Where Is That Plan? Keeping AI Project Docs Findable

You know there was a plan. You and the agent wrote it together two weeks ago — the version 1.2 features, what was in and what was out. Was it in the chat? A note? The docs folder? The other docs folder?

You find three documents with "plan" in the name. None of them is the one.

The final part of No Process — and the last post of this 20-part run on the pains of working with AI.

The short answer#

• AI work produces a lot of documents — specs, plans, research — and they scatter across chats, notes and random folders.
• Every document that matters goes in the repo, in a predictable place, with a name that sorts by date or version.
• Link each task to its document so you never search: the task line points to the plan.
• Ask the agent to save plans to files, not just to answer in the chat.

AI work produces a lot of paper#

Agents are great at writing: specs, plans, research summaries, checklists. That's good — until they pile up. The usual failure isn't too few documents. It's documents in too many places:

• In chats, where they vanish when the session scrolls away.
• In notes apps, disconnected from the code they describe.
• In the repo, but in random folders with names like plan-final-2.md.

The one you need is always in the last place you look.

One place, predictable names#

Pick a layout once and use it in every repo:

plan/
  mvp/mvp-plan.md
  update/v1.2.0/feature-tip-presets.md
  update/v1.2.0/fix-zero-bill-crash.md
research/
  2026-09-30-competitor-listings.md

Two rules do most of the work: documents live in the repo (so git, you and the agent all see them), and names sort — by version for plans, by date for research.

Plans go to files, not answers#

When you ask an agent for a plan, ask for a file:

"Write the plan to plan/update/v1.2.0/feature-tip-presets.md, then summarize it in three lines."

A plan that only exists in a conversation is as good as lost — the same lesson as memory lives in files.

Link the task to its document#

The final step removes searching entirely: each task in the state file links to its document.

{ "name": "tip presets", "phase": "3/6", "next": "build the presets screen",
  "link": "plan/update/v1.2.0/feature-tip-presets.md" }

Open the task, and the plan is one click away. Same idea as writing the run commands in the repo: put the answer where the question will be asked.

The whole series in one line#

Over twenty posts the fix has been the same: move what's in your head into small files in the repo, and look at them on one screen. Phases, next steps, stuck reasons, sessions, run commands, plans. That's the minimum process — and it's what turns a busy day into a shipped one.

Keep reading#

• No Process, Just Prompts: Why Solo AI Projects Fall Apart
• How Did I Run This Project Again? Write the Command Down
• Memory Lives in Files, Not Chats: Managing AI Agent Context