Type /tot at any point while you work. Claude reads what you're working on, suggests 3 or 4 different approaches, lets you pick, builds and scores them, and keeps the winner.
Example run · kraut-deck
One /tot call, drawn as a tree
winning pathpruned (scored too low)backtrack when a branch failsscores out of 10
The flow
What happens after you type /tot
Claude
Read context
Current task, open files, project type, recent commits.
Claude
Propose 3 or 4 options
Each with pros, cons, cost and how it will be scored.
You
Choose
Pick one, tick several, or say "all".
Claude
Build and score
One after another, or in parallel sessions.
Claude
Keep the winner
Show the score table, apply the best option, and backtrack if it fails.
Step 2 · what you'll see
The option picker
Which Merry-Cat arrangements should I build?
tick 1 or more
✓
A · Meow lead + techno
Pro
Melody stays up front
Con
Busy low end
Cost
Low, reuses v6 cat
Score
Peak dBFS, clipping, CPU
✓
B · Meow + Bhangra beat
Pro
Fresh groove
Con
Clashes with echo
Cost
Medium
Score
Same metrics as A
C · Ambient, heavy reverb
Pro
Calm, cinematic
Con
Melody may get buried
Cost
Low
Score
Melody clarity
D · Chiptune meow
Pro
Playful, light on CPU
Con
New synth patch needed
Cost
High
Score
Same metrics as A
Build A + BBuild all 4Run in parallel
Inside step 2 · how it chooses
How the agent picks the options
8
Brainstorm
Reads your task and files, then lists every approach it can think of.
5
Make them differ
Merges near-duplicates, so each option is a different idea, not a tweak.
4
Rate each one
Scores it on the four checks below and drops the weakest.
3
Show you
Shows the top 3, up to 4, with pros, cons, cost and how it will be scored.
Fits the goalDoes it solve what you asked?
Can be measuredIs there a test or number to score it?
CostTime and tokens to build it.
RiskCould it break existing work?
It only suggests. An option it cannot measure is dropped, and you always choose what gets built. Counts above are examples.
Auto-adapts
How scoring changes per project
Project type
Scored by
Audio (kraut-deck)
Offline render: peak dBFS, clipping, loudness (LUFS), CPU
Web app
Tests pass, Lighthouse score, screenshot check
Bug hunt
Bug reproduces, then the fix makes it stop
Design / copy
Side-by-side preview, you judge
Step 4 · run modes
One after another, or in parallel
Sequential
fewer tokens
Built one by one in this session. Good default for 2 or 3 small options.
Parallel
faster
Each option runs in its own session and worktree, then a judge compares them. Best for big or slow builds.
Setup · nothing to install
~/.claude/skills/tot/
SKILL.md # the instructions
scoring.md # per-project-type metrics
One folder in your home directory. It works in every project, and /tot runs it.