AI prompt tips

Every Playloop AI summary starts with your game's metadata and a free-form prompt you write. Get the prompt right and the AI reads your numbers the way a human analyst would. Get it wrong and you'll chase phantom bugs (heavy tails read as overflow) or get platitudes back. This page covers what makes a good prompt and gives genre-specific examples to copy from.

What the AI sees on every summary

Before your prompt reaches the model, Playloop prepends up to three anchors from the game settings, in this order:

Genre

The primary genre you picked + any subgenres (Steam-style). e.g. "Indie · Idle, Clicker". Anchors the AI to a coarse game shape.

Game context

Your free-form prompt, up to 5,000 chars. This is where mechanics, demo scope, currencies, and "what looks like friction but isn't" go.

Priorities (KPIs)

When you've picked event-tag priorities under Analysis tuning, three lines render here: Primary (KPIs the AI must lead with), Secondary (treat as supporting evidence), and Ignore (skip even if they cluster). Empty when you haven't tagged any events.

After that comes a digest of your build: cohort stats, session-end milestones, named-action conversion rates, top friction / praise insight clusters, and an event totals table. The AI is instructed to defer to YOUR prompt for what counts as normal, so if you say "soul counts in the billions are expected," it won't flag heavy tails as overflow bugs.

What makes a good prompt

Six things. None are required. The more you include, the sharper the summary.

1. Core loop in one sentence. What is the player doing 80% of the time? Saves the AI from inferring from event names.
2. Demo scope. Where it starts, where it ends, how long it should take. Lets the AI tell "didn't complete the demo" apart from "finished as designed."
3. Currencies and how each is earned. "X comes from Y; Z is milestone-only." Without this the AI treats every counter as a generic numeric stat.
4. Progression scale. "Late-game numbers look like ___." Saves you from chasing "overflow bugs" that are just exponential progression working as intended.
5. What looks like friction but isn't. Backtracking in metroidvanias. Dying in roguelikes. Slow openings in idle clickers. Optional side content. Call these out so the AI doesn't flag them.
6. Onboarding gates. "If [milestone X] isn't reached, players bounce." Lets the AI rank drop patterns by what actually matters.

Examples by genre

Copy any of these into your game's AI prompt and tweak. The genre prefix shows the corresponding Genre / subgenre selection in the game settings.

Necromancer's Army is an idle-clicker about grief expressed through dark progression. The player is a necromancer raising an army to find their lost son. Demo ends after the Void Pact resolves (Act 2).

Souls is the only currency. Dark Essence is milestone-granted plus Dark Harvest sacrifices (no passive accrual). The minion cap starts at 20 and rises to 30 after the Void Pact in Act 2.

Progression is exponential, late-game soul counts in the billions are expected, not a bug. Average session is 30–45 minutes; demo completion target ~50%. The first rite ("twitching fingers") is the onboarding gate; if first_rite=none exceeds 10%, that's a real cliff.

Spore Knight is a roguelike. Each run lasts 20–40 minutes; death is part of the loop, players are expected to die 5–10 times before their first win. Permanent progression (between runs) is gated by "spore tokens" earned per floor cleared.

Currencies: spore tokens (persistent) and gold (per-run, lost on death). Demo includes 3 of 8 biomes.

Heavy variance in run length is normal; failures are content, not friction. Target win rate per run: ~25% after the meta-progression curve unlocks. Don't flag death frequency as a problem unless first-time-win rate is under 50% after 5+ runs.

Lighthouse Letters is a romance visual novel. Four endings across seven chapters; ~3 hours per playthrough, ~10–12 hours for full completion across all routes. 14 choice points per route.

No fail state, every save reaches an ending. Demo covers Chapters 1–2 ending on a hard choice.

Telemetry signal we care about: route splits, time-per-chapter, and which choices players agonize over (long deliberation pauses). "Time stalled" on a choice node is engagement, not friction.

Cave of the Forgotten is a 2D metroidvania. 8–12 hour main campaign with four movement upgrades that gate map areas. Combat is parry-driven.

Demo covers the first 90 minutes through the "Wind Boots" upgrade. Expected progression: linear up to first upgrade, then exploration branches.

Backtracking is intended and should not be flagged as friction. Average tester finishes the demo in 60–100 minutes; outliers at 3+ hours are exploration enjoyers, not stuck players.

Common mistakes

• Marketing copy. "An emotional journey through a haunted forest." This tells the AI nothing about your data. Save the pitch for your store page.
• No scale anchors. If you don't say "late-game numbers are expected to reach X," the AI may flag your heavy-tailed counters as overflow bugs. Idle, roguelike, and grindy RPG telemetry especially need this line.
• Listing every event name. The AI already sees event names in the digest. Tell it what events mean, not that they exist. "consecration_accepted is the demo's climax milestone" beats a 30-line event glossary.
• Ambiguous "good" thresholds. "Most players should reach Act 2" is weaker than "Reaching Act 2 is the demo's success milestone, target 60%+ on a healthy build." Give the AI a number to compare against.
• Forgetting demo scope. Without "demo ends after Y," every dropout looks like churn. State the demo's end state so the AI can compute completion correctly.

Where to edit

Your game's prompt lives at /games/[slug]/settings → AI context. Genre + subgenres live on the same page under Details. New games get genre as a required field on the create form; the AI prompt is optional but high-leverage, capped at 5,000 characters.

Once you save, the new prompt applies to future AI summaries, clear the cached summary on a build (via regenerate) to see the new prompt in action.