Periphery Skill Manual

/save

Before: you finish a Cowork session where you reviewed Hue's CPL spike, decided to test creative variant B, and committed to check with...

The aha: Before: you finish a Cowork session where you reviewed Hue's CPL spike, decided to test creative variant B, and committed to check with Yvette about Lasso attribution. You close the tab. A week later you're prepping for the Marcon check-in and you can't remember what you decided about creative B, whether you ever talked to Yvette, or what the CPL actually was. The session disappeared. Now you type /save at the end of the conversation and Claude routes what you discussed to the right files — performance.md gets the CPL note, campaign.md gets the creative test status, and MEMORY.md gets the Yvette action item. Next session, Claude reads all of that back automatically. Your context compounds instead of evaporating.

What you need to know (2 min read)

/save is the end-of-session habit that keeps your client context fresh. You run it after you've done work in Cowork — reviewed performance, had a conversation about a campaign, drafted a brief, debriefed a meeting — and Claude writes the relevant bits to the right files.

What it updates:

Local-first by default. /save writes to your local Claude folder first. If the files are ready to share with the team, you explicitly say "yes, sync to Drive" and Claude handles it. Most work stays local — Drive is for sharing finished work.

How it works

When you run /save, Claude walks through four steps:

Step 1 — What client/project was this about? Claude looks for file paths you read or wrote, client names mentioned, account IDs referenced. If ambiguous, it asks once.

Step 2 — What kind of work? Pattern-matches the conversation against the client context file framework:

A single save can touch multiple files. That's normal.

Step 3 — Personal vs. project. Action items for you go to MEMORY.md. Decisions/state about the client go to context files. Things that are both go to both.

Step 4 — Where to write?

Three maturity tiers

/save adapts to your experience.

Learning tier (default, first 5 saves) — Conversational, file-by-file. Claude shows each file it wants to update, explains what the file is for, and asks for approval. Teaches the framework while doing the work.

Confirming tier (after 5 saves) — Single-screen confirm-all. Claude shows the full diff for all files in one view, you approve with one click.

Trusting tier (after 15 saves) — Silent auto-route. Claude writes everything and shows a one-line summary. Undo is always available via /save --undo.

You can override the tier any time: /save --tier learning if you want hand-holding back.

Try it yourself

  1. Have a real conversation in Cowork first. Review some Marcon performance data, discuss a campaign, draft something. /save needs something to save.
  2. Type: /save
  3. In learning mode, Claude will walk you through file by file. Watch where it wants to write each thing — that's how you learn the framework.
  4. Example session to try: tell Claude "I just looked at Marcon Hue's performance. CPL is up to $52 from $42 last week. Variant B is winning on CTR but A is converting better. Need to check with Yvette about Lasso attribution." Then run /save. You should see it:
  1. After it writes locally, Claude will ask if you want to sync to Drive. Try each path — "not yet" (stays local), "yes" (syncs), "always" (sets the flag).

Quick reference

DetailInfo
Command`/save [--undo \--tier <learning\confirming\trusting> \--mark-synced \--sync]`
Who uses itEveryone
Time to run30 seconds (trusting) to 3-5 min (learning)
When to runEnd of any Cowork session where you did real work
Default destinationLocal Claude folder (Drive is optional, explicit)
Flags--undo to revert, --tier to change mode, --sync to push pending local files to Drive, --mark-synced to clear flags after manual upload

Tips

Related