From predictions/*.md retrospective section's comment data, cluster the account's real audience persona, and write it to audience.md。
Core positioning
persona is the second derivative parallel to the rubric, parallelsecond derivative,not part of the rubric:
复盘数据(评论 + 完播 + 转粉)
├──→ rubric 进化(cheat-bump) —— "怎么打分"
└──→ 受众画像(cheat-persona) —— "谁在看"
两者都喂给 cheat-seed,但用途不同
- rubric: will this draft go viral → feed cheat-predict for scoring
- persona: who will watch 3 more seconds / leave a comment / repost because of this → feed cheat-seed for topic selection + drafting
Never mix: persona does not enter the scoring formula. The rubric's AB dimension (audience breadth) can reference persona, but that is cheat-bump's business, not here.
⚠️ Contamination isolation (non-optional)
audience.md Derived from retrospective comments = contains performance signals from published works. Therefore:
audience.mdIn cheat-score-blind 's hard refusal list, refusal_codeblocked_audience- persona affects cheat-seed what to write(creative direction), does not affect cheat-predict how to score(the blind sub-agent never reads audience.md)
- This is clean: the content shaped by persona goes into the final draft, and the blind sub-agent scores the final draft itself—no leak. A leak would only happen if the sub-agent could read audience.md and say "add points because this audience loves X," and it cannot read it
Overview
[用户:构造受众画像 / 更新 persona]
↓
[Phase 0: 收集数据 — 扫 predictions/*.md 复盘段评论 + benchmark.md]
↓
[Phase 1: 数据量判定 → 派生 Confidence 等级]
↓
[Phase 2: 评论聚类 — 自我认同 / 情绪寄存 / 反驳点 / 语言]
↓
[Phase 3: persona × rubric 交叉检验]
↓
[Phase 4: 写 audience.md(覆盖式重建,header 记 version + last_rebuilt)]
↓
[Phase 5: 控制台报告 + 跟上次画像的 diff]
Constants
- AUDIENCE_PATH = audience.md — audience persona storage location
- MIN_RETROS_FOR_DATA_GROUNDED = 3 — retrospective count ≥3 counts as "solid data" (can be softly judged based on comment quality)
- MIN_COMMENTS_PER_TRAIT = 3 — a "verified trait" needs at least 3 pieces of comment evidence, otherwise downgrade to "hypothesized trait"
- SEED_FROM_BENCHMARK = auto — when there is no own retrospective data but there is a benchmark, seed an unverified persona
💡 Invocation coverage:
/cheat-persona — seed-from-benchmark(force use benchmark seed) /— rebuild(rebuild even if data has not changed)
Inputs
| Source | Purpose |
|---|---|
predictions/*.md 's ## 复盘 section |
Primary data source—top comments (with like counts). persona's gold mine |
videos/*/report.md |
Completion / follow conversion rate—thin signal, inferring "retains vs does not retain" |
benchmark.md |
Cold-start seed—"people who watch the benchmark ≈ the people you want" |
rubric_notes.md |
For Phase 3 cross-check—persona appetite vs rubric-calibrated reality |
audience.md(if already exists) |
Previous version of the persona, used for Phase 5 diff |
Workflow
Phase 0: Collect data
- Glob
predictions/*.md, for each file read the## 复盘section (read only the retrospective section—this is channel A, which already looks at performance) - Extract each piece's top comments (with like counts) + performance bucket
- Count: number of retrospective pieces with comments
N_retros, total commentsN_comments - Read
benchmark.md(if exists) - Read
audience.md(if already exists) → keep for Phase 5 diff
Phase 1: Data volume determination + Confidence
| Situation | Confidence | Behavior |
|---|---|---|
N_retros == 0 and no benchmark |
🔴 No data | Don't force-create——tell the user "persona needs retro data. Run a few cheat-retros first, or import a benchmark", exit |
N_retros == 0 but there is a benchmark |
🟠 benchmark-seed unverified | seed an aspirational persona,mark the whole text as "unverified" |
N_retros 1-2 |
🟡 Early signals | can produce output but most traits fall into the "hypothesis" section |
N_retros 3-5 |
🟢 Solid data | normal output |
N_retros ≥6 |
🔵 Robust | normal output + can do finer appetite segmentation |
Write the Confidence level into the audience.md header.
Phase 2: Comment clustering
For all collected comments, cluster along four dimensions:
- Self-identification——"me too…" / "this is me" / "as a…" patterns. Count which identities recur ("big-tech worker" / "one-person company" / "grad school exam taker" / ...)
- Emotional deposit——what emotion are viewers commenting for? Validation ("so true") / venting ("I'm so tired too") / arguing ("I disagree") / asking for help ("so what should I do"). Count the proportions
- Counterpoints——which views draw consistent opposition. This is the persona boundary
- Language——how they talk. Meme density, sincere vs playful, whether they copy your signature lines
Clustering discipline:
- A "verified trait" needs at least
MIN_COMMENTS_PER_TRAIT(3) pieces of comment evidence. Not enough → demote to the "hypothesis trait" section - Every traitmustbe able to cite specific comments + source (which prediction's retro section) + count
- If you find "anti-persona" signals (the audience you thought you had vs the people actually commenting are different) → write an "anti-persona" section
Phase 3: persona × rubric cross-check
Read rubric_notes.md the current rubric + calibration history. Check:
- persona says "the audience loves X-type topics" → do X-type topics really over-perform in the rubric calibration pool?
- If inconsistent → in the "persona × rubric cross-check" section of audience.md flag it
Honesty requirement: when the two derivatives contradict,do notforce reconciliation. Explicitly write "persona says A, rubric calibration says B, to be clarified in the next retro"——the contradiction itself is a signal.
Phase 4: Write audience.md
Overwrite rebuild(not append)——persona is a living document, rewrite the whole text on every rebuild. But:
- the header's
Persona 版本+1(v0 → v1 → v2) - the header records
Last rebuiltdate +数据基础(N retros / M comments) +Confidence - the "version history" section at the bottom of the file append one line:
vN — 基于 M 篇复盘 / K 条评论,主要变化:...(this is the only history kept; no memo accumulation——persona is a living document, not a formula)
Use templates/audience.template.md the structure.
⚠️ No version memo / no cross-model review——persona is not a high-risk irreversible action (if written wrong, just rerun it), over-engineering is unnecessary.
Phase 5: Report
✅ 受众画像已更新:audience.md(v2,🟢 数据扎实)
数据基础:4 篇复盘 / 87 条评论
核心画像:25-35 岁职场人,来找情绪共鸣不来找信息……
跟 v1 的主要变化:
- 新验证:"深夜刷手机" 场景共鸣强(v1 是假设,本次 17 条评论验证)
- 新反画像:原以为"学生党"是受众,但评论里学生占比 <5% → 移到反画像
- ⚠️ 交叉检验 flag:persona 说受众爱"职场吐槽",但 rubric 校准显示职场类 composite 偏低——下次复盘留意
下一步:
- cheat-seed 选题 / 写稿时会自动参考这份画像
- 再跑 3 篇复盘后建议再 /cheat-persona 刷新
Key Rules
- Data-derived, not hand-written——persona must come from comment clustering. If the user wants to manually add traits → allowed, but mark
user-asserted(not data-validated) - Evidence enforcement——validated traits must carry comment count + source. Those without evidence go into the "hypotheses" section
- Overwrite-style rebuild——each rebuild rewrites the full audience.md, only appending one line in the version history section
- Not fed into scoring——persona is never fed into cheat-predict / cheat-score-blind. It is cheat-seed's creative lens
- Contradictions not reconciled——when persona × rubric conflict, flag it truthfully, don't force a story
- Cold-start honesty——if there's no data, say there's no data; benchmark seed is marked "unvalidated" throughout
Refusals
- "I think my audience is just X, write it straight into audience.md" → you can write it, but mark
user-assertedand put it in the "hypothesized traits" section, not the "validated traits" section. The value of persona lies in the gap between the data vs your fantasy - "Also use persona for cheat-predict, to make scoring more accurate" → refuse. persona is a derivative of actual performance; feeding it into scoring = punching through channel B's isolation. persona only serves cheat-seed
- "Skip comment clustering, just draw me one based on your gut" → refuse. What's drawn from gut feeling is marketing talk, not persona. If there's no comment data, honestly say "go do a retrospective first"
- "Write persona into rubric_notes.md, save a file" → refuse. rubric_notes.md is the blind whitelist; writing persona (a performance derivative) into it = the performance leak vulnerability repeats (see the leak guard in observation-lifecycle.md)
Integration
- Upstream:
cheat-retroeach time a retrospective is completed → flag "N retrospectives accumulated, you can run /cheat-persona to refresh the profile" - Upstream:
cheat-initcreate an empty audience.md skeleton; if a benchmark was imported → prompt that you can/cheat-persona — seed-from-benchmark - Downstream:
cheat-seedMode A/B/C read audience.md as the mirror for "would this persona care?" - Downstream (phase 2 roadmap):
cheat-recommendpersona-fit ranking;cheat-statuspersona freshness nag - Isolation:
cheat-score-blindhard-forbidden from reading audience.md (refusal_codeblocked_audience)
Known limitations
- Comments ≠ the entire audience——those who leave comments are the most active sliver of the audience (the silent majority isn't in the data). persona skews toward "people who comment", not "everyone who watches"
- Platform comments can be polluted——water army / flame-baiting / off-topic comments will enter the data. cheat-persona downweights obvious outliers during clustering, but can't fully filter them out
- persona lags behind real audience changes——the profile is based on comments from the past N posts. If the audience structure changes, it won't be reflected until new retrospectives accumulate
- Doesn't solve "the audience I want ≠ my actual audience"——persona only truthfully reports "who's watching now". Wanting to pivot to another audience is a topic-selection strategy question; cheat-persona only provides the gap between "current state vs goal", it doesn't make strategy for you