跪拜 Guibai
← All articles
Artificial Intelligence · Programmers

AI Killed Your Flow State Because You Designed a Relay Race, Not a Pipeline

By 摸鱼师moko ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Teams adopting AI coding agents often see developer satisfaction drop because the workflow fragments attention and defers critical thinking to a painful final review. Restructuring the process around parallel document pipelines and human-gated stages recovers flow and applies the most expensive resource—business judgment—at the cheapest possible moment.

Summary

The loss of flow when coding with AI isn't an inevitable cost of the technology; it's a workflow design error. The core mistake is structuring human and AI work as a relay race—AI runs autonomously for hours while the developer waits, then dumps a finished product for review. This pushes all business-intent judgment to the final stage, where deviations are most expensive to fix and context has long been lost.

The fix is to exploit the natural sequence of planning documents. While AI generates a design document, the developer reads the requirements document. While AI breaks down tasks, the developer reads the design document. Both operate in the same context, neither idles, and judgment is applied when documents are short and cheap to change.

Confirmation gates between each stage are essential. The developer must approve the requirements before AI generates the design, and approve the design before AI breaks down tasks. This keeps the human as the metronome, always holding full context, rather than chasing after an AI that has already raced ahead.

Takeaways
AI handles Layer 1 (code correctness: syntax, tests, bugs) well; it fails at Layer 2 (business intent alignment) because it lacks sufficient context or hallucinates.
Treating human-AI collaboration as a relay race—AI runs autonomously, human reviews at the end—destroys flow and makes business-intent deviations exponentially more expensive to fix.
Spec-driven development (SDD) creates a natural pipeline: while AI generates the design doc, the human reads the requirements doc; while AI breaks down tasks, the human reads the design doc.
Applying human judgment at the document stage costs a fraction of fixing the same deviation after code is written, because documents are short, readable, and cheap to edit.
Confirmation gates between stages—human approves requirements before AI designs, approves design before AI breaks down tasks—prevent the developer from becoming a chasing reviewer.
The new flow state is not about typing code but about continuously holding the problem and exercising judgment at each gate.
Conclusions

The contradiction at the heart of many AI workflows is that teams simultaneously treat human judgment as irreplaceable while designing processes that try to remove humans entirely.

Waiting for AI to finish is not the problem; the problem is that the human and the AI are not working on adjacent, dependent artifacts that keep both parties in the same context.

Moving judgment upstream to the document phase is not just cheaper—it preserves the developer's mental model, which is the real asset that batch review destroys.

The confirmation gate is not a bottleneck; it is a pacing mechanism that prevents the AI from generating downstream artifacts before the human has validated the upstream ones, which is the exact failure mode of autonomous agent pipelines.

Concepts & terms
Layer 1 vs. Layer 2 judgment
Layer 1 covers code correctness (syntax, tests, bugs, standards compliance) and can be automated with rules and agent cross-review. Layer 2 covers alignment with business intent (field reuse, boundary decisions, interaction behavior) and requires human judgment because the criteria live in the developer's head, not in a spec document.
Confirmation gate
A human approval checkpoint between stages of an AI pipeline. The AI does not proceed to the next stage (e.g., generating a design document) until the human has reviewed and approved the current stage's output (e.g., the requirements document). This keeps the human in control of pacing and context.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗