| name | tune-up |
| description | Runs a weekly self-improvement loop across 5 approval-gated steps. Makes ONE change per week to the user's Cowork AI OS. Trigger when the user says "/tune-up", "weekly tune-up", "tune up", or "improve my system". |
Tune-Up — The Weekly Self-Improvement Loop
You walk the user through 5 gates that audit their Cowork AI OS and make ONE specific improvement. This is the engine that keeps Cowork AI OS evolving.
Critical rule: ONE change per week. Not three. Not five. ONE. Compound interest, not overwhelm.
Critical rule: Approval at every gate. No autonomous changes. The user types continue, go, skip, or rollback at every gate.
Read these first
about-me/audit-log.md — last entry (for delta calculation)
about-me/memory.md — the last 7 days of entries (recent activity for the pulse)
about-me/memory-log.md — if it exists, the cold archive of older memory entries. Read it for the long-view in Gate 2 (scoring) and Gate 3 (proposing) — older patterns, prior tune-up actions, what's been tried before. Skip if absent.
about-me/about-me.md, business-brain.md, connections.md
- the 4 Cs rubric in the appendix below
GATE 1 — PULSE
Read the last 7 days of memory.md entries. Summarize in 3 bullets:
- Built: [what got created or shipped]
- Finished: [what got closed out / completed]
- Stalled: [what's been sitting open]
Show to the user. Say:
"That's your week. Type continue to keep going, or tell me if I missed something."
Wait for continue.
GATE 2 — SCORE
Run the same audit logic as /audit (see the 4 Cs rubric in the appendix below). Compute a fresh score across the 4 C's.
Compare to the last entry in audit-log.md. Compute the delta.
Show:
| C | This week | Last week | Delta |
|---|---|---|---|
| Context | X/25 | Y/25 | ±Z |
| Connections | X/25 | Y/25 | ±Z |
| Capabilities | X/25 | Y/25 | ±Z |
| Cadence | X/25 | Y/25 | ±Z |
| **Total** | **X/100** | **Y/100** | **±Z** |
Then identify the one biggest leverage gap. Be specific:
"Biggest gap: [Connections — Tasks bucket]. Specifically: you said in onboarding that your tasks live in ClickUp, but ClickUp isn't connected. Adding it would let me pull your task list for the morning brief, identify what's at risk, and draft updates."
Say:
"Type continue to see what I propose, or wait — different gap if I picked the wrong one."
Wait for continue.
GATE 3 — PROPOSE
Before picking a form, run the "What's new since last tune-up?" check (v0.10.2) below — it can surface a new connector or skill that's a better fit than anything currently on the table. Then pick exactly ONE improvement using the forms that follow.
What's new since last tune-up? (v0.10.2)
Before offering form (b) build a new skill or form (c) add a connector, check what's NEW in the official Anthropic directories since the last tune-up.
Procedure:
-
Find the last tune-up date — read about-me/memory.md for the most recent entry containing "tune-up". If never run before, use 30 days ago as the baseline.
-
Live-fetch both directories in parallel:
-
Identify "new since last tune-up": any connector or plugin that has a "New" badge OR was first listed after the last tune-up date. (For the JSON marketplace files, "New since last tune-up" means added since the last commit hash we saw — if uncertain, use last-7-days as a safe heuristic.)
-
Cross-reference with about-me/business-brain.md to score newness × relevance. A "New" coffee-business connector matters to a coach who serves coffee shops; a "New" legal connector doesn't.
-
Surface to user IN GATE 3 (before forms b/c):
"What's new in Anthropic's directories since your last tune-up:
Want to install any of these? Or should we skip to the rest of Gate 3?"
-
Honor user's answer. If they want one, branch into form (c) (connector) or form (b) (build/install a skill) using the live-fetched details. If they say skip, log it and proceed to the rest of Gate 3.
-
Log the result — append a one-line entry to about-me/memory.md:
YYYY-MM-DD: tune-up surfaced N new directory items, M accepted
If live fetch fails:
Skip this step gracefully. Don't block the rest of /tune-up. Tell the user: "Live directory check unavailable — skipping the 'what's new' step. We can still walk through forms (b) and (c) using the bundled snapshot."
Why this step exists (v0.10.2 rationale):
Anthropic adds 5-20 new connectors per week. Without this check, /tune-up only proposes from a stale bundled snapshot. With it, /tune-up becomes the weekly habit that keeps the user's stack current — without making them browse the directory themselves.
Forms — pick the one that fits the gap
(a) Update an existing custom skill
The user's own skills, built via /add-skill. NOT Cowork AI OS's bundled skills — those are read-only and update only when the Cowork AI OS plugin updates. Show before/after diff of the SKILL.md or referenced file.
(b) Build a new skill
Format a spec for the skill. Hand off to /add-skill (which checks Skill Creator is enabled, then drives the construction). The spec should include:
- name, description (the YAML frontmatter)
- when it fires
- what it reads
- what it does
- what it writes
(c) Add a connector
Name the connector and what it unlocks. Walk through enabling it.
(d) Refine a context file
Show before/after diff of an about-me/ file (about-me.md, business-brain.md, writing-rules.md, connections.md). This is also where you go when Cowork AI OS itself needs better context — refine the user's own files, not Cowork AI OS's bundled files.
Say:
"My proposal:
[The single specific change]
Why this: [leverage reasoning]
Cost: [time + complexity, e.g. '5 minutes, no new accounts needed']
Type go to make this change, or skip to defer to next Friday."
Wait for go or skip.
If skip: jump to GATE 5 (log the skip).
GATE 4 — EXECUTE
Make the change. Show the result.
If it was a file change, show the actual file. If it was a connector, walk the user through enabling and confirm it's working. If it was a new skill, hand off to /add-skill and let that skill drive.
Say:
"Change made. Confirm it looks right by [specific verification step]. Type approve to keep it, or rollback to undo."
Wait for approve or rollback.
If rollback: undo the change. Note in memory.md why it was rolled back. Skip to GATE 5.
GATE 5 — LOG
Append to about-me/audit-log.md:
### YYYY-MM-DD — Triggered by: tune-up
- **Total:** X/100 (delta: ±Y)
- **Context:** X/25
- **Connections:** X/25
- **Capabilities:** X/25
- **Cadence:** X/25
- **Top leverage gap:** [the gap from Gate 2]
- **Action taken this week:** [the change from Gate 3 + 4, or "skipped"]
Append to about-me/memory.md:
### YYYY-MM-DD — Weekly tune-up
- **Score:** X/100 (delta: ±Y)
- **Top gap:** [gap]
- **Action taken:** [change made / skipped / rolled back]
- **Next week likely focus:** [your prediction based on remaining gaps]
Then run the auto-archive check per the rules in about-me/memory.md (Auto-archive section). If memory.md now exceeds 30 entries, move the oldest overflow into about-me/memory-log.md (create the file if it doesn't exist). Silent.
Closing
Tell the user:
"Tune-up done. Score: X/100. This week we [summary of action].
Next Friday's tune-up will probably look at [your prediction]. Until then, Cowork AI OS keeps running. Have a good week."
That's it. End cleanly.
Edge cases
User isn't at the computer when this fires (scheduled task scenario):
If the schedule fires and the user doesn't respond within 30 seconds at Gate 1, save your draft audit to about-me/audit-log.md (marked as pending), append to memory.md noting "tune-up paused at Gate 1 — user not at desk." Stop. When the user returns and says "continue tune-up," resume from Gate 1.
It's their first tune-up (no prior audit-log entry):
Use the baseline from onboarding (Phase 7). If even that doesn't exist, run the audit fresh and treat this as the baseline.
Score went DOWN:
Don't be alarmed. Tune-up isn't failing — sometimes context/files drift, or the user paused something. Diagnose: which C dropped? Was it because something stale crept in (e.g. about-me/ files weren't updated when a pivot happened)? The proposal in Gate 3 should address whatever caused the drop.
User wants to make MORE than one change:
Politely refuse. Say: "One change per tune-up is the design. Six months of one change a week is 24 improvements compounding. Trying to do five at once usually means none of them stick. Pick the most important — the rest will come around in future tune-ups."
User asks "can you update Cowork AI OS itself — like the /audit skill or /onboard?":
No. Cowork AI OS's bundled skills (/onboard, /audit, /level-up, /tune-up, /morning-brief, /voice-writer, /add-skill) are read-only — they live in ~/.claude/plugins/ and update only when the Cowork AI OS plugin itself updates from /plugin update cowork-ai-os.
If something about Cowork AI OS feels off, the right channel is:
- Note it in
about-me/memory.md with ai-os-feedback: tag
- The user reports it at
github.com/automatedmarketer/cowork-ai-os/issues
- It gets fixed in the next Cowork AI OS release
What /tune-up CAN modify: the user's own about-me/ files, their custom skills (built via /add-skill), their connectors, their scheduled tasks. Plenty of room there.
Appendix: The 4 Cs framework (the audit rubric)
Score the user Cowork AI OS on a 0-100 scale. 25 points per C.
C1 - Context (0-25)
How much does Claude know about the user and their business?
- 0-8: No about-me/ files, or files have ~~placeholders unfilled.
- 9-16: About-me/ has the basics but writing-rules.md has fewer than 3 samples.
- 17-22: All 4 about-me/ files complete. Writing samples real. Memory.md has 5+ entries.
- 23-25: Above plus business-brain.md kept current and memory.md gets appended every session.
C2 - Connections (0-25)
How many of the 7 buckets (Revenue, Customer, Calendar, Comms, Tasks, Meetings, Knowledge) are reachable?
- 0-8: 0-2 buckets reachable.
- 9-16: 3-4 buckets reachable.
- 17-22: 5-6 buckets reachable.
- 23-25: All 7 reachable with safe permissions (writes need approval).
C3 - Capabilities (0-25)
How many things can Claude actually do?
- 0-8: Cowork AI OS only, never run a skill beyond /onboard.
- 9-16: Cowork AI OS skills used regularly. No additional plugins.
- 17-22: Cowork AI OS plus 1-2 Anthropic role plugins. 1-2 custom skills built via /add-skill.
- 23-25: Above plus custom skills get refined every few weeks.
C4 - Cadence (0-25)
How much works on a schedule, vs only when asked?
- 0-8: No scheduled tasks.
- 9-16: Morning brief scheduled. Nothing else.
- 17-22: Morning brief plus weekly tune-up plus at least one custom routine.
- 23-25: Multiple routines firing reliably. Computer-awake reliability solved.