AgentSkillsCN

condense

在保持内容价值的前提下,智能精简文章长度,绝不随意截断。

SKILL.md
--- frontmatter
name: condense
description: Intelligently reduce article length while preserving value. Does not arbitrarily truncate.

Condense Article

Reduce an overly long article's word count while preserving its essential arguments, unique insights, and distinctive voice.

When to Use

  • When /condense [filepath] is invoked
  • When executing a condense task from todo
  • When an article exceeds critical length threshold
  • When deep-review encounters an article over hard threshold

Principles

What to Cut

  1. Redundancy - Content that says the same thing twice

    • Repeated explanations in different sections
    • Multiple examples making the same point (keep the best one)
    • Introductory sentences that restate what follows
  2. Background LLMs already know - Standard definitions and history

    • Basic philosophical terms (consciousness, qualia, etc.) - link instead
    • Historical overviews covered in other articles
    • Common objections that have their own pages
  3. Tangential sections - Content that could be its own article

    • Extended digressions from the main argument
    • Deep dives into related topics
    • Detailed examples that overshadow the point
  4. Excessive hedging - Overqualified language

    • "It might perhaps be argued that one could say..."
    • Multiple caveats for the same point
    • Defensive qualifications that weaken rather than clarify
  5. Over-attribution - Too many citations for obvious claims

    • Multiple citations for uncontroversial facts
    • Block quotes that could be paraphrased
    • Extended literature reviews

What to Preserve

  1. Opening summary - Front-loaded key claims for LLM truncation resilience
  2. Unique arguments - The Map's distinctive contributions
  3. Relation to Site Perspective - Tenet connections (never remove this section)
  4. Author voice - Distinctive phrasings that work
  5. Critical cross-links - Connections to other Map content
  6. Further Reading and References - Keep but can trim

What to Extract (Not Delete)

Some content is valuable but doesn't belong in this article:

  • Create linked articles for detailed subtopics
  • Move extended examples to separate pages
  • Defer background to existing concept pages
  • Replace removed content with: "For details, see [[linked-article]]"

Instructions

1. Analyze Current State

First, measure the article and understand the problem:

bash
uv run python -c "
from pathlib import Path
from tools.curate.length import analyze_length
a = analyze_length(Path('[filepath]'))
print(f'Current: {a.word_count} words')
print(f'Target: {a.soft_threshold} words')
print(f'Excess: {a.excess_words} words ({a.excess_percent:.0f}% of target)')
print(f'Status: {a.status}')
print(f'Section: {a.section}')
"

Thresholds by section:

SectionTargetHardCritical
concepts/250035005000
topics/300040006000
apex/400050006500
voids/200030004000

Calculate reduction needed:

  • Target: Middle of acceptable range (e.g., 2000 for concepts, 2500 for topics)
  • Minimum: Reduce to at least soft threshold
  • Ideal: Reduce to target while preserving quality

2. Structural Review

Read the full article and identify:

  1. Essential sections - Core argument, must keep
  2. Optional sections - Supporting but not critical
  3. Redundant passages - Repeat earlier points
  4. Background to cut - Could be assumed or linked
  5. Tangents to extract - Could be separate articles

3. Plan Cuts (Show Work)

Before editing, document planned changes in a table:

SectionWordsActionReason
Opening350KeepFront-loaded summary
Background on X800Cut 60%Standard material, link to [[x]]
Main Argument1200KeepCore contribution
Historical Survey600ExtractCreate separate article
Example 1300KeepBest illustration
Example 2250CutRedundant with Example 1
Example 3200CutRedundant
Objections500Trim 30%Some repetition
Relation to Site400KeepRequired section

Expected result: Current 4500 → Target 2800 (reduction of 1700 words)

4. Apply Condensation

Edit the article following the plan:

Cutting techniques:

  • Replace paragraphs with single sentences + links
  • Combine related sections under one heading
  • Remove redundant examples (keep the best one)
  • Tighten prose:
    • "It is the case that X" → "X"
    • "In order to" → "To"
    • "Due to the fact that" → "Because"
    • "One might observe that" → (delete)
    • "It should be noted that" → (delete)

Extraction technique (when content should exist elsewhere):

  1. If creating new article: write it first with proper frontmatter
  2. If deferring to existing article: verify the link resolves
  3. Replace removed content with: "See [[linked-article]] for details on X."

Combining technique:

  • Merge sections with overlapping themes
  • Use the strongest version of repeated arguments
  • Create clear transitions between merged content

5. Verify Quality

After condensation, check:

  1. Flow - Does the article still read smoothly?
  2. Opening summary - Does it still make sense standalone?
  3. Core argument - Is the main point preserved?
  4. Tenet connection - Is "Relation to Site Perspective" intact?
  5. Cross-links - Do all internal links still work?
  6. Voice - Does it still sound like the original author?

6. Update Frontmatter

yaml
ai_modified: [current ISO timestamp]
# If significant changes to human content:
ai_contribution: [increase appropriately]

7. Log to Changelog

Append to obsidian/workflow/changelog.md:

markdown
### HH:MM - condense
- **Status**: Success
- **File**: [filepath]
- **Before**: [word count]
- **After**: [word count]
- **Reduction**: [percentage]%
- **Technique**: [summary: cut redundancy, extracted [topic] to new article, etc.]

8. Commit

code
refine(content): condense [filename] from N to M words

- [Key cut 1]
- [Key cut 2]

Preserved core arguments and tenet alignment.

What NOT to Do

  • NEVER truncate arbitrarily (cutting from end without review)
  • NEVER remove the opening summary - It's critical for LLM access
  • NEVER remove "Relation to Site Perspective" - Required section
  • NEVER sacrifice unique arguments for length - Preserve the Map's voice
  • NEVER cut all examples - Keep at least one good illustration
  • NEVER create orphan content - If extracting, ensure new article is linked
  • NEVER ignore author voice - Preserve distinctive phrasings

When to Leave an Article Long

Some articles genuinely need their length. Document and leave if:

  1. Topic complexity - The subject truly requires extended treatment
  2. Tenet support - The article is central to defending core tenets
  3. Unique coverage - No other resource covers this material
  4. High quality - Every section adds distinct value

If leaving long, add a note to the changelog:

markdown
- **Status**: Reviewed, no condensation applied
- **Reason**: [explanation]

Important

  • This skill MODIFIES content
  • Always preserve the article's core value proposition
  • Target the middle of the acceptable range, not the minimum
  • Quality > brevity - don't sacrifice clarity for word count
  • Document all significant cuts in the changelog
  • If unsure whether to cut something, keep it