Vibecoding Resume Story

Vibecoding means the division of labor is: the user supplies requirements, prompts, rules, and acceptance decisions; the AI tool writes the code. A resume entry that implies "I wrote the code" collapses under two follow-up questions. One that presents "I designed and verified everything, the tool implemented it" is both honest and a stronger signal for AI-era roles.

Non-negotiable Rules

1. Every claim must reduce to a requirement the user actually issued. Write each bullet as something the user said or decided, never as implementation work. "Defined the rule: at most 5 tasks per day, each with priority and estimated minutes" is defensible; "implemented a validation engine" is not.
2. State the division of labor openly. Name the tool (e.g. WorkBuddy) and the word vibecoding. Hiding it looks guilty; stating it reads as AI-toolchain fluency, which is the actual job skill.
3. No implementation jargon the user cannot explain in plain words. If the user cannot answer "what does L1/L2/LLM-as-Judge mean" in one sentence, the term does not go on the resume. Translate everything into requirement language.
4. Metrics only from real runs, at real granularity. One live test scoring 8/10 is written as "one end-to-end live run, scored 8/10" — never "stable 8-9 points". Inflated metrics are the first thing that gets probed.
5. Pain point comes from the user's own experience. "I hit this problem myself while doing X" beats "the whole industry has this problem" — it is unverifiable in the good way and shows user empathy.
6. Acceptance is the user's job title. The repeated verbs are: defined, required, directed, verified on a real device, rejected, iterated. Never: coded, implemented, developed the algorithm.

Entry Structure (4-5 bullets, each at most 2 lines)

1. Background & goal: the user's own pain point in one line; division of labor in one line; shipped state (open-sourced, deployed, usable without a key).
2. Rules & constraints: the hard requirements issued as prompts, in plain numbers (counts, ranges, budgets).
3. Evaluation & verification: how output quality is checked (rule pass then model score, threshold, bounded retry, audit log); the deliberately-bad test path used to prove the check works; the live-run evidence with its real single-run score.
4. Iterations & delivery: the timeline as rounds of "requirement → verify → next requirement" (e.g. MVP overnight → live API → mobile polish → desktop → open-source launch), with the acceptance rule "rejected until it passed on a real device".
5. Optional — reduced decision cost: templates, one-tap adopt, one-tap export; the user's original "reduce decision cost" framing.

Interview Prep Table (always deliver with the resume text)

For every bullet, provide the plain-language sentence the user can say out loud. Example mapping:

Resume says	User says in interview
"daily tasks capped at 3-5, each with priority and estimated minutes, total within the day's budget"	"I told it: max five things a day, each needs a priority and rough minutes, and the total can't exceed the time I actually have"
"score below 7 retries, at most twice"	"first the hard rules block it, then the model grades it; under 7 it regenerates, twice at most, otherwise it keeps the best attempt"
"deliberately-bad sample path to verify interception"	"I made it generate a bad plan on purpose the first time, to watch the gate actually catch it — that's how I trusted the mechanism"

Red Flags To Scan For Before Sending

• Mixed-language accidents ("runtime 长问题") or terms the user did not write themselves.
• Contradictions between bullets (e.g. "independently completed" next to "AI-assisted").
• The same fact told twice in different bullets (assign each fact one home: mechanism in the rules bullet, evidence in the verification bullet, timeline in the iteration bullet).
• Any number that cannot be produced on demand (scores, counts, timings).
• Bolder claims in the resume than in the project's public README — interviewers do compare.

Example

Bad, because it claims implementation work:

实现了电商脚本生成引擎，包含合规校验模块。

Good, because every clause reduces to something the user actually decided:

定义电商短视频脚本规则：单条不超过 60 秒，卖点不超过 3 个，合规词先过再写； 用 AI 工具实现后，我逐条验收输出并沉淀 27 类违禁表述词库，过审率从 70% 提到 80%。