Feedback: show skills and no-skills together, and link where to get the skills

[leggetter]

From Gareth, on the first draft:

showing skill/no-skill figures together rather than a toggle (to show the advantage
of using the skills and most folks won't toggle) - worth including a link to where
folks can get the skills, too.

The toggle argument is right. The page has a Benchmark / Without skills segmented
control, and a reader who never touches it sees one arm and no comparison. The skills
delta is a stated goal of the project and it is currently behind an interaction most
people will not perform.

The link is obviously right and cheap: nothing on the page says where to get the
skills. npx skills add hookdeck/agent-skills, or the repository.

One caution on the framing. The suggestion is to show the comparison "to show the
advantage of using the skills". The current measurement does not support that framing:

Pair	+skills	no skills	delta
Claude Sonnet 5	15/15	14/15	+1
Codex GPT-5.6	12/15	12/15	0
Codex GPT-5.4-mini	10/15	13/15	-3

And after Loop 1, even those numbers are suspect: they come from single-attempt runs,
and #14 shows single attempts in this suite are unreliable.

So showing both arms together is the right design and would currently show a mixed
result, including one negative. That is fine, and arguably better than a toggle that
hides it, but it should be a deliberate choice rather than a surprise: the page would
be publishing "our skills made this model worse" in its default view.

Depends on #14. Worth designing now, worth publishing once the numbers are
trustworthy.

Done when both arms are visible without interaction, and the page says where to
get the skills.

[leggetter]

Decided: both suites side by side, no toggle, fractions kept.

Shape, confirmed against the real data:

                Claude Code            Codex              Codex
                 Sonnet 5             GPT-5.6          GPT-5.4-mini
            without   with       without   with     without   with
  Build      10/11    11/11        8/11    10/11     9/11     6/11

Two notes on implementation and on what it will show.

It is smaller than it looks. groupColumnsByFamily in src/lib/evals/scoreboard.ts already builds a two-tier header, family over its columns. The table renders one column per family today only because EvalsApp filters rows to a single experimentSuite before passing them in. Removing that filter and dropping the segmented control gets the layout above; the header, the weaker-model marker and the column widths already handle multiple columns per family.

What it will publish. The current deltas are +1, 0 and -3, so most cells will show the same value twice and one model will visibly score worse with our skills. Gareth's framing was "to show the advantage of using the skills"; the honest version of this design shows the advantage where there is one and the cost where there is not. That is the right call for a page whose argument is that its numbers are checkable, but it should be a decision rather than a surprise on launch day.

Still depends on #14: both columns come from single-attempt runs, and Loop 1 showed those are unreliable in this suite.

Also still open from this issue: nothing on the page says where to get the skills. npx skills add hookdeck/agent-skills.