Settings → Integrations had three- to four-fold redundant status displays per card. The new pill is the one place a connection status appears: top-right, Dracula-tokenized chip, "Connected · 12 min ago" relative-time pattern. Locale-aware (EN + DE), mobile-safe (whitespace- nowrap, abbreviated "min" / "h" / "d" forms so it fits a 360px card). Reusable for v1.4.20's Apple Health card on iOS. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
3.5 KiB
Phase A4 — AI prompt: no default-positivity opener (v1.4.19)
Completed 2026-05-10 ~12:59 CEST.
Marc's mandate
"I read AI insights and notice the SAME phrase opens almost every response: 'Datengrundlage ist sehr stark / Your data foundation is strong'. The user doesn't even know what data is shared, so this opener is awkward. Mention data quality only when actually problematic."
What landed
One commit on origin/main:
| Commit | Subject |
|---|---|
b5e9a95 |
fix(ai): no default-positivity opener about data quality (mention only when limited) |
Prompt change — src/lib/ai/prompts/insight-generator.ts
Added GROUND RULE 7 to both EN and DE prompts. The rule:
- Forbids opening with a compliment about data quantity / quality.
- Defines the only conditions under which to mention data quality:
n<7readings in the analyzed window,recencyDays>14since the last entry, or a coverage gap that biases the comparison. - When data is fine: dive straight into the analysis.
- Lists banned-opener phrases verbatim in both locales so the model cannot paraphrase the same sentiment ("Your data foundation is strong", "Datengrundlage ist sehr stark", "You have a solid baseline", "Großartiger Datensatz", "Du hast eine solide Baseline", "Great dataset").
Version bump
PROMPT_VERSION 4.16.1 → 4.19.0. Feedback aggregation
(feedback-attribution.ts reads the constant) can now distinguish
responses generated under the new rule from older payloads.
Tests
New file src/lib/ai/__tests__/no-default-positivity-opener.test.ts
pins the rule, the trigger thresholds, the banned phrases, and the
dive-straight-in instruction in both locales (9 tests, all green).
Existing PROMPT_VERSION assertions in
medical-reference-prompt.test.ts relaxed from /4\.16\.\d+/ to
/4\.\d+\.\d+/ so future minor bumps don't sweep up these checks.
Verification
npx vitest run src/lib/ai/__tests__/insight-generator-prompt.test.ts src/lib/ai/__tests__/medical-reference-prompt.test.ts src/lib/ai/__tests__/no-default-positivity-opener.test.ts— 38/38 pass (was 28; +9 new + 1 reused after relaxation).pnpm lint— 0 errors / 12 pre-existing warnings.pnpm typecheck— only pre-existing error from A5/A6's untrackedintegration-status-pill.test.tsx. My code is clean.- Full
pnpm test— 6 pre-existing failures in A1'sbp-in-target, A3'sinsights-polish, and A5/A6'si18n-locale-integrity/integrationPillkeys. None mine.
Smoke-test plan (deferred to post-deploy)
Once v1.4.19 ships, generate a fresh insight against Marc's data and
confirm the first sentence of summary does NOT mention
"Datengrundlage" / "data foundation" unless n<7, recencyDays>14,
or a coverage gap is present. Cached payloads from older
PROMPT_VERSION will still carry the old opener — promptVersion on
the cached row distinguishes them.
Cross-agent race
src/app/page.tsx (A3 territory) was auto-bundled by the
pre-commit hook at the moment of git commit — same shared-cwd
race documented across C1, B5a, B5b, B6 of prior marathons. My
intended scope is the three files in src/lib/ai/. Verified
post-commit on origin/main: the AI prompt + tests are correct;
page.tsx carries A3's edits, which are A3's to defend. v1.4.20
worktree adoption (deferred from v1.4.15 → v1.4.16 → v1.4.18) is
still overdue.