Condense
VexJoy AI Agent with Intelligent Routing - /do routes plain-English requests to the right specialist agent and gates the work with reviews, tests, and a learning loop.
npx -y skills add notque/vexjoy-agent --skill condenseAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
What its author says it does
Copied from the file, not written here
Maximize information density: preserve all instructions, remove prose filler.
SKILL.md
3.6 KB, as published. Nobody here has run it
Condense
Strip prose filler from .md files. Preserve every instruction. This skill practices what it preaches.
Phase 1: SCOPE
Identify targets.
- Single file: User names a path. Read it.
- Glob: User gives a pattern (
agents/*.md). Expand, list matches, confirm with user. - 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.
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:
- Read the full file. Record word count.
- Rewrite in place applying the rules below.
- 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:
- YAML check: Confirm frontmatter parses.
python3 -c "import yaml; yaml.safe_load(open('<file>').read().split('---')[1])" - Report: Show
| File | Before | After | Reduction |table with word counts. - 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.