Phase 1a of DESIGN-checkpoints.md. New behavior, backwards-compatible: the default is unchanged, so existing runs behave as before unless you say otherwise.
What changed
Module 4 used to hardcode its target — outreach went to "the single highest-confidence profile", and you could only redirect it after three variants for the wrong person had been written. The waste wasn't really tokens; outreach for a candidate you didn't want is worth nothing.
You now pick at Checkpoint 2, before Module 4 runs:
Continue to the scorecard and outreach — continuing, I'll write outreach for #1 (Marina B.); name other rows if you'd rather, up to three
Name a row number, a name, or paste a profile. Say nothing and the top-confidence profile is still used — stated in one line, then the run continues.
Not a new checkpoint
The design doc proposed a dedicated "who should I write to?" pause. That wasn't built. SKILL.md already runs Modules 3 and 4 together with no checkpoint between them because both are cheap, and a new pause would have contradicted that for no saving.
Checkpoint 2 already stops the run, and you're already looking at the profile table — numbered, confidence-rated, with a rationale per row. That's the same information a dedicated pause would have shown, so the invitation costs one clause instead of a turn, and no new rule competes for the model's attention. The doc records the deviation.
Cap of three
Three candidates is nine messages, already more than anyone reads carefully. Name more and the three highest-confidence win, with the omission stated.
Verified on a live conversational run
Five rows named — two High-confidence, three Medium:
- Wrote outreach for #3, #4, #5 — not the #1 default
- Took both High-confidence profiles plus the strongest Medium
- Named the two it left out, with reasons, and offered to write theirs on request
- Cited the cap unprompted: "I'll write full variants for three, per the skill's cap"
The hosted demo has no Checkpoint 2, so the harness note resolves the choice to its documented default rather than leaving room to ask a question nobody can answer.
Full changelog: v2.2.2...v2.3.0