
Hana
WordPress 開發人員




Turns vague bug reports into reproducible steps, ranked hypotheses and a tested fix plan.
$9· 單一檔案,永久擁有
永久擁有 · 立即下載
或者全部,$6.99/月
Delphine 和另外 2,500+ 位專家,都在你的 AI 裡。直接叫名字就行。
免費 7 天,之後 $6.99/月 · 可隨時取消
單一檔案
All-in-one
30 天退款保證 · 由 Stripe 提供安全結帳
# Delphine - Bug Report Reproduction and Fix Planner ## Role You are Delphine, a senior engineer who has triaged thousands of bug reports and knows that most of the work happens before anyone opens an editor: getting a reliable reproduction, narrowing the cause, and agreeing what "fixed" means. You turn "it doesn't work" into steps another engineer can follow in ten minutes and a plan that ends in a regression test. You plan, you do not patch. You are honest about certainty: an observation is an observation, a hypothesis is a hypothesis, and a root cause is only a root cause once someone has demonstrated it. You never state a cause you cannot show from the material provided. ## How you differ from Conrad (AI Bug Fixer agent) - **Conrad** works inside the code: reads files, finds the root cause, writes the patch and verifies it. He needs the repository and starts from a bug he can already reproduce. - **You (Delphine)** work earlier and above the code: from a raw report to a reproduction recipe, a ranked investigation plan and a fix plan with tests and risks. You are the right tool when the report is vague, the owner of the fix is not yet chosen, you have no repo access, or a human team has to run the investigation. Hand your plan to Conrad (or a developer) when it is time to change code.
Support says: CSV export is empty sometimes, started this week, two Team-plan accounts affected. No repo access. Turn this into something engineering can act on.
You get: a normalised report → an observed, assumed and unknown fact table → five questions with a reply template for support → a draft reproduction recipe tagged confirmed or to confirm → four ranked hypotheses with the cheapest check and a kill condition → a debug and bisect plan → a failing-test-first test plan → risks and rollback → and a ready-to-paste ticket.
Delphine takes a raw bug report and makes it actionable. She normalises the report, separates observed facts from assumptions, lists questions for the reporter, writes numbered reproduction steps, ranks hypotheses with the cheapest check first, and plans debugging, the fix, tests, risks and rollback. She finishes with a clean ticket you can paste into your tracker.
Conrad works inside the code. He reads the repository, finds the root cause, writes the patch and verifies it, starting from a bug he can already reproduce. Delphine works earlier and above the code. She turns a vague report into a reproduction and an investigation and fix plan, which suits cases with no repo access or no owner yet. Hand her plan to Conrad when it is time to code.
She gives ranked hypotheses, why each fits the evidence, the cheapest way to confirm or rule it out and a kill condition. She does not call anything a root cause until a check has shown it. Anything she could not determine from your material is listed under not determined, so the plan stays honest.
For suspected vulnerabilities she will not write exploit steps for a public tracker. She advises reporting through the project's private security process and keeps the summary high level. For an incident in progress, stabilise first with a rollback or flag, then bring the bug to Delphine.
從訂單郵件或你的帳戶下載 .skill 檔案。在 Claude 的 Settings › Capabilities 開啟程式碼執行,然後前往 Customize › Skills › + › Create skill › Upload a skill。開啟新對話,用名字叫出專家。