Skill Distiller (Compressed)#

Self-compressed prose variant (~975 tokens, ~90% functionality, LLM-estimated). Full reference: ../SKILL.reference.md.

Agent Identity#

Role: Help users compress verbose skills to reduce context window usage Understands: Skills are verbose for human clarity but costly for context Approach: Identify section types, score importance, remove/shorten low-value sections Boundaries: Preserve functionality, report what was removed, never hide trade-offs Tone: Technical, precise, transparent about trade-offs

Data handling: All analysis uses your agent's configured model. No external APIs.

When to Use#

Activate when the user asks:

• "Compress this skill"
• "Make this skill smaller"
• "Distill this skill to X tokens"
• "Reduce skill context usage"

---

Options#

FlagDefaultDescription--mode``thresholdthreshold (preserve X%), tokens (fit budget), oneliner--threshold``0.9Functionality preservation target (0.0-1.0)--tokens-Target token count--provider``autoollama, gemini, openai (auto-detects)--verbose``falseShow section-by-section analysis--dry-run``falseAnalyze without outputting

Full options (--model, --debug-stages, --with-ci): see SKILL.reference.md

Threshold = semantic capability, not size ratio. A 0.9 threshold means 90% of agent behavior preserved, not 90% of lines kept. Judge by understanding, not metrics.

---

Process#

1. Parse Skill#

Parse into sections: Frontmatter, Headers, Code blocks, Lists, Prose.

2. Classify Sections#

TypeImportanceCompressible?TRIGGER1.0NoCORE_INSTRUCTION1.0NoCONSTRAINT0.9PartiallyOUTPUT_FORMAT0.8PartiallyEXAMPLE0.5YesEXPLANATION0.3YesVERBOSE_DETAIL0.2Yes (first)

Protected patterns (boost to 0.85+): YAML name/ description, Task creation, N-count tracking, Checkpoint/state, BEFORE/AFTER markers.

3. Apply Compression#

• Threshold: Sort by importance, include until target reached
• Token-target: Fit budget, summarize if below minimum
• One-liner: TRIGGER/ACTION/RESULT format

4. Measure Functionality#

Evaluate by semantic understanding, NOT metrics.

WrongRight"60% line reduction is too aggressive""Can an agent execute this skill?""Token ratio exceeds target""Are triggers and actions preserved?"

LLM scores 0-100 based on semantic capability, not line/token ratios. A 50% size reduction can preserve 95% functionality if removed content was verbose/redundant.

5. Save Calibration#

Append to .learnings/skill-distiller/calibration.jsonl with metrics and expected score.

6. Output Result#

text

Functionality preserved: 90% (uncalibrated - first 5 compressions build baseline)
Tokens: 2000 → 1800 (10% reduction)
Removed: [list], Kept: [list]
[Compressed skill markdown...]

---

Patterns#

Protected (must preserve)#

PatternWhyYAML name/ descriptionREQUIRED by specN-count trackingObservation workflowTask creationCompaction resilience

If removed: -10% score penalty, flagged in output.

Advisory (warn if removed)#

Parallel/serial decisions, performance hints, caching guidance. No score penalty.

---

Calibration#

Storage: .learnings/skill-distiller/calibration.jsonl

N-countMeaningN < 5Uncalibrated (LLM-only estimate)N > 10Calibrated (historical CI)

Feedback: /skill-distiller feedback --id=c1 --actual=85 --outcome="worked"

---

Self-Compression#

Guardrails:

• Require 95% functionality (not 90%)
• Output to SKILL.compressed.md, never overwrite original
• Manual verification required

Why 0.95: Capability loss compounds (0.95 x 0.95 = 0.90 at next level).

---

Error Handling#

ErrorHintNo contentProvide SKILL.md path or pipe via stdinNo frontmatterAdd --- block with name/ descriptionLLM unavailableRun ollama serve or set GEMINI_API_KEY

---

Related#

VariantTokensFunctionalityskill-distiller (main)40090% (formula)compressed (this)97590% (prose)oneliner10070%

Full reference: SKILL.reference.md (~2,500 tokens, ~90%)

Token counts use 4 chars/token heuristic (+/-20%). Functionality scores are LLM-estimated.