Next Step Planner
Next Step Planner
A daily AI planner that turns your routine into a ranked, explainable "what to do next" list.
Sleep, work, commute, meals, exercise, chores, leisure, right-now context — feed it what you're comfortable sharing, and it reasons through what makes sense to do next. Every suggestion comes with a confidence level and a reason you can question. Report what you actually did, and it re-ranks on the spot. Nothing gets saved or reminded without your explicit confirmation first.
Built as a Claude Code skill. Self-corrects instead of pretending to be certain.
Why this exists
Most "productivity AI" either guesses silently or demands a rigid daily log. This agent does neither:
- Explainable, not magic — every suggestion says why, based on which parts of your routine it actually has data for.
- Self-correcting — tell it what you really did, and it updates the ranking immediately, on the record.
- No silent memory — nothing becomes a "long-term habit" without you confirming it, even after repeated patterns.
- No silent reminders — reminders are proposed, shown to you in full, and only scheduled after a second explicit confirmation.
- Honest about failure — if a tool isn't available or a call fails, it says so plainly instead of pretending it worked.
What it does
You → short, skippable check-in on 7 routine areas + current context
Agent → up to 5 ranked next-step hypotheses
├─ High / Mid / Low confidence (heuristic, not a probability)
├─ Traceable reasoning
├─ Data coverage (which routine areas it actually knows)
├─ Conflict flags (e.g. "tired" vs "deadline looming")
└─ Up to 3 follow-up questions that would change the ranking
Example
Input (partial check-in):
Slept 6.5h, worked since 9am with no break, lunch skipped, no exercise today, deadline at 6pm.
Output:
1. [High] Take a 15-min break before continuing work
Reason: No break in 6+ hrs of work; skipped meal raises fatigue risk before a hard deadline.
Coverage: sleep✓ work✓ meals✓ exercise✓ commute✗ chores✗ leisure✗
2. [Mid] Eat something light now, not after the deadline
Reason: Skipped lunch + afternoon deadline stacking cognitive load.
Conflict signal: deadline urgency vs. fatigue signals — flagged, not resolved for you.
Follow-up: How close are you to finishing the 6pm deliverable?
This is a hypothesis, not a prediction of what you'll actually do.
Self-correction loop
"Around 3pm, after lunch, I actually kept working instead of taking a break."
↓
Agent re-ranks this session immediately using that fact
↓
Asks: was this a one-off exception, a recent shift, or your real pattern?
↓
A single result never auto-becomes a long-term habit — even a repeated
pattern only becomes a *candidate*. You confirm before anything is stored.
Memory model
| Layer | Holds | Persists across chats? |
|---|---|---|
| Short-term | This session's current activity, recent commitments, today's exceptions | No |
| Medium-term | Recent outcomes, corrections, temporary routine shifts | Only if you confirm an expiry |
| Long-term | Explicitly confirmed fixed routines, stable preferences, constraints | Only if you confirm an expiry |
Before writing anything, the agent shows you the exact summary, the tier, the purpose, where it's stored, and how to correct or delete it — every time. No memory tool available? It just says so and works from this conversation only.
Reminders, done honestly
- Proposed, not scheduled — full plan (action, date/recurrence, time, timezone, lead time) shown for your second confirmation.
- Scheduled — only after you confirm, and only if the host actually has a working reminder/calendar tool. Success is only reported after a verifiable receipt from that tool.
No tool, no permission, or a failed call → you get a clear "not scheduled" result and a copy-pasteable plan, never a false "done."
Scope & limits
- Adult self-analysis only, with consent — never analyzes anyone else
- No mind-reading, no passive device monitoring, no behavior manipulation
- Not for medical, employment, credit, insurance, legal, or law-enforcement decisions
- Outputs are hypotheses, not calibrated probabilities or behavioral guarantees
Data sharing
No routine data, hypotheses, or reminder plans go to the skill's seller during normal use. Your chosen Agent host/model provider processes what you give it. Scheduling fields are only sent if you confirm a reminder and the host has that tool. Memory summaries are only sent if you explicitly opt into memory. Actual retention and deletion depend on your host and provider's own terms — this skill can't control that part.
Install
Runs as a Claude Code skill. Drop the SKILL.md into your skills directory, or load it through your Agent Client's skill installer.
This is a self-organization aid, not a scientifically validated behavioral model.
