence while staying close
to least squares.
Prompt + grammar consise
explain huberregression from scikit-learn. Be extremely concise. Sacrifice grammar for the sake of concision.
HuberRegression (sklearn): linear reg w/ Huber loss; quadratic for
small residuals, linear for large → robust to outliers. Tuned by
`epsilon`; solves via iterative reweighted least squares; supports
`alpha` (L2) + `fit_intercept`, `max_iter`, `tol`.
Notice it?
Is it me, or does it feel like the variant with less grammer gives me many extra details that the normal "be consise" prompt doesn't give? Feels like dropping grammar might indeed work as a forcing mechanism.