R&D Management 12 min read

AI Skips the Workflow and Writes Correct Code—What Human Value Remains in 2026?

The article examines a real auto‑sign project where a large model directly edited code, bypassing the intended OpenSpec‑based workflow, and argues that while AI can produce usable first drafts, developers still provide essential value through boundary setting, acceptance arbitration, source truth maintenance, organizational memory, and workload reduction decisions.

Frontend AI Walk
Frontend AI Walk
Frontend AI Walk
AI Skips the Workflow and Writes Correct Code—What Human Value Remains in 2026?

Introduction

A large model can directly modify code without running the full AI‑assisted workflow when the domain is familiar, a reference implementation exists, and the request is simple. The author verified this behavior personally.

What You Expect vs. What Actually Happens

Expected Process

你:描述需求 + @工作流
 → Agent 严格执行阶段 0→5
 → OpenSpec 契约、4a 验证、4b 验收
 → 交付报告 ✅
 → 你才能放心合并

Real Process (auto‑sign case)

你:描述需求 + "请使用 ai-workflow"
 → Agent 读了 CHANGELOG 和 sign-auto.vue
 → 直接改 standard-file.vue、utils/index.js
 → 没写需求对齐、没写 OpenSpec 文档,没跑 4a/4b、没出 delivery‑report...
 → 你一看:逻辑好像对,能联调(直接写的产物)

The model + built‑in tools often skip a custom workflow under the conditions above. This is not an illusion; it is a reproducible outcome.

Are Workflow, OpenSpec, and Superpowers Obsolete?

Short answer: No, but their battlefield has shifted.

"Let AI write smarter." – early selling point

By 2026 the model and agents already handle the "smart" part. The workflow now sells accountability, repeatability, degradability, and organizational contracts.

Three Task Categories and Three Approaches

Solo work, familiar module, modify 1‑2 files, easy rollback – Direct coding is cost‑effective; a full workflow is overkill.

Reference implementation exists, clear verbal requirements – Direct coding is often sufficient; a retroactive run can supplement the workflow.

Team collaboration, audit, compliance, financial or legal impact – Direct coding is insufficient; a full workflow is mandatory.

In the auto‑sign facial‑recognition overhaul the case fell between the second and third categories: the code was correct, but missing checks such as submitFileSurvey parameters, backend support, and E2E failures should have been captured as Open Questions or deferred acceptance criteria.

Human Role Beyond Prompt Writing

Prompt engineering is only the entry point. The irreplaceable human contributions are organized into five decision areas:

1. Boundary Definition

Decide what must be reviewed by a person versus what an agent can resolve (e.g., isSkipSign handling types 1/2 vs. 3).

2. Source of Truth

Choose the authoritative document (dialogue, PRD, OpenSpec, backend YApi) and store it consistently under openspec/changes/<id>/ rather than ad‑hoc output folders.

3. Acceptance Contract

Define what “correct” means; static code checks can be automated, but real‑world integration tests require human or QA verification.

4. Organizational Memory

Hard gates like IMP‑036 and IMP‑038 ensure that future engineers or agents cannot skip critical stages, preserving knowledge across personnel changes.

5. Trade‑off Decisions

Know when a full workflow is unnecessary and apply a “When NOT to Use” decision tree to avoid over‑engineering.

The "Initial Draft" Trap

A seemingly perfect first draft can create a false sense of completeness, leading to missing acceptance matrices, open questions, delivery reports, and E2E evidence. As models improve, this trap deepens because the draft looks more like a finished product.

Comfort zone:
 需求 → 直接改代码 → 结论还行 → "工作流可有可无"

Blind spots:
 - 没有 AC 矩阵 → 联调失败时不知道漏了哪条
 - 没有 Open Questions → 后端字段名错了要重新猜
 - 没有 delivery‑report → 三个月后没人记得为什么删了验证码弹窗
 - 没有 E2E 证据 → 上线后人脸在部分机型失败,责任说不清

The workflow’s purpose is to prevent technical debt from being mistaken for completed work.

Iterating the Workflow

What Not to Do

Stack more skills expecting the agent to obey better – diminishing returns.

Run the full 0→5 pipeline for every line change – ritualistic.

Hard‑bind Superpowers, OpenSpec, and custom workflow – agents may selectively ignore constraints.

What to Do (Current Practices)

Hard Gate : IMP‑036 coding before/after door control – declares that incomplete workflow is unacceptable.

Path Standard : IMP‑038 stores files under openspec/changes/<id>/workflow/ – avoids dual truth sources between output and OpenSpec.

Retroactive Mode : retroactive · skipStage3Coding – accepts the reality of “write‑first‑document‑later”.

Size Awareness : Apply “When NOT to Use” and limit execution to only necessary stages – lets humans retain reduction decisions.

The ultimate goal of workflow evolution is not 100% agent compliance but team awareness of when compliance is mandatory.

Actionable Advice for Different Audiences

Individual Developers / Prototyping

Describe requirements directly to AI for code changes.

At minimum, list three acceptance criteria, even without a full OpenSpec.

Before merging, ask: “If I switch machines or agents, can I still prove this works?”

Team / Production / Compliance (e.g., facial, payment, signing)

When using the workflow, enforce persisting change + workflow + 4b artifacts.

Allow retroactive runs but never mark everything green without an AC matrix.

Store the source of truth in openspec/project.md, not in oral agreements.

Maintainers of AI‑Workflow Tools

Strengthen enforcement, retroactive support, and path standards.

Focus on decision trees that clarify when a full run is unnecessary.

Use real projects like the auto‑sign case for regression testing instead of only updating skill text.

Conclusion

Stronger models don’t invalidate methodology; they shift the rationale from “help AI write right” to “help the organization stay accountable, auditable, and iterative.”

First drafts can be good for exploration but don’t replace a complete delivery process.

Human value lies in boundary setting, acceptance arbitration, source‑truth maintenance, and workload‑reduction decisions.

Workflows should evolve to become more precise, stricter, and lighter—not heavier.

Never treat a flawless first AI output as proof that the workflow is optional; stronger models increase the need for traceable accountability.

Appendix: auto‑sign Timeline (Fact Anchors)

First round : Declared ai‑workflow → direct coding → no verification persisted.

Reflection : Questioned why the workflow didn’t run fully.

IMP‑036 : Full‑run hard gate (fullRunMode, check‑pre‑coding‑gate).

Retroactive : retroactive 0→5, static 9/9 AC, E2E deferred.

IMP‑038 : Documentation moved to openspec/changes/standard-contract-face-auth/.

Original Source

Signed-in readers can open the original source through BestHub's protected redirect.

Sign in to view source
Republication Notice

This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactadmin@besthub.devand we will review it promptly.

AutomationAI programminghuman-in-the-loopsoftware workflow2026OpenSpec
Frontend AI Walk
Written by

Frontend AI Walk

Looking for a one‑stop platform that deeply merges frontend development with AI? This community focuses on intelligent frontend tech, offering cutting‑edge insights, practical implementation experience, toolchain innovations, and rich content to help developers quickly break through in the AI‑driven frontend era.

0 followers
Reader feedback

How this landed with the community

Sign in to like

Rate this article

Was this worth your time?

Sign in to rate
Discussion

0 Comments

Thoughtful readers leave field notes, pushback, and hard-won operational detail here.