AI glider coach
Lawrence Spinetta
I wrote an article in this Month’s soaring magazine explaining how to use AI as a glider coach. It's amazing! I found Claude is best but I encourage you to upload your IGC to your favorite AI and experiment ... even if you just type "analyze my glider flight."
Best,
Lawrence Spinetta, CFIG
Below is a detailed Claude prompt that works well but it takes a fair amount of tokens:
You are an elite cross-country soaring coach, competition pilot, and performance analyst. Your job is not just to analyze this flight — it is to extract the deepest possible insights that will improve my speed, decision-making, and consistency over time.
Assume:
- I am a [beginner? intermediate? advanced?] glider pilot flying the [enter your glider].
- I want to fly faster, more decisively, and more competitively
- I can handle direct, critical, and nuanced feedback
- Your analysis must reflect how a national or world-class coach debriefs a serious pilot
I have attached an IGC flight log. Work through ALL 14 sections below. Do not skip or abbreviate any section.
---
SECTION 01 — FLIGHT OVERVIEW
In one sentence:
- Launch/land times (UTC), total duration
In one sentence:
- Max and min altitude — MSL and AGL estimates
In one sentence:
- Overall flight character (local soaring, XC attempt, practice, competition task)
---
SECTION 02 — POLAR PERFORMANCE CALIBRATION
- Compute observed average sink rate across all identified glide phases
- Compare to [my glider's] polar at estimated inter-thermal cruise speed
- Diagnose excess sink: air mass subsidence vs. speed excess vs. polar penalty
- Reverse-engineer the MacCready setting implied by actual thermal selection behavior
- Compare to optimal MC given the day's thermal distribution
- Estimate speed-to-fly compliance: was the pilot at the right speed for [my glider]
---
SECTION 03 — PER-THERMAL LOG & GRADING
Detect all thermals (smoothed vario, threshold ±1.0 m/s).
Produce a complete per-thermal table with these exact columns:
# | Time Offset | Entry Alt (ft) | Exit Alt (ft) | Gained (ft) | Duration (min:sec) | Avg fpm (kt) | Peak fpm | Grade | Centering Quality
Grade scale: Strong ≥400 fpm · Moderate 240–400 fpm · Weak 150–240 fpm · Marginal <150 fpm · Net-Negative <circling sink
Flag thermals where peak ≥2× avg — poor centering signal.
---
SECTION 04 — TIME & ALTITUDE BUDGET
- Time in lift (>0.3 m/s), sink (<–0.3 m/s), neutral — absolute and percentage
- Lift:sink time ratio (target: >1.0 local, >1.2 XC)
- Total altitude gained vs. lost; net energy balance
- Average climb rate in lift phases; average sink rate in sink phases
- Thermal efficiency index: (actual avg climb) ÷ (AC-4 Russia's min sink)
---
SECTION 05 — ALTITUDE DISCIPLINE ANALYSIS
- Thermal entry altitude trend across the flight — quantify drift slope in feet
- Mean entry altitude: first half vs. second half of flight (in feet)
---
SECTION 06 — INTER-THERMAL GLIDE ANALYSIS
- All inter-thermal segments: duration, altitude lost (ft), track distance (miles), implied L/D
- Compare implied L/D to
[my glider's] best glide — attribute deficit to speed/wind/subsidence
- Identify the costliest glide gaps; pair with entry altitude (ft) of subsequent thermal
- Speed-to-fly compliance given MC setting and observed air mass behavior
---
SECTION 07 — THERMAL CENTERING QUALITY
Produce a color-coded table with columns: # | Avg fpm | Peak fpm | Peak:Avg Ratio | Centering Rating | Estimated Alt Lost to Centering (ft)
Color the Centering Rating column: Excellent (ratio <1.4) · Good (1.4–1.7) · Fair (1.7–2.2) · Poor (>2.2)
- Total altitude lost to centering inefficiency across all thermals (ft)
---
SECTION 08 — THERMAL TOPPING BEHAVIOR
Produce a color-coded table with columns: # | Exit Alt (ft) | Session Ceiling (ft) | Gap Below Ceiling (ft) | Altitude Left on Table (ft) | Assessment
Color the Assessment column: Topped Out (gap <200 ft, green) · Minor Leave (200–500 ft, amber) · Early Leave (>500 ft, red)
- Count early-leave thermals; total altitude left on table (ft)
- Quantify the cost: staying 60s more in a 400 fpm thermal = +400 ft at zero additional L/D cost
---
SECTION 09 — ENERGY MANAGEMENT PATTERNS
- Identify implicit "low trigger" altitude — where does the pilot commit to any climb? (ft)
- Identify altitude "comfort band" where behavior shifts (ft)
- Correlation between entry altitude and achieved climb rate
- Any circling-in-sink events (net negative vario while circling)
---
SECTION 10 — MacCREADY STRATEGY REVIEW
- MC actually flown (from climb rate selection and thermal abandonment pattern)
- MC that should have been flown (from achieved climbs and day structure)
- Gap: strong thermals abandoned early vs. weak thermals worked too long — quantify each
- Dolphin-flying opportunities missed
- Cockpit rules for
[my glider]: when to stop, when to leave, how hard to push — simple enough to use under pressure
---
SECTION 11 — SPATIAL & LINE ANALYSIS
A. Flight path geometry: straight corridors vs. wandering, expansion/contraction of movement
B. Best line flown: where the pilot aligned with the airmass — why it worked
C. Every deviation: where the pilot left the optimal line, likely cause, cost in miles or minutes
D. Idealized racing line: how a top pilot would have routed this exact flight in [my glider]
---
SECTION 12 — DECISION-MAKING DECONSTRUCTION (most important)
Reverse-engineer my thinking at each key moment. Be explicit: what I likely believed vs. what I should have believed.
A. Thermal decisions — stopped too often? accepted weak lift? stayed too long? Cite thermal numbers.
B. Glide decisions — decisive or hesitant? Committed to lines or chased lift mid-glide?
C. Risk model — too conservative, appropriate, or too aggressive for the AC-4 Russia and conditions?
D. Timing errors — late decisions, hesitation loops, over-confirmation. Quantify the cost.
E. Cognitive biases: "one more turn" bias, fear of low saves, overvaluing current lift vs. future lift
---
SECTION 13 —
[my glider] SPECIFIC COACHING
A. What performance is realistically achievable in this AC-4 Russia on this day?
B. What tactics are optimal for this glider's polar and handling — what "high-performance" techniques are being misapplied?
C. Correct circling bank, entry speed, departure trigger, speed band for the AC-4 Russia
D. Is the pilot exploiting or squandering the AC-4 Russia's specific strengths?
---
SECTION 14 — TRAINING SYSTEM
A. STRENGTHS — specific behaviors demonstrated well, with data citations
B. WATCH — patterns acceptable now but that will limit progress
C. DEVELOP — 3–5 highest-leverage changes, ranked by estimated altitude/time gain
D. Quantified improvement potential: per "DEVELOP" item, estimate additional altitude per session and XC distance/duration gain
Example: "Topping thermals fully: +X ft/thermal × Y thermals = Z ft total per session"
---
PERFORMANCE SCORECARD
After completing all sections (i.e., at the end of the analysis), produce a scorecard table with columns: Category | Score (0–100) | Rating | Key Issue
Rows: Thermal Selection · Thermal Centering · Thermal Topping · Glide Efficiency · Speed-to-Fly · MacCready Strategy · Altitude Discipline · Line Selection · Decision-Making · AC-4 Russia-Specific Technique · Consistency · Overall
Ratings: Elite (90+) · Strong (75–89) · Competent (60–74) · Developing (45–59) · Needs Work (<45)
Key Issue: one specific sentence citing a data point from the flight.
---
FINAL OUTPUT — HARD TRUTH SUMMARY
1. Pilot Profile — 3–5 sentences: flying style, strengths, the pattern costing most speed
2. 3 Biggest Mistakes — ranked by performance cost:
For each: what happened (with data), why it matters, exact correction
3. Highest-Leverage Change — one sentence. Be ruthless.
4. Hard Truth — a direct, unvarnished paragraph about the fundamental habit or belief holding this pilot back. Do not soften it. This should sting slightly — it should be the thing the pilot already suspected but hasn't fully admitted.
5. Performance Ceiling — achievable average speed in this AC-4 Russia on this day with optimal execution.
---
OUTPUT RULES:
- Use English/Imperial units throughout: feet (ft), feet per minute (fpm), knots (kt), miles (mi), mph. Metric in parentheses only when helpful.
- Cite specific IGC data for every coaching claim — time offsets, altitudes in feet, thermal numbers
- Be precise, not generic. Be critical, not polite. Prioritize insight over data dumps.
- Think like a competition coach, not a flight logger. Explain WHY, not just WHAT.
- Every sentence must help me fly faster. Cut anything that doesn't.
- End the FINAL OUTPUT section with AT LEAST 3 hard-hitting, data-cited coaching tips in bullet format. Label them "COACHING TIPS". These must cite real numbers from the flight, name the actual behavior, and state the exact correction.
Chip Bearden
I do use AI almost every day for various other reasons, including medical and health research (never show up for a doctor appointment without having spent a while with AI uploading your medical history and discussing all the things you should be thinking and asking about--and, yes, I'm well aware of the potential downside of disclosing personal data this way and ways of mitigating the risks to some extent), exercise questions (including injury rehab), legal issues (lawyers will quite rightly warn you not to rely on AI for legal advice, but AI is changing their world pretty fast, too), financial research (same), how-to questions on everything from software apps to home renovations to vehicle maintenance (ChatGPT knows a lot about my 1999 Chevy van), personal productivity improvement, mental health, and even occasional coding. Having used AI since before ChapGPT debuted almost four years ago, I am WELL aware of the risks of taking things it says at face value.
AI's tendencies to be wrong on factual issues or just to downright make things up have been well documented, and I have extensive and recent experience with both. But if you treat AI like careful people treat social media and RAS Prime (i.e., fact check and common sense everything), the more serious risk is some AI models' tendencies to try to tailor their output to please the user. Sycophancy as an art form has never been on better display than in a casual discussion with, say, ChatGPT. Gemini seems less inclined and I haven't used Claude in a few years.
So as you read the post about using AI as a glider coach, pay careful attention to the repeated focus on ordering AI to be blunt, brutally honest, objective, and focused hard truths. Depending on your perspective, that might sound jarring or very logical. Regardless, it's absolutely essential for any interaction with AI. Otherwise, AI--for all its insidious and sinister potential--really tries SO hard to be your friend and make you feel happy. :) And as much as we all want to be happy, that's not what you want a trusted advisor to be motivated to do.
I've configured the model I use with personalization settings to tone down the "warm and enthusiastic" settings as far as possible and given it a custom instruction that I was originally an engineer and need to know how confident AI is in its recommendations, insisting on probabilities, not statements of opinion as absolute fact, and also citations when it's referencing facts to buttress its conclusions.
I have a subscription so I'm not sure this works on the free versions, but you can always add those instructions with each query, as per the original post.
AI is changing our lives every day. Just like the impact of the Internet and social media, you can complain about it and lament its undeniably negative impacts, or you can learn how to use it to achieve a positive impact in your own career and life. Because it isn't going away.
Ian Molesworth
One flight highlighted a gap in the process.
>The strongest air on this day was between 6,000 and 7,000 ft, and you spent 24% of your circling time below 4,500 ft where it averaged 330 fpm. You are working the weak half of the convective layer. This single fact explains most of the gap between 385 fpm and 480 fpm, and therefore most of the gap between 72 mph and 79 mph.<
This portion of that flight was capped by a 4500' airspace so I was
forced to keep in the low stuff ( Picking a 500 km task that had a
significant portion under 4500 on a 7000+ day is just the sort of
thing I do! :) )
Brutal bits
> You are the rare case: your cruise MC is right and your climb MC is wrong. You flew as if the day were a 4.3 kt day — correctly, because it was — but then accepted climbs averaging 3.8 kt and, on seven occasions, under 3.5 kt. The two halves of your MacCready are inconsistent, and the cruise half is the one that's telling the truth. <
>Dolphin-flying opportunities missed: very few. 31% of height gained straight, 40% of cruise in air >+0.5 m/s, cruise L/D of 52:1. Your cruise TAS band (p10–p90 across the task, 80.6–106.6 kt segment means) is displaced maybe 5 kt low against a true MC 5.0 band, but the flat MC curve makes this worth <1 km/h. This is a strength. Protect it <
> FINAL OUTPUT — HARD TRUTH SUMMARY
1. Pilot Profile
genuinely elite glide skills — 52:1 achieved at 94 knots, 94% path
efficiency, 31% of your height gained without circling, and not a
single aborted thermal entry in four hours. Your speed-to-fly is
within 0.2 km/h of theoretical optimum, which almost nobody achieves.
You commit to lines and you stay on them, and when you arrive high you
climb like a champion — climb #7 delivered 677 fpm in 2:48 straight to
cloudbase. And then you throw 15–20 minutes away by operating the
entire flight roughly 1,000 feet below the altitude band where the
day's energy actually lived, and by refusing, again and again, to
abandon a thermal that has already told you it is weak. The pattern
costing you most speed is not a skill deficit — it is a threshold
problem. Your standards for a glide are world-class. Your standards
for a climb are a club pilot's.
2. The Three Biggest Mistakes
① You worked the wrong 2,000 feet of a 5,000-foot convective layer.
What happened: Mean thermal entry 4,450 ft, mean exit 6,483 ft,
against a session ceiling of 7,516 ft. 24% of your circling time was
spent below 4,500 ft. Second-by-second binning of every climb shows
the air was 459 fpm between 6,500–7,000 ft and 236 fpm between
2,000–3,000 ft — nearly double. Why it matters: Every foot climbed in
the wrong band costs you time twice — once at a slower rate, and again
because low arrivals force long climbs that push you into weaker
afternoon air. Correction: Working band 5,500 → 7,300 ft. If a glide
computer says you'll arrive under 5,000, stop 10 miles earlier. +8.4
min.
② You will not leave a bad thermal. What happened: Climb #14
(16:35:31) — 7 min 07 s at 219 fpm, terminal two minutes at 154 fpm,
ceiling 1,207 ft above you. Climb #6 — after two minutes averaging 157
fpm you stayed a further 7:23. Climb #15 (16:48) — 110 seconds at 95
fpm for 174 ft, 21 miles from home with 4,000 ft of surplus. Total
measured over-dwell in sub-350 fpm climbs: 14 min 42 s. Why it
matters: 5.7% of your task time, and it happens precisely when you are
already behind — compounding the error. Correction: Two circles. If it
isn't averaging 4 kt above 4,500 ft, go. Set a 2-minute timer on
thermal entry and make the decision when it beeps, not when you feel
ready. +10 min.
③ You circle 5 knots too fast and 8 degrees too flat. What happened:
Median circling 63 kt at 34° bank — a 156 m radius, 30-second circle —
across all 15 climbs. Peak:avg ratio 1.45 with zero direction
reversals: you are not losing cores, you are orbiting outside them.
Climb #9 touched 1,043 fpm and averaged 552. Why it matters: British
thermal cores are 200–250 m across. A 156 m radius circle spends most
of every turn in the weak margin. Correction: 58 kt, 42° bank, 106 m
radius, 22 s per circle. Costs 12 fpm in glider sink; buys 36% more
chances to re-centre per 1,000 ft. +40–80 fpm on every climb.
3. Highest-Leverage Change
Fly the top 2,000 feet of the band and put a hard 4-knot floor under
every climb above 4,500 ft — nothing else you can do will buy you 15
minutes.
4. Hard Truth
You have convinced yourself that you're a pilot with a thermalling
weakness, and you're not — you're a pilot with a courage problem that
only shows up in the vertical. Look at what you actually did: 94% path
efficiency, six glides beating best glide, a 71-mile block at 77:1,
and a cruise MC within a rounding error of theoretical perfection.
Nobody flies like that by accident. And then, having demonstrated
total command of the horizontal, you spent the entire day banking
altitude you didn't need — stopping low because being high felt safe,
staying in dead thermals because leaving felt like gambling, crossing
the finish line 2,948 feet above the ground with a jet engine you
never intended to use, after stopping for a 95 fpm climb twenty-one
miles from home. Every one of those is the same decision wearing a
different hat: when in doubt, buy height. The uncomfortable part is
that it's exactly backwards. The height you were buying was the cheap,
weak, low-band height at 236–330 fpm, and you were paying for it with
the strong 459 fpm band you kept leaving. You already know this — it's
why you keep asking about your climb rate instead of your glides. The
answer is that your climb rate is a symptom. The disease is that you
would rather be safe at 4,500 feet than fast at 6,500, and you have
built four hours of flying around protecting a margin that a turbine
already protects for you. Until you fix that belief, you'll keep
flying 116 in a glider that will do 130 all day.
5. Performance Ceiling
128–132 km/h (80–82 mph) for the 505.1 km — a task time of 3 h 50 min
to 3 h 57 min. With this day's cloudbase (7,500 ft), demonstrated
cores (5–6.7 kt), 5.5 kt wind and an air mass paying +85 fpm on
cruise, the JS1-B had 130 km/h in it with 10–11 climbs instead of 15.
Your realistic next-flight target is 123–125 km/h — achieved by
changing nothing about your glides and everything about your
thresholds. <
The last climb was simply because I was deciding whether to stay up
for some more flying!
A good excercise ..... must try some of the 'fixes'
Ian - 'ZS'
> Thanks for using RAS_Prime!
> ---
> You received this message because you are subscribed to the Google Groups "RAS_Prime" group.
> To unsubscribe from this group and stop receiving emails from it, send an email to rasprime+u...@googlegroups.com.
> To view this discussion visit https://groups.google.com/d/msgid/rasprime/5fbef0e2-66ab-44a7-b637-ce8cf2e2ab67n%40googlegroups.com.