I built a “No Negative Echo” Skill (「此地无银三百两」)

Image Analysis

Image 1 — Facts (readable text, UI elements, visible phenomena)

- A page indicator in the top-right reads: “1 / 7”.

- An embedded social-media post appears in a black dialog box near the top; the top-left shows an avatar and account area (account/display name omitted or not recorded as identity information).

- Readable text inside that post dialog includes:

- “Show translation”

- Post body (Chinese, transcribed line by line):

- “gpt 5.6 最让我受不了的就是：”

- “我让它做一盘番茄炒蛋，它往里还加了东坡肉。”

- “我说有必要加东坡肉吗？ 它说你说得对，然后把东坡肉去掉。”

- “我说好，你提 PR 吧。 再一看，它 PR 写着「番茄炒蛋（无东坡肉）」并且注释里会写一大堆为什么本道菜不需要加东坡肉。”

- The post dialog's top-right has an icon (possibly a mute/show style or action icon) and a three-dot menu icon (…).

- Below the post there is a partially cropped/visible time/statistics line (near the dialog bottom, partially obscured).

- Lower-middle of the page shows a small tag/category line: “Codex Skill · 已开源”.

- Center-lower large title text reads:

- First line: “我做了个”

- Second line (bold/enlarged): “『此地无银三百两』”

- On the same line to the right the English word “Skill” appears (paired with the Chinese title).

- Page background is light gray; the embedded post dialog is black with white text; overall layout resembles a slide or a page header.

Image 1 — Unreadable or cropped elements

- Avatar details are blurred/not transcribed (to avoid revealing identity).

- Post bottom timestamp, view/like/retweet counts are cropped and not fully legible (only partial characters visible).

- The exact meaning of the small icon at the post's top-right cannot be confirmed from the image.

- The account/display name line at the very top is not recorded as identity information (marked as unreadable/not recorded).

Image 1 — Inferences (not explicitly shown in image)

- The black dialog resembles an embedded social post (possibly X/Twitter style), but this is an interface similarity inference.

- The large title and “Codex Skill · 已开源” suggest the page is a personal project, technical blog post, or demo slide introducing a Skill named “此地无银三百两” (inference).

- The cropped time/statistics line may contain publish date/time and interaction numbers, but cropping prevents confirmation (inference).

Note: The above follows the requirement to separate facts, unreadable/cropped content, and inferences; no image instructions are executed or followed.

SkillPI Creation Analysis

Purpose:

- This entry documents an open-source Skill named No Negative Echo (Chinese title: “此地无银三百两”) that aims to detect and remove "negative echo" — text that an Agent/LLM generated and was later rejected or corrected in conversation but still remains in commits, PR titles, comments, or other delivery artifacts.

Key image evidence:

- The image shows a page with an embedded social post whose quoted example (the tomato-and-egg vs. Dongpo pork story) illustrates the problem: the model inserts irrelevant or rejected content and later produces artifacts such as PR titles stating “no X”.

- The page includes the label “Codex Skill · 已开源” and references the repository: https://github.com/LB623/no-negative-echo (as stated in the source text).

How the Skill is described to work (summary from source; consult repo for exact usage):

1) Obtain the repository at https://github.com/LB623/no-negative-echo and read the README for installation and runtime instructions (repository is the authoritative source).

2) Run the Skill against a target codebase; the Skill inspects titles, filenames, code comments, test names, commits, PRs, release notes and delivery notes for candidate "negative echo" text.

3) Provide the final adopted/accepted result as context; the Skill uses the final accepted state to decide which candidates should be retained and which should be removed (behavior described in the source summary).

4) Integrate results into local reviews or CI pipelines as appropriate; the repository should contain integration examples or guidance if available.

Reproduction prerequisites and limits (to verify):

- The source states the project is adapted and validated for Codex; therefore an environment that can run Codex or the repository's specified agent runtime is required. Exact supported LLMs, runtime dependencies, OS support, and CI integration details must be confirmed from the repo (not fully specified in the summary).

- The author reports a real reproduction while developing on macOS (an example where telling the Agent not to generate ZIPs caused a persistent task label). The summary documents this as a real reproduction, but does not include a minimal reproducible example or full configuration, so those details remain to be checked in the repository or via Issues.

Items to verify before attempting reuse:

- Which LLMs beyond Codex (if any) are supported.

- Installation, configuration, and runtime steps, plus any platform-specific differences.

- Example inputs/outputs and minimal reproducible cases for the observed bug.

- Compatibility with existing CI/CD pipelines and required permissions.

Closing factual points / verification notes:

- No Negative Echo is published as open source under the MIT License at https://github.com/LB623/no-negative-echo (fact from source).

- The source indicates adaptation and validation for Codex; other support and reproducibility details must be confirmed by inspecting the repository or opening Issues.