Skip to main content

RC1 02 Native Evaluation Image Implementation Plan

For agentic workers: REQUIRED SUB-SKILL: Use superpowers:subagent-driven-development (recommended) or superpowers:executing-plans to implement this plan task-by-task. Steps use checkbox (- [ ]) syntax for tracking.
Goal: Replace img/rc1-highlights-v2/02-native-eval.png with a pure-white 3:4 product infographic that combines three-generation reliability, Native-vs-Classic Token and duration reductions, and RC1-vs-Beta16 duration improvement. Architecture: Generate one raster candidate from the approved copy and layout, then copy it into the existing stable asset path. Normalize neutral near-white background pixels to exact #FFFFFF without changing colored content, and verify dimensions, corners, text, and Git scope before committing and pushing. Tech Stack: Built-in OpenAI image generation, PNG, PowerShell, .NET System.Drawing, Git.

Global Constraints

  • Output path is exactly img/rc1-highlights-v2/02-native-eval.png.
  • Canvas ratio is exactly 3:4 portrait; no non-uniform scaling or vertically stretched typography.
  • Background and all four corners are exact #FFFFFF.
  • Reuse the repository Comet logo without changing its aspect ratio.
  • Copy is declarative and professional; do not use rhetorical questions or “不是……而是……” constructions.
  • Reliability values are Classic 89.58% / 75%, Beta16 95.83% / 87.5%, and RC1 100% / 100% for strict pass@1 / pass³.
  • Efficiency values are Native vs Classic: 总 Token −76.8%, 模型耗时 −47.4% and RC1 vs Beta16: 平均模型耗时 −22.3%.
  • Footer is 16 个真实任务 · 每个任务运行 3 次.
  • Do not claim RC1 reduces Token versus Beta16.

Task 1: Generate and select the infographic

Files:
  • Read: logo/favicon.png
  • Read: img/rc1-highlights-v2/02-native-eval.png
  • Modify: img/rc1-highlights-v2/02-native-eval.png
Interfaces:
  • Consumes: approved copy and layout from docs/superpowers/specs/2026-08-31-rc1-02-native-eval-design.md.
  • Produces: a single 3:4 PNG at img/rc1-highlights-v2/02-native-eval.png.
  • Step 1: Inspect the current target and logo
Use the image viewer for both files. Confirm the current target is the edit subject and logo/favicon.png is the brand reference.
  • Step 2: Generate one infographic candidate
Use the built-in image generation tool with this exact production prompt:
  • Step 3: Copy the selected generated PNG into the target path
Run:
  • Step 4: Inspect the workspace copy
Use the image viewer on img/rc1-highlights-v2/02-native-eval.png. Reject and regenerate if any required string is wrong, typography is vertically stretched, the composition is crowded, or the generation contains an unapproved claim.

Task 2: Normalize, verify, commit, and push

Files:
  • Modify: img/rc1-highlights-v2/02-native-eval.png
  • Verify: docs/superpowers/specs/2026-08-31-rc1-02-native-eval-design.md
Interfaces:
  • Consumes: selected Task 1 PNG.
  • Produces: verified asset commit pushed to origin/rc1.
  • Step 1: Normalize neutral near-white background pixels
Run:
  • Step 2: Verify geometry and border pixels
Run:
Expected: Ratio is 0.75 and Corners is #FFFFFF, #FFFFFF, #FFFFFF, #FFFFFF.
  • Step 3: Perform final visual inspection
Use the image viewer at original detail. Verify all approved text, normal typography proportions, pure-white background, correct hierarchy, and readable comparison labels.
  • Step 4: Verify Git scope
Run git status --short, git diff --check, and git diff --stat. Expected: the implementation diff contains only img/rc1-highlights-v2/02-native-eval.png; the already committed spec and plan do not appear as unstaged changes.
  • Step 5: Commit the image
Run git add -- img/rc1-highlights-v2/02-native-eval.png and git commit -m "fix: expand RC1 native evaluation comparison".
  • Step 6: Push and verify the remote head
Run git push origin rc1, git fetch origin rc1, git rev-parse HEAD, git rev-parse origin/rc1, and git status --short --branch. Expected: local HEAD equals origin/rc1, and the worktree is clean.
最后修改于 2026年8月31日