Skill443 repo starsupdated 5d ago

condense

The condense skill strips prose filler from markdown files while preserving all instructions, rules, and technical content. Use it to reduce token overhead in documentation by running a mechanical whitespace cleanup pass followed by LLM-guided rewriting that removes redundancy, motivational framing, and obvious explanations while maintaining structure, code blocks, tables, and every actionable directive.

View source Repository: vexjoy-agent

Install in Claude Code

Copy

git clone --depth 1 https://github.com/notque/vexjoy-agent /tmp/condense && cp -r /tmp/condense/skills/code-quality/condense ~/.claude/skills/condense

Then start a new Claude Code session; the skill loads automatically.

Definition

SKILL.md

# Condense

Strip prose filler from .md files. Preserve every instruction. This skill practices what it preaches.

## Phase 1: SCOPE

Identify targets.

1. **Single file**: User names a path. Read it.
2. **Glob**: User gives a pattern (`agents/*.md`). Expand, list matches, confirm with user.
3. **Batch (10+ files)**: Dispatch parallel agents, one per file.

**Mechanical pre-pass** (deterministic, run before LLM condensing): strip trailing whitespace and consecutive blank lines that inflate Opus token counts. The script handles the mechanical reduction so the LLM phase focuses on prose density.

```bash
python3 scripts/check-whitespace.py --fix <target-file-or-dir> # 0=clean, 1=violations fixed
```

Run on the scoped targets (defaults to `agents/**/*.md` and `skills/**/*.md` when no path given). Then proceed to the LLM pass on the same files.

**Gate**: At least one target file identified and readable; mechanical pre-pass run.

---

## Phase 2: CONDENSE

For each file:

1. Read the full file. Record word count.
2. Rewrite in place applying the rules below.
3. Record new word count.

### Rules

**KEEP** (never cut):
- Every instruction, rule, gate, phase, step
- Tables, code blocks, commands, paths
- YAML frontmatter (do not alter)
- Structure: headers, numbered lists, phase ordering
- Technical terms naming specific things
- Reference loading tables
- Error handling sections
- Non-obvious "because X" reasoning

**CUT**:
- Redundant restatements of the same rule
- "Because X" on obvious rules
- Motivational framing ("this will help you", "it is important to note")
- Filler phrases: "in order to", "it should be noted that", "it is worth mentioning"
- Examples that repeat what the phase already says
- Paragraphs saying the same thing from different angles -- merge to one

**STYLE**: Short sentences. Active voice. Concrete words. If you can cut a word without losing an instruction, cut it.

### DELETE TEST

Before cutting any sentence: "If I remove this, does the reader lose an instruction, rule, or decision?" No = cut. Yes = keep.

### Boundaries

Do not reorganize sections, change meaning, add ideas, alter paths/commands, drop tables or code blocks, or modify YAML frontmatter values.

---

## Phase 3: VERIFY

For each condensed file:

1. **YAML check**: Confirm frontmatter parses.
```bash
python3 -c "import yaml; yaml.safe_load(open('<file>').read().split('---')[1])"
```
2. **Report**: Show `| File | Before | After | Reduction |` table with word counts.
3. **Instruction check**: Grep original for key terms (phase names, gate names, commands). Confirm each appears in condensed version. If any missing, restore from original.

**Gate**: YAML parses. No instructions lost. Reduction reported.

---

## Error Handling

**No prose to cut**: Report 0% reduction, move to next file.

**Instruction removed**: Re-read original, restore missing instruction, re-verify.

**YAML broken**: Restore original frontmatter verbatim, re-condense body only.

**Non-.md file**: Skip with warning.