Gamebook screenshot analysis

GameBook Screenshot Analysis

Turn the user's Golf GameBook screenshots into a structured on-course round the coach can use. GameBook has no export/API, but the data is on the user's phone as screenshots — read them directly.

The one rule that governs everything

GameBook reliably records only score-per-hole. Putts, fairways, GIR, chips, bunkers, scrambling, up-and-down and sand saves are only as complete as what the golfer tapped in during the round. Never analyze a dimension that wasn't really tracked. A "0.0%" sand-save/up-and-down/scrambling with no supporting hole data means "not entered," not a genuine zero. Flag coverage; do not invent.

Run this off the main thread

This reads several images (large tokens). In Claude Code, the main agent should dispatch ONE fresh subagent (Task/Agent tool, general-purpose) whose prompt is: "Follow the gamebook-screenshot-analysis skill end to end on these image paths: <paths> and return only the final summary." The subagent does the work and returns just the summary. If you are that dispatched worker, proceed.

Reading the screens (legend)

• Header: the big number is gross (it may sit partly behind the "Net score" label — it's still the round total; confirm it equals Out+In). "Net score" + a coloured +/- chip = net and net-to-par. "Position x/y" = group finish.
• Scorecard rows: HCP, Par, Score, Net, then Putts, Fairways, GIR, Bunkers, Chips, Penalties, with an Out/In total column. A blank cell means not entered → null, never 0.
• Fairway icons: ✓ = hit, → = missed right, ← = missed left; blank on a par 3 = na. The Out/In fairway total (e.g. 3/5) counts only tracked driving holes.
• GIR icons: ✓ = green hit; an arrow (↑ ↓ ← →) = missed that way (still a miss). If the per-hole card and the Statistics dial disagree, trust the card (this round: card 1/18 vs dial 8%).
• Statistics → Scores gives bogey/double/worse counts and par-type averages — use them to cross-check the per-hole read, not as the source.
• The 0.0% trap: Scrambling / Up-and-Down / Sand save at 0.0% with no hole-level chip/bunker data means not tracked → coverage none, not a real 0.

Workflow

1. Group the images into one round. A round is usually 2 scorecard halves (front/back) plus optional Statistics pages. The scorecard is the source of truth; Statistics pages are used only to validate and to read miss-direction splits. If only the scorecard is provided, that's fine — you still get the reliable scoring dimension.

2. Extract per hole from the scorecard: par, score, and when present putts, fairway (hit/miss_left/miss_right/na on par 3s), gir (bool), bunkers, chips, penalties. Use null for any hole where a value isn't shown. Read the course par, CR, slope, and the header gross/net/position.

3. Assign coverage per dimension (full/partial/none): full if tracked on ~90%+ of eligible holes (fairways exclude par 3s), partial if some, none if zero. Any Statistics-page rate (sand save, up-and-down, scrambling) with no hole-level backing is none. If a Statistics number contradicts the scorecard, keep the scorecard and add a notes entry.

4. Self-check before saving. Confirm hole scores sum to gross, hole pars sum to course par, and there are 9 or 18 holes. If anything fails, show the user the holes you're unsure about and fix the read — do not save a wrong round.

5. Save by calling gamebook(action="save", round=<record>). The record shape is in the tool docs; the tool computes the scoring block and stores it (rolling last 5). If it returns saved: false, resolve the problems and retry.

6. Report progress. Call gamebook(action="compare"). Summarize the latest round and its direction of travel vs prior rounds — confidently on scoring, and on any other dimension only where comparable is true — then hand off to golf-coaching for the practice prescription.

Output format (this skill's return value)

## GameBook round — <date> (<course par>, gross <gross> / net <net>)

<one-line headline: to-par and the scoring shape>

**Scoring (reliable):** +<to_par>; <bogey>/<double>/<triple_plus> spread;
par-3 <+x.xx>, par-4 <+x.xx>, par-5 <+x.xx>.

**Tracked this round:** putts <coverage>, fairways <coverage>, GIR <coverage>,
short game <coverage>. <one line naming what's too sparse to judge>

**Progress vs last <n> rounds:** <to-par direction + par-type direction; putts
direction only if comparable; "accuracy not comparable — not tracked in enough
rounds" otherwise>

<Notes: any scorecard-vs-stats discrepancies>

Keep it factual. No drills here — golf-coaching prescribes from this.