ra-复盘
Goal
Turn "looking at data" into "accumulating assets." Each review produces three things: a classification conclusion, evidence-based attribution, and at least one deposit. The only pipeline for enriching the material library (quotes/cases/concepts), methodology, and course inspiration hinges on this process.
Workflow
1. Locate the Work
- Data ledger:
01-Content Production/Data Statistics/<Platform>/Work Data.md(matched by title) - In-library body: Video →
Video Workstation/Completed/handover document orTranscript/; Article →Article Workstation/; Graphic →Graphic Workstation/Completed/ - Topic card:
00-Topic Pool/(if no card exists, create one at the end of the review)
2. Fetch Data
- Latest metrics and increments: AUTO block in the ledger
- Historical curve:
automation/daily_snapshots/own/<Platform>/*.json(daily snapshots) - Follower count (for M calculation): Corresponding platform
Account Data.md - When ledger and snapshot conflict, the snapshot JSON prevails
3. Hit Classification (Reuse Standards, Don't Invent)
At runtime, read 03-Benchmark Hit Library/Hit Classification Standard.md for the latest definitions and thresholds. Applicable to own accounts as well:
- R = Core metric of the work / Median of the account's last 20 posts (use all if fewer than 20 and note it; for 抖音 use likes, for 小红书 use likes + saves)
- M = Likes / Followers, calibrated by Tier (follower size layer) threshold
- Show the calculation process before giving the conclusion; when sample size is small, state it as a "tentative conclusion"
4. Content Body (Pull on Demand, Don't Pre-Fetch)
- Full text available in library → read directly (works produced by the new system are all in the library)
- Historical 抖音/小红书 works: Pull single details via TikHub MCP
- Own videos needing transcript:
ra-Transcript Extraction Skill(Volcengine ASR, cost ~¥0.4/hour) - Tweets/Posts: Official API or
https://r.jina.ai/<Tweet URL>to pull full text
5. Attribution
- Against the content body, answer: what is the hook, what did the first 7 seconds/cover image deliver, how is the structure, what is the traffic nature of the topic, posting timing
- When there is a comparable benchmark, run
automation/scripts/hot_monitor_local_job.sh read boomsto filter cloud-based samples of the same topic, high R/M, and similar scale; if transcript or metric history is needed, then runautomation/scripts/hot_monitor_local_job.sh read work <Cloud Work ID> 01-Content Production/Video Workstation/.internal/Benchmark Hit Knowledge Assets/serves only as historical deep knowledge supplement, not as the true source of real-time metrics; it contains source information; continue to respect privacy walls when referencing- Prohibit blind attribution: each attribution must point to a specific sentence/image in the content, or a specific inflection point in the data (which day it rose, which metric rose)
6. Deposit (Golden Rule Step)
| Deposit Type | Destination | Inclusion Criteria |
|---|---|---|
| Validated hooks/titles/expressions | Material Library/Quote Library.md | Opening sentence/title of highly graded works (with data: "sentence" |
| Reusable operational process | Material Library/Case Library.md | Tool/tutorial-type works' scenario/operation/result/reusable points |
| Validated analogies/explanation frameworks | Material Library/Concept Library.md | Cognitive building blocks used in explanations (e.g., "whiteboard analogy") |
| Platform patterns | Methodology/ corresponding file | Append entries with date, do not overwrite old conclusions; mark single samples as "tendency", only write as pattern when three or more samples align |
| Course inspiration | 02-Course System/Course Topic Library/Inspiration Log.md | Informative content with good data |
| One-line review conclusion | Topic card "Review" section | Required every time (date + classification + one-line attribution) |
7. Output
Provide the review report within the conversation: classification (with calculation) → attribution → deposit list (what was written, to which file). The report itself is not stored—assets go into the library, the report stays in the conversation.
Hard Rules
- Each review must deposit at least one asset; if there truly is nothing to deposit, explicitly state the reason (e.g., "data is flat with no new expression")
- Quotes must be deposited with data; do not accept sentences that "feel good"
- Distinguish between "single-case tendency" and "multi-case pattern" in attribution; only the latter goes into the methodology body
- Suspicious rate-limiting/violation flags: first record in the topic card's review conclusion; platform rule library updates await the audit pipeline activation
- Own works are not subject to the source privacy wall (that rule is for plagiarized sources only)