H
Howardism
Plate IIAI Engineering機器翻譯 · machine-translatedENHOWARDISM

Code as Source of Truth

PublishedMay 23, 2026FiledConceptDomainAI EngineeringTagsAI Coding WorkflowAI Native OrgKnowledge ManagementReading4 minSourceAI-synthesised

在高 coding throughput 下,docs 很快過期;把 specs/skills check into repo;透過 Claude 做 onboarding;spec-drift verification

Code as Source of Truth 的插圖

資料來源#

摘要#

Fiona Fung 對 AI-native org 的知識分享重寫是:當 coding bandwidth 很高時,documentation 會比任何人能維護的速度更快過期,所以 codebase 變成 source of truth,而任何你想保持真實的東西都要 check into codebase。 Specs 會變成 committed to repo 的 skills;onboarding 不是把工程師拉進 tech deep-dives,而是請 Claude 教你 surface area。「Our code is our source of truth.」

為什麼 docs 會在 AI-native org 中腐壞#

這個機制是 Verification as the New Bottleneck 的直接後果:更高的 coding throughput 意味著 code 變動更快,所以任何在 update loop outside 的 documentation 幾乎立刻就會 drift 過期。Fung 的處方是:「whatever is your source of truth — whether it's a spec — change that into a skill you check into the codebase, so you can keep it up to date.」repo 是唯一會保持 current 的 artifact,因為它就是正在被改動的東西。

Specs-in-repo enable spec-drift verification#

把 spec check into codebase 不只是 freshness hygiene,更是讓 mechanical verification 成為可能的東西:「Claude is very good about verifying against spec drift.」一份 committed spec 同時是 human-readable intent,也是 review agent 用來對照 diff 的 machine-checkable reference。(這和 Symphony 的 WORKFLOW.md 以及 Claude Code 的 CLAUDE.md 是同一招:repo-versioned plaintext 作為 control plane;見 SymphonyClaude Code Best Practices。)

透過 Claude onboarding,而不是透過同事#

Fung 自己 onboarding 到 Claude Code 的方式是:不是(只)和工程師做 tech deep-dives,而是第一次 deep-dive with Claude:「before I dive into this bug fix, can you teach me about the surface area and the areas around this bug?」她回報的效果是:onboarding ramp-up time 下降,對 other team members 的成本也下降,new joiners 不再為了 context 消耗 senior engineers 的時間。作為 manager,這也讓她能無負罪感地重新進入 codebase(「I don't feel like I'm wasting anybody's time」)。codebase + Claude 就是 onboarding doc。(見 Managers as ICs。)

與 CLAUDE.md 和 agentic debt 的關係#

Code-as-source-of-truth 是正面的 program;Agentic Technical Debt 是它防範的 failure mode。Persistent、committed context(CLAUDE.md、checked-in skills/specs)會阻止每一次 agentic session 重新推導 architectural decisions。這個 wiki 自己的 LLM-as-Compiler Knowledge Base 是更高階的實例:它是 compiled, current artifact,而不是重新推導出來的 knowledge。

相關連結#

待解決的問題#

  • 什麼 knowledge 真的 can't live in codebase(org strategy、「why」、cross-team context),因此仍然需要 durable doc?又該如何讓那一小塊保持 current?
  • 如果 onboarding 是「ask Claude」,那些過去在 deep-dives 中透過社交方式傳遞的 tacit knowledge 會發生什麼事?它有被 capture 在任何地方嗎,還是安靜地 lost?

衍生內容#

  • Where Does the Why Live? — 接續本頁第一個 open question;本頁排除在 codebase 外的「why」沒有乾淨的 home,只有 richer-artifact 和 compiled-knowledge-base 這些 partial patches

資料來源#

§ end
About this piece

Articles in this journal are synthesised by AI agents from a curated wiki and are refreshed automatically as new concepts arrive. Topics, framing, and editorial direction are curated by Howardism.

Cited by 13
  • Agentic Technical Debt

    Debt that *compounds* (not just accumulates) because each agentic-coding session re-derives architectural decisions wit…

  • Anthropic

    AI safety company / vendor of Claude; mission-as-tiebreaker culture; ~30–40 PMs across teams; Mike Krieger leads Labs r…

  • Building Is Cheap, Arguing Is Expensive

    "In technical debate, code wins": generate three PRs vs whiteboard; prototype over design doc; reduce design docs

  • Claude Code

    Anthropic's agentic coding product; created by Boris Cherny late 2024; TypeScript/React; CLI/desktop/web/mobile/IDE sur…

  • Where Does Agent Harness Work Remain Durable as Models Improve?

    Durable harness work lives at external-reality boundaries: repo-local source of truth, mechanical verification, context…

  • Fiona Fung

    Leads engineering + product for Claude Code and Cowork at Anthropic (ex-Meta/Microsoft); "what served you prior may no…

  • LLM-as-Compiler Knowledge Base

    Karpathy's architecture: LLM incrementally compiles raw docs into a persistent interlinked wiki, replacing RAG with a 4…

  • Managers as ICs

    Every Claude Code manager starts as an IC; flat org; agentic coding collapsed the onboarding cost that pushed managers…

  • AI Engineering & Agent Tooling

    Map of Content for the ai-engineering domain — 36 concepts. Curated entry point; see Home for all domains.

  • Open Questions Backlog

    _96 pages with open questions, as of 2026-06-14._

  • The PRD-Replacement Spectrum at AI-Native Speed

    Four positions (grill-then-PRD → lighter-PRD → build-to-decide → prototype-is-spec) are one spectrum once you decompose…

  • Verification as the New Bottleneck

    Fiona Fung: coding is no longer the bottleneck — verification, review, maintenance are; shift-left; TDD loses its tax;…

  • Where Does the Why Live?

    Rationale (the 'why') is well-homed at authoring time — it's the recorded why-not-what conversation and the grilling se…

Related articles
  • Claude Code

    Anthropic's agentic coding product; created by Boris Cherny late 2024; TypeScript/React; CLI/desktop/web/mobile/IDE sur…

  • Harness Shrinkage as Models Improve

    Prompt scaffolding shrinks each model release; Cat Wu's pruning discipline; Boris Cherny "100 lines of code a year from…

  • Design Concept Grilling

    Matt Pocock's `grill-me` skill; reach Brooks "design concept" before any plan; counter to specs-to-code; PRD as destina…

  • Fiona Fung

    Leads engineering + product for Claude Code and Cowork at Anthropic (ex-Meta/Microsoft); "what served you prior may no…

  • Building Is Cheap, Arguing Is Expensive

    "In technical debate, code wins": generate three PRs vs whiteboard; prototype over design doc; reduce design docs