Read candidates.md → sort by composite → output top N recommendations, each with scoring details + anchor comparison + recommendation reason.
Overview
[用户:推荐选题]
↓
[Phase 0: 检查 candidates.md 存在性] ← 不存在则引导,不报错
↓
[Phase 1: 解析 candidates 列表]
↓
[Phase 2: 过滤(tier / 安全性 / 已发过)]
↓
[Phase 3: 排序 by composite + 找锚点]
↓
[Phase 4: 输出 top N + 每条的 rationale + 锚点对比]
Constants
- TOP_N = 5 — by default recommend top 5
- STRATEGY = stable+experimental — when recommending ≥2, follow cadence-protocol.md 's "1 safe-score + 1 experimental" strategy; when recommending 1, only recommend the top safe-score
- POOL_PATH = candidates.md — candidate pool path
- EXCLUDE_PUBLISHED = true — exclude already published (deduplicate with
predictions/*.md) - EXCLUDE_REJECTED = true — exclude those the user actively skipped (
tier=skip) - REQUIRE_SCORED = true — only recommend already scored ones — avoid recommending material not yet read
- DUPLICATE_CATEGORY_LOOKBACK — derived from
state.target_publish_cadence_days: within max(3, cadence_days × 3) days, do not recommend candidates of the same category that have already been published (avoid aesthetic fatigue)
💡 Override when calling:
/cheat-recommend — top: 3 — filter: safe
Inputs
| required | source |
|---|---|
candidates.md |
user project root |
predictions/*.md |
used for deduplication |
.cheat-state.json |
current rubric_version |
Workflow
Phase 0: candidate pool existence check
Read candidates.md:
| status | handling |
|---|---|
| file does not exist | do not error. Output guidance: see "no candidate pool guidance" below |
| file exists but is empty (< 1 entry) | same as above |
| file exists and is non-empty | proceed to Phase 1 |
no candidate pool guidance(core: do not let the user be discouraged the first time they encounter cheat-recommend):
你目前没有候选池(candidates.md 不存在或为空)。
绝大部分人没有候选池——这很正常。四个建立方式,挑一个:
1. 🌱 [推荐] 跑 /cheat-seed
一次性的种子动作:3 个问题(兴趣 / 调性 / 红线)→ 拉公开热点 + Claude brainstorm
→ 输出 15 候选让你挑 5 → 默认顺带写 5 个 draft。5 分钟搞定。
- 没发过历史的:纯 brainstorm(兴趣 × 热点)
- 发过历史的(init 时已 import):brainstorm 会基于"你过去做过什么"给推荐
说:"找选题" 或 "seed"
2. 🔥 [日常补充] 用 /cheat-trends 抓 20 条带打分的候选
说:"抓热点" — 从 weibo-hot / zhihu-hot / b站热门 / HN / 你配的源各拉 N 条
适合已经跑过 /cheat-seed、想日常补充候选池的用户
3. ✍️ 手动建:把候选标题贴进 candidates.md,每行一条
我会自动给每条粗打分
4. 📋 从 Notion / RSS 导入:跑 /cheat-init --mode add-pool 配置 adapter
你也可以跳过候选池,直接给我具体稿子说"启动预测"。
> /cheat-seed vs /cheat-trends 的区别:
> - seed 是种子动作(含 brainstorm + 可选 draft),适合"我从零开始没选题"
> - trends 是日常多 adapter 抓取(不 brainstorm 不写 draft),适合"日常补充候选池"
complete guidance → exit, do not continue to subsequent phases.
Phase 1: parse candidates
According to candidate-schema.md 's "Markdown representation" format, parse each H3 entry:
### [tier1] 标题
- **id**: a3f2c1d4e5b6
- **composite (v2)**: 8.47 — ER=4 HP=4 QL=5 NA=3 AB=5 SR=3 SAT=3
- **predicted bucket**: 5-30w
...
extract each entry's id / title / tier / composite / dimension_scores / note。
fault tolerance:candidates.md format has been manually modified by the user → ask the user for the schema,do not silently ignore unrecognized entries。
Phase 2: filter
1. EXCLUDE_PUBLISHED=true → 扫 predictions/*.md 的 header,提取所有 id;从候选池过滤掉
2. EXCLUDE_REJECTED=true → 过滤 tier=skip
3. REQUIRE_SCORED=true → 过滤 composite=null(未打分的不推荐)
4. filter 参数:
- tier1: 只保留 tier=tier1
- all: 不过滤(tier1+2+3)
- safe: 排除 tier=risky
- risky: 仅显示 tier=risky(用于"我今天就想发风险议题")
Phase 2.5: Buffer color override (highest priority)
Read state.shoots + state.target_publish_cadence_days to compute buffer color (cadence-protocol.md):
| Buffer color | recommendation strategy override |
|---|---|
| 🔴 Red | only recommend top 1 safe-score— do not recommend experimental. Reply: "buffer is already 0/1 posts, the risk of missing an update on the next publish day is high, today you must shoot ≥1 safe-score. Below is the top 1 safe-score (no experimental recommendations)" |
| 🟠 Orange | standard 1 safe + 1 experimental, but prompt "recommend prioritizing shooting the safe-score" |
| 🟢 Green | standard 1+1 (default) |
| 🔵 Blue | refuse to recommend. Reply: "your buffer already has N posts, cadence-protocol specifies pausing shooting when there is a backlog. Publish the stock + review first. To manually override, say 'I just want to shoot'" |
flexible mode (target_publish_cadence_days=null) |
do not apply buffer override, standard strategy |
Phase 3: sort + select 1 safe + 1 experimental (by STRATEGY)
1st entry (safe-score)
- sort by
compositein descending order - filter out
tier=risky(safe-score requires safe topics) - Filter out
categoryand the most recentDUPLICATE_CATEGORY_LOOKBACKdays' already-published/already-pushed duplicates (to avoid aesthetic fatigue) - take top 1
Item 2 (experimental)
- Find in candidates.md:
- dimension combination vs most recent already-published sampleslargest difference(to increase calibration information), or
- contains an explicit pattern/dimension hypothesis (e.g. "MS=5 A/B comparison"), or
- tier=risky but the user is proactively willing to try (override with
--filter riskyoverride)
- composite is not necessarily top—but has "information value"
- if there is no suitable experimental candidate in the candidates pool → reply: "There are no obvious experimental samples in the candidate pool; here are 2 solid picks for you"
remaining (TOP_N - 2) items
fill in descending composite order, marked "(backup)".
anchor
for each one find 1-2 with close compositealready publishedworks as anchors (read from predictions/*.md ). Prefersame-durationanchors (by state.typical_duration_seconds ±20%)。
Phase 4: Output
🎯 候选池推荐(rubric: v2 / buffer: 🟢 绿 / cadence: 隔日更)
📌 第 1 条 — **稳分**(推荐立即拍):
**[tier1] [👍 9.18] "为你好"高密体系**
- 维度:ER=5 HP=5 QL=4 NA=4 AB=5 SR=5 SAT=4
- 粗预测桶:30-100w(中枢 ~60w)
- rationale:ER+SR 双 5 顶配,"高密度家庭议题"普适且分享安全
- 锚点:仓鼠 (composite 9.41, 实绩 124w) — 同走"理论框架+具象样本"路线
- 风险:议题厚重,不适合连续 2 篇都打这种
🧪 第 2 条 — **实验性**(验证特定假设):
**[tier1] [👍 8.71] 哈哈长度**
- 维度:ER=3 HP=5 QL=5 NA=4 AB=5 SR=4 SAT=5
- 粗预测桶:30-100w(中枢 ~55w)
- **测试目标**:v2.1 候选维度 MS+TS 双 5 vs 谁问你了同 ER/HP/QL/SR 但 MS+TS 低 3
- 信息价值:拍这条能强证据/弱推翻 v2.1 升正
- 锚点:谁问你了 (composite 8.24, 实绩 11.7w)
(备选 top 3):
3. ……
4. ……
5. ……
下一步:
- 选稳分 + 实验性各拍 1 条 → 改写 script → "启动预测"
- 只拍 1 条 → 选稳分(buffer 颜色越红,越应该选稳分)
- 想抓更多候选 → 说"抓热点"
- 都不满意 → 说"过滤改 all"看其他 tier 或 "regen"
if buffer color is 🔴:
🔴 buffer 警戒:你 buffer 已 0/1 篇,**下个发布日可能断更**。
按节奏协议,只推 top 1 稳分(不推实验性):
**[tier1] [👍 9.18] "为你好"高密体系**
- ...(同上稳分格式)
今天必须拍这条。挑 5 条候选 → "抓热点"。
if buffer color is 🔵:
🔵 buffer 积压:你 buffer 已 N 条,**暂停推荐**。
按节奏协议,先发存货 + 复盘。
- 已拍未发:N 条(最早一条 X 天前拍的)
- 待复盘:N 条
说 "已发布 ..." 出队,或 "复盘" 处理待复盘项。
如果你坚持要拍新的,回 "我就要拍",我会推 top 1 稳分。
Each one must have: dimension scores (so the user can challenge the scoring) + anchors (so the user can calibrate the credibility of the composite) + rationale (so the user can understand the recommendation logic).Not allowed to output only a composite ranking without explanation—that is a black box.
Key Rules
- Do not error, provide guidance. Missing candidates is the default state, not an error
- Do not recommend unscored ones. REQUIRE_SCORED=true is an honesty threshold—recommending material you have not read is astrology
- Must include anchors. A composite of 8.47 means different things on different accounts; anchors ground the abstract number in real samples
- Must include rationale. One sentence—why is this one stronger than the second?
- deduplicate published. Do not recommend already-published ones (the user may explicitly override)
Refusals
- "Just give me the highest composite, no need to explain the reason" → refuse. Showing the score + anchors is the only chance to discover a "mis-score"
- "Rescore every entry in candidates.md" → route to
/cheat-scoredo it one-by-one; batch rescoring is part of/cheat-bump, not within the scope of recommend - "Sort by predicted bucket, not by composite" → ask for the reason. bucket is a discretization of composite; sorting by composite is sorting by bucket, and the difference is the within-bucket order—if the user really wants to sort by "bet expected value", it needs to be multiplied by average plays, which is another independent scoring dimension
Integration
- Upstream:
/cheat-trendspulls external hot topics into candidates.md → recommend automatically sees them - Downstream: after the user picks one they write the script →
/cheat-predict(the candidate's rough composite does not enter prediction; prediction rescores) - In coordination with
/cheat-status: status shows "the candidates pool has N tier1 items unpublished"; recommend provides specific recommendations