Three Prompts

I want to turn something I already solved with you into a skill I never
have to re-explain.

Before you do anything, ask me these one at a time and wait for each answer:
- What did we just do in this conversation (or the one I paste below) that
  I will need again?
- What do I call it when I ask for it next time? Give me a name I would
  actually type.
- Where do my skills live? Point me to the folder if you already know it,
  otherwise ask.

Then read back the actual steps we took, pulled from this conversation's
real actions, not a guess at what should have happened.

Write the skill with three sections:

1. WHAT IT DOES - one line, so I remember why I built this in six months.
2. HOW TO RUN IT - the steps, in the order that actually worked. Call out
   anything I had to correct you on, so version one already knows.
3. GOTCHAS - empty for now. This is where the failures go once I hit them.

Before you save it, tell me:
- The one step most likely to break the first time someone else runs this.
- Whether this belongs as its own skill or as a step inside a skill I
  already have. Check what I already have before adding a new one.

Save it, give me the exact trigger phrase, then run it once in front of me
so I can watch it fire.

I want to know if a task I keep doing by hand is worth turning into
something that runs on its own.

Before you do any work, ask me these one at a time:
- What is the task, in plain words, start to finish?
- How often does it happen, and how long does it take me each time?
- Which apps or tools does it touch? Email, calendar, a spreadsheet,
  Chrome, a specific app.

Then walk the task apart into its steps and score it honestly:

1. Which steps are a decision only I can make, and which are just
   repeating what I did last time?
2. What percentage of the whole task could run without me? Say the number.
   Do not round up to make the case for building it.
3. If it is under 80 percent, tell me straight it is not worth automating
   yet, and name the one step holding the number down.
4. If it clears 80 percent, write the workflow as a numbered list of exact
   actions, one line per action, the way I would hand it to a new hire.

Then tell me:
- The one failure mode most likely to happen the first time this runs
  unattended, and what should happen when it does. Stop and tell me, or
  keep going and log it.
- Where this workflow should log what it did, so I can check on it
  without watching it run every time.

Build it only after I say go.

I have one task that is really three or more separate jobs wearing a
trench coat. Help me split it and run it in parallel instead of one at
a time.

Before you do anything, ask me:
- What is the task, and what does done look like?
- Can you already see the natural way to split it, or do you need me to
  describe the pieces?

Then split it into pieces that do not need each other's context to
complete. Tell me the split before you launch anything, and flag if any
piece actually depends on another. Those cannot run in parallel, say so.

Launch one agent per piece, in parallel. For anything you are about to
judge or grade rather than just build, use a separate agent to review it
than the one that built it.

When they all finish, bring the results back together yourself and tell
me:
- What each agent actually did, in one line each.
- Where they disagreed or produced conflicting output, if anywhere.
- The one piece that took the longest, and whether splitting it further
  next time would help.