Import Notes

import-notes

Import the user's existing research notes (scattered across other notes apps, folders, or pasted directly into the prompt) into this wiki, file them in the right places per this repo's current conventions, and update the relevant indexes.

When to trigger

The user says any of (or anything semantically similar):

• "import notes" / "integrate my existing notes" / "pull these notes in"
• "migrate from the old project / another notes app"
• Hands you a folder path, a batch of files, or a big blob of pasted content and asks you to "classify and file it per the rules"

Input forms

Two input modes are supported (mix is fine too):

1. Folder: the user gives a directory path containing anything the Agent can read (.md / .txt / .pdf / .docx / .html / web clippings / code comments, etc.). Walk and read.
2. Inline prompt: the user pastes note content directly into the conversation. Split into individual items by blank lines or headings.

If the user gives neither a path nor inline content, proactively ask.

Workflow

Step 1: Learn the current rules of this repo

Do NOT hardcode classification rules. The only premise of this skill is: this repo already has a set of live conventions. The actual routing must be read by the Agent from the repo on the fly:

1. Read CLAUDE.md — learn the reading order, update rules, and do-not-read areas
2. Read Overview.md — learn the project skeleton, task navigation, and doc-organization conventions
3. Read each module's Overview (Data/DataOverview.md, Training/TrainingOverview.md, Eval/EvalOverview.md, ResearchNotes/ResearchNotesOverview.md, etc.) — learn what already exists and the naming conventions
4. If the repo has additional XXX-conventions.md / supported-XXX.md style docs, read those too

Based on the above, form "this repo's current classification rules" as the basis for this import.

Step 2: Collect the source material

• Folder input: list all files, read each one
• Inline input: split into items by heading/blank line
• Assign an internal id to each item (e.g. src-01, src-02…) for later reference when presenting the plan

Step 3: Classify each item

Using the rules learned in Step 1, decide for each item:

• Which directory / which file it goes to
• Create a new file, or append to a section in an existing file
• Whether to split it (if an item spans multiple topics, split into multiple items and route each)
• Whether to add a new [[ ]] link in the corresponding Overview's index

If an item cannot be classified confidently, do not force it in; park it and wait for the user to decide.

Step 4: Output the migration plan (dry-run, default required)

Present it as a checklist:

Import plan for this run (N items total):

[New]
  src-01 → ResearchNotes/Models/xxxx.md
  src-03 → ExperimentalTracks/20260405-yyy.md

[Append]
  src-02 → Overview.md § Current main-line method
  src-05 → Data/DataOverview.md § supported-datasets

[Index updates needed]
  ResearchNotesOverview.md § Models
  Overview.md § Task navigation

[Needs your decision]
  src-04: covers both training tuning and eval metrics; suggest splitting into two, but please confirm the split point
  src-07: looks like a diary, unsure whether to import

Step 5: Wait for confirmation, then execute

• By default you must wait for user confirmation before writing
• If the user invoked batch mode (see below), skip confirmation and execute on the Agent's current judgment. Items parked as "needs decision" must still be asked.
• When writing: at the top of each imported note (or at the top of the section it was appended into), keep a one-line provenance note:

  > Imported from: <original filename or "inline prompt"> (YYYY-MM-DD)
  

• Update the corresponding Overview's index: add a [[ ]] link + one-line summary

Step 6: Remind the user they can re-import later

When done, tell the user:

Import complete. If you have more batches of notes later — another folder, another export, or scattered notes you think of later — you can always call this skill again and do it in multiple passes.

Batch mode

If the user explicitly says any of the following on trigger:

• "batch mode"
• "batch"
• "no per-item confirmation"
• "just run through it"

Then:

• Skip the user-confirmation part of Step 4 (but still print the plan for audit)
• Items that clearly can't be classified still need to be parked and asked — don't cram them somewhere

Edges and cautions

1. Do not modify the source: source files / source prompt are read-only — no deletion, no moving
2. Dedup: before writing, quickly check whether the target file already has a highly similar entry; if so, append / merge by default rather than creating a duplicate file
3. Do not auto-import obviously unrelated content (e.g. personal diary, code snippets unrelated to this project) — list them and ask
4. Respect the "do not read by default" areas: archive/, ExperimentalTracks/abandoned/, long experiment logs, and similar areas marked as do-not-read-by-default in this repo's rules — do not write into them (unless the user explicitly directs)
5. Naming consistency: new files should match the naming style of existing files in the target directory (case, separators, date prefixes, etc.)