
多数人做副业的第一步就错了:把 AI 当成"提效工具",结果产出速度从 1 篇/天变成 20 篇/天,人反而成了整条链路的瓶颈——选题靠人、发布靠人、回复靠人、复盘靠人。副业能不能持续,取决的不是产能,而是"无人值守时长"。突击营的核心命题因此不是"怎么用 AI 写东西",而是"怎么把一条价值流做成可调度、可核算、可熔断的自动化系统"。
铁律 | 判据 | 反面案例 |
|---|---|---|
自动化率 | 单条价值流 ≥ 80% 环节可无人执行 | 需要真人出镜的口播号 |
边际成本 | 每新增一条交付,增量成本趋近于 token 费 | 定制化咨询 |
合规余量 | 平台规则、资质、版权三层无灰区 | 搬运剪辑、医疗建议 |
选品先过这三关,再谈技术。技术能放大流水线,但放大不了一条本身不成立的生意。
需求捕获(趋势/关键词) → 生产(多Agent协作) → 质检(规则+LLM评审) → 分发(限流/合规) → 数据回流
↑___________________ 反馈调优 ___________________|工程上三条硬约束:状态落盘(进程挂了能续跑)、幂等键(重跑不重复发布)、成本埋点(每条交付物的 token 与耗时都要记,否则算不清 ROI)。
# pipeline.py
import sqlite3, json, hashlib, time
from contextlib import contextmanager
DB = "side_hustle.db"
def init():
with sqlite3.connect(DB) as c:
c.execute("""CREATE TABLE IF NOT EXISTS jobs(
job_id TEXT PRIMARY KEY, stage TEXT, payload TEXT,
cost REAL DEFAULT 0, created REAL, updated REAL)""")
c.execute("""CREATE TABLE IF NOT EXISTS published(
idem_key TEXT PRIMARY KEY, job_id TEXT, ts REAL)""")
@contextmanager
def db():
conn = sqlite3.connect(DB); conn.row_factory = sqlite3.Row
try: yield conn; conn.commit()
finally: conn.close()
STAGES = ["topic", "draft", "review", "publish", "done"]
def idem_key(job_id, stage, content):
return hashlib.sha256(f"{job_id}:{stage}:{content}".encode()).hexdigest()[:24]
def load_or_create(job_id):
with db() as c:
row = c.execute("SELECT * FROM jobs WHERE job_id=?", (job_id,)).fetchone()
if row: return dict(row)
rec = dict(job_id=job_id, stage="topic", payload="{}",
cost=0.0, created=time.time(), updated=time.time())
c.execute("INSERT INTO jobs VALUES(:job_id,:stage,:payload,:cost,:created,:updated)", rec)
return rec
def advance(job_id, stage, payload, cost=0.0):
with db() as c:
c.execute("UPDATE jobs SET stage=?,payload=?,cost=cost+?,updated=? WHERE job_id=?",
(stage, json.dumps(payload, ensure_ascii=False), cost, time.time(), job_id))# agents.py
import os, json
from openai import OpenAI
llm = OpenAI(api_key=os.getenv("DEEPSEEK_API_KEY"),
base_url="https://api.deepseek.com")
def call(system, user, model="deepseek-chat", temp=0.7):
r = llm.chat.completions.create(
model=model, temperature=temp,
messages=[{"role": "system", "content": system},
{"role": "user", "content": user}])
return r.choices[0].message.content, r.usage.total_tokens
def agent_topic(niche: str, history: list[str]) -> dict:
sys = "你是内容选题分析师。输出 JSON:{title, angle, keywords[3], est_demand(1-10)}。只输出 JSON。"
usr = f"赛道:{niche}\n近期已发选题(避免重复):{history[-20:]}"
raw, tok = call(sys, usr, temp=0.8)
return {"data": json.loads(raw.strip("```json \n")), "tokens": tok}
def agent_draft(topic: dict, style_ref: str) -> dict:
sys = ("你是资深内容作者。按给定风格写正文,结构:钩子→3个信息增量点→可执行清单。"
"禁止编造数据与政策,不确定处标注[待核实]。")
usr = f"选题:{topic['title']}\n角度:{topic['angle']}\n风格样例:{style_ref}"
text, tok = call(sys, usr)
return {"text": text, "tokens": tok}# qa.py
import re, json
BANNED = [r"保证(收益|回本)", r"(医疗|投资)建议", r"百分百", r"最(低|高)价"]
REQUIRED = {"min_len": 600, "max_len": 3000}
def rule_gate(text: str) -> list[str]:
errs = []
for p in BANNED:
if re.search(p, text): errs.append(f"违禁表述:{p}")
if not REQUIRED["min_len"] <= len(text) <= REQUIRED["max_len"]:
errs.append("长度越界")
if text.count("[待核实]") > 2: errs.append("待核实项过多")
return errs
def llm_review(text: str) -> dict:
sys = ("你是内容质检员。对文本按 信息密度/可执行性/合规风险 三项各打1-10分,"
"输出 JSON:{density, actionable, risk, reason}。只输出 JSON。")
raw, tok = call(sys, text, temp=0.1)
s = json.loads(raw.strip("```json \n"))
s["pass"] = s["density"] >= 7 and s["actionable"] >= 7 and s["risk"] >= 8
return {"score": s, "tokens": tok}# roi.py
PRICE_PER_1K = 0.001 # 按实际模型单价配置,单位:元
DAILY_BUDGET = 5.0 # 日预算上限,超支即熔断
def job_cost(tokens: int) -> float:
return tokens / 1000 * PRICE_PER_1K
class Budget:
def __init__(self, limit=DAILY_BUDGET):
self.spent, self.limit, self.day = 0.0, limit, time.strftime("%Y-%m-%d")
def charge(self, tokens: int) -> float:
today = time.strftime("%Y-%m-%d")
if today != self.day: self.spent, self.day = 0.0, today
c = job_cost(tokens); self.spent += c
if self.spent > self.limit:
raise RuntimeError(f"日预算熔断:已花 {self.spent:.2f} 元")
return c
def unit_economics(job: dict, revenue: float) -> dict:
return {"成本": round(job["cost"], 4), "收入": revenue,
"毛利": round(revenue - job["cost"], 4),
"ROI": round(revenue / max(job["cost"], 1e-6), 2)}# publish.py
import hashlib
def should_human_review(score: dict) -> bool:
return score["risk"] < 9 or score["density"] < 8 # 边缘样本走人工
def publish(job_id: str, text: str, channel: str) -> str:
key = idem_key(job_id, "publish", text + channel)
with db() as c:
if c.execute("SELECT 1 FROM published WHERE idem_key=?", (key,)).fetchone():
return "skipped_duplicate"
# 此处接各平台 API / 半自动导出
_dispatch(channel, text)
c.execute("INSERT INTO published VALUES(?,?,?)", (key, job_id, time.time()))
return "published"def run_one(niche: str, budget: Budget):
job_id = hashlib.md5(f"{niche}{time.time()}".encode()).hexdigest()[:12]
job = load_or_create(job_id)
if job["stage"] == "topic":
t = agent_topic(niche, history=[]); budget.charge(t["tokens"])
advance(job_id, "draft", t["data"], job_cost(t["tokens"]))
job = load_or_create(job_id)
if job["stage"] == "draft":
d = agent_draft(json.loads(job["payload"]), style_ref="…")
budget.charge(d["tokens"])
advance(job_id, "review", {"text": d["text"]}, job_cost(d["tokens"]))
job = load_or_create(job_id)
if job["stage"] == "review":
text = json.loads(job["payload"])["text"]
errs, rv = rule_gate(text), llm_review(text); budget.charge(rv["tokens"])
if errs or not rv["score"]["pass"]:
advance(job_id, "done", {"rejected": errs or rv["score"]})
elif should_human_review(rv["score"]):
advance(job_id, "done", {"pending_human": rv["score"]})
else:
advance(job_id, "publish", {"text": text}, job_cost(rv["tokens"]))
job = load_or_create(job_id)
if job["stage"] == "publish":
publish(job_id, json.loads(job["payload"])["text"], "wechat")
advance(job_id, "done", {})
return load_or_create(job_id)环节 | 技术选型 | 副业视角的关键点 |
|---|---|---|
选题 | 趋势 API + 向量检索去重 | 去重比生成更值钱 |
生产 | DeepSeek / 多模态模型 | 长文用 chat,配图用图像模型 |
质检 | 规则正则 + LLM-as-judge | 规则挡硬伤,模型挡软伤 |
分发 | 平台 API / RSS / 半自动 | 先做半自动,别碰风控灰区 |
回流 | SQLite + 简易看板 | 记录"哪类选题 ROI 高" |
运维 | 定时任务 + 预算熔断 | 没有熔断的自动化是负债 |
第 1 周:跑通全链路,人工全检,目标是发现规则的漏洞而非追求产量。 第 2 周:积累 30 条样本,校准 LLM 评审阈值(阈值靠样本定,不靠感觉)。 第 3 周:接入预算熔断与幂等发布,把"无人值守"从 2 小时拉到 24 小时。 第 4 周:按 ROI 排序选题类型,砍掉毛利为负的品类。
自动化不等于免责。三条红线:平台规则优先于技术可行性(批量发布触发风控的账号价值归零);合规审查不能全托给模型(医疗、金融、法律类内容必须人工终审);成本熔断必须硬编码(一个死循环能在一夜之间烧掉一个月利润)。
副业做成的标志不是"AI 帮我写了很多",而是你关掉电脑一周,流水线仍在产出、成本仍在预算内、质量仍在阈值上。突击营练的,就是把这个状态调出来。
原创声明:本文系作者授权腾讯云开发者社区发表,未经许可,不得转载。
如有侵权,请联系 cloudcommunity@tencent.com 删除。