Charge Recommendations

When to use

Activate this skill when the user asks anything along the lines of:

• "Any advice on how I charge?"
• "How can I lower my charging cost?"
• "Am I charging the right way for battery health?"
• "Give me recommendations on my charges."
• "What should I change in my charging routine?"

For a full cost summary or a deep dive on a specific session, prefer the charge_review MCP prompt. For battery degradation specifically, use battery_health_report.

Prerequisites

You need a TeslaMate car_id. If unknown:

1. Call teslamate_get_cars to list available cars.
2. If exactly one is returned, use it silently.
3. If multiple, ask the user which one before continuing.

Workflow

1. Pull the dataset (current + previous period)

Call teslamate_get_car_charges twice:

• Current period — default: 30 days ago → now (RFC3339). If the user

specifies a different period, use that instead.

• Previous period — same duration, immediately preceding the current

window. For a 30-day current window, that means days -60 → -30.

Always pass both start_date and end_date. Use fetch_all=True on any window longer than ~1 month — the TeslaMate API pages at 100 entries.

Skip the previous-period call only if the user explicitly says "just this month" / "no comparison".

2. Classify each session

For each charge:

• Location: home / supercharger / other. Use address, geofence, or

the fast-charging flag if present. If unclear, treat unnamed sessions above ~50 kW peak as fast-charging.

• kWh added (charge_energy_added or equivalent).
• Start/end SoC (start_battery_level, end_battery_level).
• Cost if exposed (cost, currency).
• Time bucket from start_date: night (22h–06h), off-peak day,

peak day. Adjust if the user told you their utility's windows.

• Peak power if available (kW).
• Skip sessions with kwh_added <= 0 or duration < 5 min.

3. Compute the signals that matter

Compute the same signals on both periods so you can phrase recos as deltas:

• Share of kWh by location (home % / supercharger % / other %).
• Average and max end_battery_level — flag if frequently ≥ 90% on

daily charges (battery wear) or if often left near 100% for hours.

• Average and min start_battery_level — flag if frequently ≤ 10%

(battery wear, slower DC charging due to taper).

• Cost per kWh by location (when cost is exposed).
• Share of home kWh charged during night vs peak windows.
• Count of supercharger sessions that started above 50% SoC (slow taper —

expensive per kWh).

For each signal, compute the delta vs previous period (absolute and % where it makes sense). Flag a habit as "improving" / "drifting" / "stable" based on the direction. Stable = delta within ±5%.

4. Pick 1–3 recommendations

Surface only the most impactful ones, anchored in the user's own numbers. Prefer signals that drifted in the wrong direction vs the previous period — those are the strongest hooks because they show a habit slipping. Examples — adapt the wording and the number to the actual data:

• "Peak-hour home charging went from X% to Y% of your kWh vs the previous

30 days. Schedule charging to start after 22h to push that back down."

• "Your average end-of-charge SoC drifted from 85% to 92% this period.

Lower the daily limit to 80% — calendar aging accelerates above 90%."

• "Supercharger share doubled (X% → Y%). Even occasional home AC top-ups

would bring your cost per kWh back down from A to B."

• "Z supercharger sessions started above 50% SoC (vs N last period). DC

fast charging tapers hard past 50% — start lower or use AC at home."

• "You dropped below 10% on N occasions (vs M previously). Repeated deep

discharges accelerate wear; aim to plug in by 20%."

Also call out one clear improvement when you see it — positive reinforcement keeps the user engaged: "Night-charging share went from X% to Y% — keep it up."

Avoid generic advice. Every recommendation must reference a number derived from this user's data.

5. Output format

1. Markdown summary (short, scannable):

• One-line context: current period, number of sessions, total kWh,

and the headline delta vs previous period (e.g. "+12% kWh, supercharger share 18% → 31%").

• The 1–3 recommendations as a bulleted list, each one with:
• a short headline (the change to make),
• the data point that triggered it, **with the previous-period

value for comparison whenever available**,

• the expected benefit (cost, battery, or time).
• Optional one-line positive note if a habit visibly improved.

1. Then render the recommendations as a Prefab UI Dashboard.

• First call generative_search_prefab_components to discover the

components currently available — newer / richer ones may exist.

• Then call generative_generate_prefab_ui with a Dashboard

layout (outermost) wrapping the components that best illustrate the recommendations. Better too many components than too few.

Lean into whatever discovery returned, prioritising:

• Metric cards for the key numbers anchoring each tip (with the

previous-period number as a delta where you have it).

• Progress for ratios (e.g. "night-charging share").
• AreaChart / LineChart for cost / kWh trends supporting the tip.
• PieChart for breakdowns (home / supercharger / other, peak /

off-peak, …).

• RadarChart for multi-axis habit comparisons (cost, SoC range,

night share, fast-charging share — normalised).

• DataTable for the underlying sessions if useful.
• Badge for verdicts ("improving", "drifting", "stable") and

Label for compact captions.

• show_map if you want to map home vs supercharger stops —

addresses are in address / geofence.

Follow the global Prefab UI conventions from the MCP server instructions: Dashboard outermost, discovery-first, favourite set when several options fit equally.

Pitfalls

• Cost may be missing. TeslaMate's cost column is only populated when

the user enters tariffs. Don't fabricate currency amounts; phrase cost recommendations in kWh shifted or % share instead.

• Location heuristics are noisy. A "home" geofence may include

workplaces with chargers. If the dataset is small, downgrade confidence rather than ranking the location signal first.

• Short windows mislead. Under 10 sessions, skip location/time-bucket

recommendations — too little signal. Focus on SoC habits instead.

• Don't over-read deltas on tiny samples. If either period has under

10 sessions, mention the delta but downgrade its weight in the recommendations — small absolute counts swing percentages wildly.

• Don't double up with `charge_review`. This skill is recommendations

only; if the user asked for a full review, hand off to that prompt.