From data to narrative

The translation problem

You have done good work on the data. Now the harder problem starts. A coach wants two things in 30 seconds. A scout wants comparisons across competitors. A broadcaster wants a story they can tell in a half-time graphic. A board wants the risk in one paragraph and the headline number in big text. The same analysis, four audiences, four registers, four sets of vocabulary, four standards of acceptable simplification.

This is the layer where most analysts spend the most time and get the least credit. It is also the layer that AI handles meaningfully well — provided you keep your hands on the wheel.

The four audiences, in detail

The coach

Wants: two or three sentences that change a selection or a tactical decision before kick-off. Has thirty seconds. Knows the sport deeply. Has zero patience for jargon they have to translate. Cares more about confidence levels than you might think — a coach reads "this is probably true" very differently from "this is definitely true".

Tone: direct, sport-native, no hedging spirals. Acceptable: "Their right-back gets dragged 10m higher when they're chasing. Target the space behind." Unacceptable: "The opposition's right defender's positional data suggests a possible vertical displacement under negative-state conditions."

The scout

Wants: a comparison. Player X versus the alternatives in the same role, at the same price, in similar systems. The scout's mental model is league-wide and multi-season; your analysis has to slot into that.

Tone: cool, comparative, ratio-driven. Numbers are useful here. Caveats about sample size, level of league, system fit are mandatory.

The broadcaster

Wants: a story that survives a graphic. Three numbers, one storyline, one image-friendly comparison. Has to be true; also has to be interesting.

Tone: narrative, accessible, simplified — but not wrong. Watch for the bot's instinct to over-claim here. The best broadcasters are precise; the bot, given a broadcast brief, will drift towards superlatives unless you stop it.

The board

Wants: the risk and the upside in one paragraph. Will skim. Reads in bullet form, in big-font headlines, and in the first sentence of any paragraph. Wants to know what a decision will cost and what it will return.

Tone: executive, neutral, conclusion-first. The structure is upside-then-risk-then-recommendation, not the other way around. Numbers in pounds or euros, not in xT.

One analysis, four briefings

Take the worked example from Lesson 3 — Team A's match. We saw four things: created more chances, pressed harder for an hour, made a tactical switch in the second half around the substitution, and the right-back was the player to watch. Now we translate this for each audience.

Team A press hard for the first hour, then drop into a mid-block — usually around the 60th minute. Their right-back gets pulled high in the second half; the space behind him is exploitable. Set our left winger to attack that channel after the hour.

Team A's right-back (Player 11) is operating high in possession (heat map cluster 10–15m above the defensive line in the second half) but loses possession in his own half at a rate that's above average for a starting Championship right-back (7 turnovers in this match, 3 leading to opposition shots). The role is closer to a wing-back than a traditional full-back. Sample is one match — would need 5–10 to confirm.

Story: Team A is a first-hour team. They press, they create, they tire.

— 14 shots vs 7 in this match
— 9.2 PPDA in the first half (intense press) → 14.6 in the second (mid-block)
— Their last 5 goals have come before the 65th minute

Visual: a heat map split by half, showing the higher defensive line in H1 and the deeper line in H2.

Recommendation: Team A's performance pattern is consistent — strong first hour, fade in the second. Their right-back is the highest-upside player to target in any swap or signing decision involving that role, but a one-match read is not enough to commit to a transfer position. We recommend tracking him across the next 8 matches; estimated cost to do that £4,000 in additional scouting time and external data. Update at the end of November.

Notice that the same underlying facts produce four very different documents. The coach briefing is three sentences and an instruction. The scout briefing has a sample-size caveat. The broadcaster briefing leads with a storyline; the numbers serve the story. The board briefing leads with a recommendation and ends with a cost.

How to ask the bot to translate well

Three rules of thumb make the bot a better translation partner.

Name the audience and the brief in the same sentence. Not "rewrite this for a coach". Say: "Rewrite this as three sentences for a head coach who has 30 seconds before kick-off. Sport-native vocabulary, no caveats, one tactical instruction at the end."

Show, don't just tell. If you have an example of the format a particular audience expects — a previous match-prep note, an old scouting one-pager — paste it in alongside the new analysis. The bot will copy the structure and register.

Always do the technical version first, then the audience versions. Ask the bot for the analyst-level summary, check it for accuracy, then translate from that. If you ask the bot to translate directly to a broadcaster brief, the loss of nuance happens in one step and is hard to spot. With an intermediate analyst-level version, you have a known-good reference to compare back against.

The three failure modes of AI-written sport copy

False certainty. The bot will say "Team A is a first-hour team" without saying "based on a five-match sample". The caveat gets dropped in translation. Add to your system prompt: "Carry sample-size caveats through every translation. If a stat was based on one match in the source material, it must say so in any audience version."

Broadcast-cliché drift. Ask the bot to write for a broadcaster and it pulls towards Sky Sports voice — "phenomenal pressing intensity", "world-class second-half adjustment". The broadcasters you actually respect are precise. Add: "For broadcast briefings, use specific verbs and concrete numbers. Avoid superlatives unless the data supports them."

Lost technical nuance. Ask the bot to write for a coach and it will simplify a tactical observation to the point of being wrong. "Their right-back gets dragged high" is a simplification of "their right-back's defensive line averages 10–15m higher in the second half versus the first when their team is behind". The first might be wrong (he gets pulled high all the time, not just when behind). Have the bot write a one-line "what was simplified out" alongside each audience version. Lossy compression is fine if you know what was lost.

Looking ahead

Lesson 5 is the wrap. We take the work we have done — system prompt, real data, four-audience translation — and put a use policy around it. We also look at where this is going next: video, computer vision, biometrics, multimodal. And we say honestly what AI is not going to replace in sport analytics any time soon.