Three-Layer Directory Fix: How to Stop AI from Losing Your Work

AI Cannot See Your Desktop

The author discovered that AI only sees file paths and names — not icons, thumbnails, or folder nesting. A file named "New Document 3.docx" is meaningless to AI, while "20260712-公众号钩子四种-定稿.md" instantly tells AI the date, topic, and status. The key insight: directories are not for human tidying; they are a coordinate system for AI.

Three Layers, None Optional

The author structures the workspace into three distinct layers:

Layer 1 — Raw Materials (素材层): Unprocessed inputs — good articles, fleeting ideas, reference materials, topic inspirations. These are raw ingredients, not ready to use.

Layer 2 — Work-in-Progress (产物层): AI-generated outputs — first drafts, formatted HTML, generated images, tables. These are usable but not finalized.

Layer 3 — Finalized Assets (沉淀层): Verified, reusable deliverables — final copies, approved angles, templates, standards, proven workflows. This is the "capital" reused for future work.

Mixing layers causes confusion: the author once published a half-finished draft because draft and final versions shared a folder. After separation, the layer alone reveals the file's maturity.

Naming: Three Mandatory Components

Every filename must contain date, topic, and status — no exceptions. Format: YYYYMMDD-topic-status.ext. Examples:

20260924-远程控制日常-初稿.md

20260924-远程控制日常-定稿.md

20260924-远程控制日常-排版.html

20260915-选题库9月版-v2.md

Rules learned from mistakes:

Use eight-digit dates for correct sorting.

Write topics in plain language; avoid abbreviations like "WB-KB-v3" that become unreadable months later.

Status must be one of four: 初稿 (draft), 待改 (needs revision), 定稿 (final), 归档 (archived). This single rule saves the most time.

Bad names: 新建文档 3.docx, 最终版2真的最终版.md. Good name: 20260924-公众号排版-定稿.html — recognizable by both future you and AI.

Knowledge Base: Directories Manage Location; Knowledge Base Manages Continuity

Directories answer "where is it?" on the current machine. Knowledge bases answer "is it still there after this chat closes?" The author learned that even a perfect local folder structure fails when switching devices or starting a new AI session — the AI becomes amnesiac unless context persists.

Rule: Raw Materials and Work-in-Progress can stay in local directories. Finalized Assets must enter the knowledge base. These are the capital; losing or corrupting them wastes all prior effort.

Cloud storage ≠ knowledge base. Cloud drives store files for human download. Knowledge bases store context for AI to pick up instantly. "Cloud drive is a warehouse; knowledge base is a workbench. Warehouse items must be moved to use; workbench items are at hand."

Attaching a knowledge-base document to a task tells AI: "Continue from this, don't reinvent context." One action saves ten minutes of re-explanation.

A critical safeguard: the knowledge base requires explicit "accept" before overwriting any document. The author caught AI silently rewriting a passage into formal corporate tone; without acceptance, the original remained untouched.

Format standards for the Finalized Assets layer: editable text in Markdown, tabular data in CSV, publish-ready output in HTML. Wrong container choice creates downstream friction.

Three Common Mistakes (All Committed by the Author)

Only Raw Materials, no Finalized Assets: Endless accumulation, zero extraction. Hundreds of files stored, none reusable.

Work-in-Progress left local, never promoted to knowledge base: "I'll review later" leads to loss on device switch.

Multiple parallel versions per topic: Each edit saved as a new file; months later three versions conflict. Solution: one file per topic, iterate in place, distinguish versions via date in filename.

Closing Thought

Organizing directories is really organizing your own mind. Clarifying what is raw material, what is half-finished, and what is capital makes your intent explicit — so AI finally knows where to look. The author now simply instructs: "Pull the public-account angle from the Finalized Assets and write to that." AI retrieves it more accurately than the author. 136 articles once sat as dead data on a disk; now they function as a second brain — not because AI got smarter, but because they finally got coordinates.