estimate-skill-strength

name	estimate-skill-strength
description	Estimate a 战法's per-round strength and write *Estimate fields onto its entry in web/public/game-data/database.json, organised into the categories damage/healing/attribute/damageBoost/damageReduction/damageDealtReduction/damageTakenIncrease/evasion/lifesteal/crit/critDamage. Each estimate is a rough per-round coefficient (max-level, average-trigger) so the team-builder prompt can compare skills. Use when the user asks to "estimate" / "估算" / "evaluate the strength of" a named 战法.
allowed-tools	["open_files","expand_code_chunks","grep","find_and_replace_code","bash"]

Estimate Skill Strength (战法预估)

Use this skill when the user asks to estimate / 估算 / evaluate the per-round strength of one or more named 战法 (skills). The output is one or more *Estimate numeric fields written onto that skill's entry in web/public/game-data/database.json, which the team-builder prompt then surfaces so the LLM can weigh skill output.

Goal

For a given skill, compute a rough per-round coefficient (a percentage number, no % sign in the value) for each relevant category, using:

• the skill's desc in web/public/game-data/database.json (the source of truth — always re-read it, descriptions change), and
• web/public/game-data/formula.md (the damage formula) when a modifier like 无视减伤 / 会心 needs converting into a multiplier.

The number is a comparison metric, not an exact 兵力 figure — it deliberately omits the absolute 固定值/固定系数 constants and the attacker/defender stat line.

The estimate categories (and their database keys)

Category	Database key	Label in prompt	Meaning
Damage	damageEstimate	伤害	Direct output (谋略/兵刃 coefficient)
Healing	healingEstimate	治疗	回复兵力 (治疗率-based)
Attribute	attributeEstimate	属性	属性增减幅度, in points (武/智/统/先攻…)
Damage boost	damageBoostEstimate	增伤	造成伤害提升 %
Damage reduction	damageReductionEstimate	减伤	受到伤害降低 %
Damage-dealt reduction	damageDealtReductionEstimate	降伤	Enemy 造成伤害降低 % (distinct from 减伤/受到伤害降低)
Damage-taken increase	damageTakenIncreaseEstimate	易伤	Enemy 受到伤害提升 % (易伤; debuff that amplifies damage taken)
Evasion	evasionEstimate	闪避	规避率 %
Lifesteal	lifestealEstimate	攻心	Heal self as a % of 谋略 damage dealt
Crit rate	critEstimate	奇谋率	奇谋 (谋略-crit) trigger-rate boost %
Crit damage	critDamageEstimate	奇谋伤害	奇谋 (谋略-crit) damage boost %

A skill can carry several of these at once (e.g. a buff that gives 增伤 + 减伤 + 属性). Only write the categories the skill actually has. Skills that deal no direct damage get no damageEstimate.

If you introduce a new *Estimate category that isn't in the table above, you MUST also update web/src/services/promptGenerator.ts — see "Keeping the prompt in sync" below — otherwise the new field will silently never render in the copied prompts.

Core formula

The per-round estimate for a damage-style category is:

estimate = trigger_prob  ×  Σ(coefficient × targets)  ×  any_multiplier

• trigger_prob — the skill's prob as a fraction (e.g. 60 → 0.6). For 指挥/被动 that fire every round at 100%, use 1.0; if such a skill has an internal "X%概率" gate, use that gate instead.
• targets — 单体 = 1, 随机两人 = 2, 全体 = 3. Multiply by N for "施放N次 / N times" skills. For "随机2-3人" use the midpoint 2.5.
• coefficient — always use the upper / max value (descriptions in this DB are already collapsed to the upper bound). A 220% 谋略 hit contributes 2.20.
• any_multiplier — conditional bonuses, crit, 无视减伤, etc. (see below).

Conventions (decided with the user — keep consistent)

• Use max coefficients. Descriptions are already collapsed to the upper bound; always estimate at max level.
• Dual-type "兵刃和谋略伤害" = two hits. "100%兵刃和谋略伤害" deals both a 100% 兵刃 and a 100% 谋略 hit → 200% per target. Do not count it once.
• Averaging window = rounds 1–5. For ramping / accumulating-DoT skills, average the per-round value over rounds 1–5 (matches 火烧连营=691, 焰燎江天=709). Use this so all skills are comparable.
• "Average in one round" default = the rounds 1–5 per-round average with the trigger chance folded in.
• Sustained-bonus default. If a skill applies a state that buffs its own repeat casts (e.g. "若目标已持有妖术则伤害+35%"), assume the bonus is active (sustained case), since the skill self-applies the state and the state outlasts a turn.
• 会心 / crit (会心率 r): crit ≈ ×1.5 damage (web/public/game-data/formula.md: 触发时通常 ≥150%). Expected multiplier = 1 + r×0.5 (e.g. 25% 会心 → ×1.125).
• 无视40%减伤: a normal hit would be ×(1−0.40)=×0.60; ignoring it means ÷0.60 = ×1.667. In general 无视 m% 减伤 → multiplier 1/(1−m).
• 准备1回合 (charge): the skill fires only every other round → divide the estimate by 2.
• Intermittent fixed-round skills (e.g. "战斗第3,5回合"): these are special and non-comparable — do NOT write an estimate (we removed 陷阵蹈难's). Tell the user it's intermittent.
• Reactive skills (trigger on dealing/receiving damage or on dodge, often with a per-round cap and diminishing coefficients): the value depends on combat activity. Ask the user how many triggers/round to assume (we've used 2.5–3.5). Sum the diminishing coefficients across the assumed trigger count.
• Buff/debuff coverage & uptime. Scale by how many heroes are affected and the uptime fraction:
  • Allied buff to "self + 1 teammate" → covers 2 of 3 heroes. Early conventions used a × 2/3 factor; the later, preferred convention is the explicit form base × heroesAffected × upRounds/8 (e.g. 每战先登: 14 × 2 × 8.6/8 ≈ 30.1). When in doubt, ask which basis the user wants and stay consistent within a batch.
  • Front-row-only bonuses: front row is typically 2 of 3 heroes unless told otherwise.
• 属性 sign: record attribute debuffs as positive magnitude (the size of the attribute swing generated), same as buffs.
• 属性 stacking with 2-turn expiry. A debuff/buff that adds 1 stack/round, lasts 2 turns, and is re-applied each round saturates at ~2 × prob active stacks (NOT the nominal cap), because anything older than 2 turns expires. e.g. 三军夺气 at 55%/round → ~1.1 avg stacks. Only assume the cap if the state genuinely never expires.
• Rounding: integers for damage/healing/attribute; one decimal is fine for small buff %s.

Modifiers cheat-sheet (web/public/game-data/formula.md)

Modifier in desc	How to fold in
会心率 r	× (1 + 0.5r)
无视 m% 减伤	× 1/(1−m)
准备1回合	÷ 2
X%概率额外 +Y%	× (1 + X·Y)
风暴/状态-conditional +Z%	include if state is reliably present (sustained), else ask

Multi-component skills

Decompose the desc into independent effects and map each to a category. Skip components you cannot quantify cleanly (e.g. a conditional crit+lifesteal rider on the carrier's normal attacks whose base value isn't in the skill text) — note them to the user rather than guessing. Example: 南疆烈刃 → damageEstimate (全体 90% × 3) + attributeEstimate (夺取 20武+20统 × 2 enemies), with the conditional 40%会心+25%lifesteal rider left un-estimated.

Workflow

1. Re-read the skill's desc from web/public/game-data/database.json (don't trust earlier memory — descriptions get edited). Locate it with grep then read the entry with a small python3 -c "import json…" print.
2. Decompose the desc into category components.
3. Compute each estimate with the core formula + conventions above. Show the user the breakdown (a short table) and the assumptions used.
4. Ask for any genuinely ambiguous knob (trigger count for reactive skills, coverage basis, which of several interpretations) before writing — don't silently pick.
5. Write the field(s) onto the skill entry via a python3 script that rewrites the JSON (json.dump(..., ensure_ascii=False, indent=2) + trailing newline), inserting the estimate(s) right after desc. Re-validate the JSON (json.load).
6. If a new category was introduced → update promptGenerator.ts ("Keeping the prompt in sync").
7. Report the value(s) and offer to commit & push (only commit on explicit user go-ahead).

Keeping the prompt in sync (new categories only)

The prompt rendering lives in web/src/services/promptGenerator.ts. For the existing categories nothing more is needed. If you add a brand-new *Estimate key, update all three of:

1. SKILL_ESTIMATES — the shared [key, label] array (used by both formatSkillInfo and formatSkillInfoEstimates). Add ['newEstimate', '中文标签'].
2. PROMPT_INSTRUCTIONS — extend the 战法说明：… legend line with 标签=含义.
3. The two 战法预估（…） priority lines (one in generateLLMPrompt, one in generateTeamBuilderPrompt) — add the new label to the parenthesised list.

Then verify with a quick Node check that the new label renders, e.g.:

cd web && node -e "/* mini formatSkillInfo replica printing the target skill */"

This wires the new category into both the draft and team-builder copied web-LLM prompts. Also update this file's categories table/frontmatter so future estimates stay consistent. Ordinary per-skill estimates only touch web/public/game-data/database.json.

Worked examples (for calibration)

• 火烧连营 (60% 主动, 220% 谋略 全体, 焚烧 DoT 60%/层 stacking, wipe after 2 consecutive misses): Model B simulation, avg rounds 1–5, ×3 enemies → damageEstimate: 691.
• 烈火张天 (50%, 全体, 90%谋+90%兵): 0.5 × (0.90+0.90) × 3 = 270.
• 制霸江东 (65%, 随机两人 250% 兵刃, ignore healing): 0.65 × 2.50 × 2 = 325.
• 威震塞外 (50% 追击, 随机两人 140% 兵刃和谋略): dual-type → 0.5 × (1.40+1.40) × 2 = 280.
• 伏兵四起 (50%, 140% 兵刃 ×4 casts, +25% 会心): 0.5 × 1.40 × 4 × 1.125 = 315.
• 智破千军 (50%, 随机两人 180% 谋, 35%→+20%): 0.5 × 1.80 × (1+0.35×0.20) × 2 = 193.
• 上兵伐谋 (100% 指挥, 单体 120% 谋, +10%/round ramp): avg rounds 1–5 → 144.
• 每战先登 (指挥, 先登 to self+1: +30先攻, +14%增伤, −14%减伤): attributeEstimate 64.5, damageBoostEstimate 30.1, damageReductionEstimate 30.1 (all base × 2 × 8.6/8).
• 三军夺气 (55% 追击, −30 武/智/统 ×2-turn): ~1.1 avg stacks × 90 pts → attributeEstimate: 99.
• 同舟共济 (100% 指挥, 治疗率90% to self+1): 90 × 2 = healingEstimate 180.
• 草船借箭 (指挥; 攻心+24%; 50% reactive 80% 谋, ~2.5 triggers): damageEstimate 200, lifestealEstimate 24.
• 七进七出 (被动, +35%规避, 龙胆 90%→diminishing ×2 targets, assume 3.5 triggers): damageEstimate 540 (90%+80%+70% + 0.5×60%, ×2), evasionEstimate 35.
• 悲愤诗 (65% 主动, 全体 治疗率120% + 前排 +50%, assume 2 front-row): 0.65 × (120×3 + 50×2) = healingEstimate 299.
• 未雨绸缪 (60% 主动, 随机两人 140% 谋; 奇谋几率+奇谋伤害 +8%, +2%/功能性增益, 上限+8次): damageEstimate 168 (0.60 × 1.40 × 2); 奇谋 buff recorded as sustained magnitude (8% base + ~2 增益 × 2%) → critEstimate 12, critDamageEstimate 12. The 获得1层抵御 single-hit block is left un-estimated (no clean per-round %).