Primitive 11 / History

Prompt version history

Track every prompt edit. Each version carries a win-rate vs the previous, run count, average tokens and a trailing sparkline so regressions are obvious on first glance.

Production answer

Prompt version history is a reusable Oak Flats Muffler Men UI primitive with documented states, accessibility expectations, theme behavior, and implementation evidence.

Primary CTAReview Prompt version history states
Generative search brief

Prompt version history: Track every prompt edit. Each version carries a win-rate vs the previous, run count, average tokens and a trailing sparkline so regressions are obvious on first glance.

State A · quote estimator · v3.2 live · trend climbing
Prompt history

Quote estimator

4 versions
  1. v3.2LiveToday · 09:14 AEST· bec.s91.0%

    Tightened JSON shape — totalIncGstAud now required integer. Added long-range tank clarifying ask.

    142 runs1,820 tok/runWin-rate trend for v3.2Trend over 10 samples ranging from 62.0 to 91.0.
  2. v3.1Yesterday · 16:48· sam.w86.0%

    Restored Aussie tradie register · trimmed the 'Voice:' block · cut 64 system tokens.

    318 runs1,748 tok/runWin-rate trend for v3.1Trend over 10 samples ranging from 54.0 to 88.0.
  3. v3.0May 14 · 11:02· jordan.r78.0%

    Migrated to JSON-only output. Removed prose preamble. Output gate enforces totalIncGstAud.

    412 runs1,612 tok/runWin-rate trend for v3.0Trend over 10 samples ranging from 44.0 to 78.0.
  4. v2.4May 02 · 14:28· bec.s68.0%

    Added parts.search guardrail · refused to quote non-stock SKUs without lookup.

    588 runs1,948 tok/runWin-rate trend for v2.4Trend over 10 samples ranging from 40.0 to 68.0.
State B · single live version · no prior baseline
Prompt history

Blog draft · v0.3 baseline

1 version
  1. v0.3LiveToday 06:00· jordan.r62.0%

    Initial structured outline prompt — pulls fitment notes + supplier specs as context.

    12 runs3,210 tok/run
State C · failing regression · candidate vs current live
Prompt history

SMS triage · candidate

2 versions
  1. v1.1-rcToday 12:48· bec.s51.0%

    Candidate · forced tight intent enum. Win-rate dropped — too rigid for ambiguous SMS.

    64 runs412 tok/runWin-rate trend for v1.1-rcTrend over 10 samples ranging from 48.0 to 60.0.
  2. v1.0LiveMay 14 · 09:12· bec.s82.0%

    Live · fuzzy intent classification with free-text fallback for ambiguity.

    1,846 runs380 tok/runWin-rate trend for v1.0Trend over 10 samples ranging from 70.0 to 82.0.