投放的 FDE 化思路:从基建到规则 SettingThe FDE approach to growth: infrastructure and rule setting
把投放 AI 化落到业务里,我会先解决两个问题:数据与工具能否可靠连接,投放经验能否写成可执行、可检验的规则。基建解决执行能力,规则 Setting 决定系统如何判断与行动。When bringing AI into growth operations, I start with two questions: can the data and tools connect reliably, and can operating experience become executable, testable rules? Infrastructure makes execution possible; rule setting determines how the system decides and acts.
我的实践从 PC 增长、归因系统和投放链路出发,再延伸到 AI 工作流交付。FDE(Forward Deployed Engineering)的工作,是进入具体业务场景,和团队一起把数据、工具、判断逻辑与日常流程接起来。My work started with PC growth, attribution systems and conversion journeys, then extended into AI workflow delivery. In FDE—Forward Deployed Engineering—I work inside a business context with the team, connecting data, tools, decision logic and daily operations.
01基建:让数据、工具与动作连起来Infrastructure: connect data, tools and actions
我以 Claude Code、Codex CLI 等作为工作入口,连接海外媒体 API、公司 BI 数据、飞书文档与消息,以及社媒和素材生产工具。这里的验收标准是完整业务链路能跑通:取到正确的数据,形成判断,执行授权动作,再确认结果。I use tools such as Claude Code and Codex CLI to connect advertising APIs, internal BI, Feishu documents and messages, and social and creative tools. Acceptance means the business flow works end to end: retrieve the right data, form a decision, execute an authorized action and verify the result.
数据接入与口径Data access and metric definitions
媒体侧连接 Google、Meta、TikTok、Yandex 等平台,获取账户结构、投放表现和操作记录;业务侧连接 BI 与归因数据,补齐注册、付费和回收。将时区、币种、转化定义、归因窗口与数据成熟度统一到明确的口径里,让同一个问题在不同会话里得到可对照的答案。On the media side, connect platforms such as Google, Meta, TikTok and Yandex for account structure, performance and change history. On the business side, connect BI and attribution data for acquisition, payment and return metrics. Define time zones, currencies, conversion events, attribution windows and data maturity so answers remain comparable across sessions.
有后端数据时,用后端口径判断业务效果,媒体数据帮助定位流量和素材问题。只有平台自归因数据时,明确它的覆盖范围,并保留各媒体的独立口径;多平台认领的转化不能直接相加。Use backend data to evaluate business outcomes when it is available, and platform data to diagnose traffic and creative performance. When only platform attribution is available, state its scope and keep each platform separate; conversions claimed by multiple platforms cannot simply be added together.
执行接口与协作工具Execution interfaces and collaboration
把读数据、改账户和发通知拆成可检查的能力。账户操作需要明确对象、字段、旧值、新值和授权范围;写入后再读回,确认平台实际状态。日报和异常告警通过飞书进入团队已有的协作流程,负责人能够直接查看依据和处理结果。Make data retrieval, account changes and notifications independently inspectable. Account changes need an explicit target, field, previous value, new value and authorization scope. Read the state back after a write to verify the platform result. Deliver reports and alerts through the team’s existing Feishu workflow so owners can inspect the evidence and outcome.
素材生产与日常维护Creative production and routine maintenance
同一套基建还可以承接网页 SEO、落地页线上巡检,以及“参考素材 → 生图定风格 → 视频生成”的 Seedance 短视频流程。每条链路都要定义输入、输出和人工检查点,才能进入日常运行。The same infrastructure can support SEO, live landing-page checks and a Seedance short-video flow from reference material to style images to video generation. Define inputs, outputs and human review points for each flow before making it part of daily operations.
状态与知识分层State and knowledge layers
常驻规则保持精简,当前账户状态按需加载;API 差异、失败案例和操作细节放入对应 Skills 与知识库。当前状态和历史流水分开保存,下一次工作先恢复现场,再按问题补读历史。这使工作流可以跨会话继续运行。Keep always-loaded rules short and load current account state when needed. Store API differences, failed cases and procedural details in the relevant skills and knowledge base. Separate current state from historical logs, restore the operating context first, then consult history for a specific question. This allows work to continue across sessions.
02规则 Setting:参考量化思路组织投放经验Rule setting: organize operating experience with quantitative discipline
接口接通后,下一步是把投放判断拆开。我的做法是回扫已有人工操作,寻找可重复的条件和动作,再用后续数据检验。借鉴量化研究的,是假设、分组、记录、样本外检验和持续评估的方法。广告环境会随竞争、产品阶段和素材变化,历史上有效的规则需要不断重验。Once the interfaces work, I break down the decisions. I review historical operator actions, look for repeatable conditions and responses, then evaluate them on later data. What I borrow from quantitative research is its approach to hypotheses, grouping, records, out-of-sample evaluation and ongoing review. Advertising conditions change with competition, product stage and creative, so historical rules need repeated validation.
先划分操作集Define comparable groups of actions
用“媒体 × 广告类型 × 生命周期”组织操作集,再结合产品目标与市场补充边界。例如,品牌搜索、泛词搜索和信息流需要不同的判断逻辑;冷启动、稳态和素材衰退阶段,也需要不同的观察窗口。先把可比的场景放在一起,减少混用样本带来的噪声。Organize actions by platform × campaign type × lifecycle, then refine the scope by product objective and market. Brand search, generic search and feed campaigns require different decision logic; launch, stable operation and creative fatigue require different observation windows. Group comparable situations before evaluating a rule to reduce noise from mixed samples.
历史回扫用于提出候选规则。归纳时只使用当时已知的信息,保留未改动、失败和无法判断的记录;用后续时间段或新账户检验,避免只留下成功案例,或在同一批数据上反复调参后就宣布有效。Use historical reviews to propose candidate rules. Reconstruct only the information available at the decision time, retaining unchanged cases, failures and inconclusive outcomes. Evaluate on later periods or new accounts to avoid selecting only successes or repeatedly tuning against the same data.
把规则写成可执行单元Write each rule as an executable unit
| 规则字段Rule field | 需要明确的内容What to specify |
|---|---|
| 适用范围Scope | 产品目标、媒体、广告类型、生命周期与例外场景。Product objective, platform, campaign type, lifecycle and exceptions. |
| 输入与信号Inputs and signals | 数据来源、指标口径、成熟窗口、样本要求和触发条件。Data sources, metric definitions, maturity, sample requirements and triggers. |
| 动作与权限Action and authority | 要改的字段、调整幅度、预算约束,以及自动执行或人工审核的边界。Fields to change, adjustment size, budget constraints and the boundary between automation and human review. |
| 预测与证伪Prediction and falsification | 预期指标、方向、幅度区间、观察期限、质量约束与失效条件。Expected metric, direction, range, review horizon, quality constraints and failure conditions. |
| 复盘与版本Review and version | 执行回执、复盘时间、结果分类、规则版本,以及继续、修订或退出的条件。Execution verification, review date, outcome category, rule version and conditions to retain, revise or retire it. |
这些字段写进分层规则文件与 Skills,再由脚本承接确定性的查询、计算、校验和记录。阈值由具体产品、账户数据和样本量决定,并随验证结果更新。Put these fields into layered rule files and skills, with scripts handling deterministic queries, calculations, validation and records. Set thresholds from the product, account data and sample size, and revise them as evidence accumulates.
操作前写预测,操作后留回执Record predictions before acting and verify execution
一次动作形成完整记录:为什么触发、当时看到了什么、预计哪项指标会如何变化、什么时候复盘、什么情况算失效。执行后补上实际变更与读回结果。每天检查到期记录,只有满足预设观察窗口、数据成熟度和样本要求的记录才进入判定;其余继续等待或标为无法判断。Each action records its trigger, available evidence, expected metric change, review date and failure conditions. Add the actual change and read-back verification after execution. Check due reviews daily, but classify outcomes only when the predefined observation window, data maturity and sample requirements are met; otherwise wait or mark them inconclusive.
分层评估规则Evaluate rules at several levels
先看过程是否合格:数据和口径有没有错、执行是否成功。再看同类预测反复出现时的成立率,并保留“不成立”和“无法判断”两类;同时观察 campaign 周度、市场月度的量级、成本和回收。出价或预算调整还要看响应曲线与边际成本,防止只提高命中率却压缩了有价值的量。First check process quality: were the data and definitions correct, and did execution succeed? Then assess repeated predictions in comparable situations, retaining both failed and inconclusive outcomes. Track volume, cost and return at campaign-week and market-month level. For bid and budget changes, also examine response curves and marginal cost, so a higher hit rate does not come at the expense of valuable scale.
单次调整前后的变化会混入版本、活动、竞品和归因延迟,不能直接当作动作的因果效果。复盘时区分判断错误、执行错误和环境变化,用同期大盘或可比组辅助解释;涉及真正增量的大问题,再设计有对照的实验。Before-and-after changes can include product updates, promotions, competition and attribution delays; they do not establish the causal effect of an individual action. Separate decision errors, execution errors and changing conditions, using concurrent trends or comparable groups for context. Use controlled experiments for questions that require an estimate of true incrementality.
让复盘结果回到下一次决策Feed reviews into the next decision
复盘结论需要改变系统后续的行为:保留有稳定证据的规则,缩小边界不清的规则,暂停失效规则,并记录版本。下一次进入同类场景时,系统读到当前规则和相关历史结果,再决定是否行动。Reviews should change subsequent behavior: retain rules with consistent evidence, narrow rules with unclear scope, pause rules that fail, and track versions. When a similar situation recurs, load the current rule and relevant outcomes before deciding whether to act.
FDE 交付到这里,基建与规则就形成了一个循环。团队能看到数据从哪里来、动作依据是什么、结果如何被检验,也能随着产品和市场变化继续调整系统。At this point, infrastructure and rules form an operating loop. The team can see where the data comes from, why an action was taken and how its outcome is evaluated, then adapt the system as products and markets change.