SuperAPI AI Directory

anibon-macro-density

Writing & research

View repository

Updated 10-16 05:54

About this skill

Anibon Macro Density

Enforces aggressive timestamp merging so the final output stays compact (≤ 20 YouTube comments for a 9-hour stream) without losing any major context.

The Problem This Skill Fixes

Observed baseline failure (2026-07-18, stream Chdc68BJnNo):

  • 70 raw parts generated for a ~9-hour stream
  • Most sections had 1 timestamp each
  • Same ongoing talk topic got re-stamped every 5-minute chunk
  • Result: verbose, spam-level comment quota usage

Core Rule

One timestamp = one MACRO activity shift, not one sub-topic mention.

A 9-hour talk stream should produce ≤ 30 raw timestamps total. After Summarizer merging → target 15–20 YouTube comments.

Density Hierarchy

What changed Stamp?
Completely different game title ✅ Yes
New guest speaker joins/leaves ✅ Yes
Stream switches from gaming → talk (or vice versa) ✅ Yes
Entirely new news story with no connection to prior topic ✅ Yes
Sub-topic tangent within same ongoing conversation ❌ No — fold into existing
Same topic continues across chunk boundary ❌ No — output 0 timestamps
Minor donation read during ongoing talk ❌ No — only stamp if it derails the topic 5+ min
Speaker mentions a new detail of the same story ❌ No — update prior description

Merging Rules for Subagents

  1. Continuous talk = 1 stamp. If Boat is still discussing the same news story, political topic, or game reaction across multiple chunks → that is 1 timestamp regardless of how many sub-angles he covers.
  2. 0 is correct. Outputting 0 timestamps for a chunk is the right answer when the chunk is a continuation. Do not feel pressure to produce output.
  3. Comma-list, not multi-line. If two genuinely distinct topics occur in same chunk, fold them into one description: วิเคราะห์ Prince Group, ติงช่องอมรินทร์ดราม่าเขมร — one line.
  4. No sub-topic lines. [Talk] พูดถึง... followed by another [Talk] 3 minutes later in same chunk = violation. Merge.

Summarizer Density Rules

After Map stage collects all subagent results, the Summarizer MUST:

  • Target ≤ 20 final parts for streams up to 9 hours.
  • Pack each part toward the 3,500-byte ceiling before splitting.
  • Merge any two adjacent parts whose combined bytes stay under 3,200 — always.
  • A final part with < 4 timestamps is a smell; merge it with its neighbor unless byte limit prevents it.

Chunking Parameter Override

For next stream: use --block 600 --overlap 60 (10-min chunks) instead of default --block 300 --overlap 30. This halves the number of subagents and naturally enforces macro density.

Red Flags — Output 0, Do Not Stamp

  • Chunk starts with Boat still mid-sentence on prior topic.
  • Only content is chat replies within an existing talk flow.
  • Chunk is reaction to the same video/article already stamped in prior chunk.
  • Donation that lasts < 2 minutes before returning to main topic.

Real-World Impact

Without this skill: 70 parts, 70 YouTube comments needed.
With this skill: ≤ 16 parts, 16 comments needed.
Token cost and review time: roughly 4× cheaper.