القائمة
Monitors and optimizes token usage across sessions. Audits instruction files, learnings, memory, and source code for bloat. Run periodically or on-demand to keep costs minimal.
التثبيت باستخدام Codex أو Claude انسخ هذا Prompt والصقه في Codex أو Claude أو مساعد آخر ليراجع صفحة Skill ويثبّتها لك.
استنادا إلى تصنيف SOC المهني
LLM-maintained development wiki for software projects. Two modes: CONCEIVE (develop software concept from idea to architecture) and TRACK (manage ongoing development with sprints, decisions, deployments, learnings). Creates and maintains an Obsidian-compatible wiki that accumulates project knowledge across sessions. Triggers on: 'dev wiki', 'wiki', 'init wiki', 'start project', 'new project', 'project wiki', 'knowledge base', 'track development', 'development plan', 'sprint planning', or when starting a new software project. ⛔ NOT: за sync на СЪЩЕСТВУВАЩ codebase wiki от source (ползвай wiki-sync).
Safe, project-aware fix workflow for the 4-layer ecosystem. Enforces impact-check BEFORE editing, surgical change, verify AFTER, and logs the gotcha — so fixes don't break other things and the lesson is captured. Use when the user says 'поправи', 'оправи', 'fix this', 'счупи се', 'fix the bug/error', or asks to fix anything in a project under {{CODE_PATH}}. NOT for finding an unknown root cause (use systematic-debugging first).
Autonomous self-improvement loop — adversarial audit → grouped fixes → fresh-eyes verify → repeat until dry, with PROOF of completion at every step. Use when the user says "improve the system", "self-improvement loop", "намери и поправи слабостите", "пусни loop на самоподобрения", "одит и оправи всичко", "направи го перфектно", or wants a full find→fix→verify cycle where "done" must be demonstrated, never claimed.
Generic knowledge base ingest skill for any Karpathy-method research wiki. Automatically processes new RAW files (raw/transcripts/*.md) into wiki summaries, updates index.md, hot.md, knowledge_graph.json, NotebookLM, and processed.json. Use when: "обнови wiki", "провери за нови файлове", "ingest", "нови raw файлове", or when user adds .md files to a raw/transcripts/ folder. Configure per-project via KB_CONFIG.md in the wiki base directory. Works for: Claude Trading, Claude Code Research, AI Video, or any new research wiki.
Builds a compact JSON knowledge graph for a software project, enabling ~70x token savings compared to scanning source files. Creates knowledge_graph.json (meta + architecture + critical_rules + clusters) and per-cluster Obsidian .md files for RAG-style access. Triggers on: 'graphify', 'knowledge graph', 'map the project', 'build graph', 'create graph', 'граф на проекта', 'знания граф', or when starting work on a new project without an existing graph.
Bootstrap a new project with the full 4-layer memory architecture (Pinecone + Obsidian Wiki + GitNexus + CLAUDE.md). Use when user says "setup 4-layer for <project>", "add memory architecture", "onboard <project> to memory ecosystem", "init memory for new project", or similar. Creates reproducible infrastructure that cuts session startup tokens 40-180x across projects.
| name | token-optimizer |
| description | Monitors and optimizes token usage across sessions. Audits instruction files, learnings, memory, and source code for bloat. Run periodically or on-demand to keep costs minimal. |
| invoke | /optimize-tokens |
You are a token optimization specialist. Your job is to minimize the tokens consumed per session while preserving ALL useful information. You achieve this by running audit scripts and applying fixes.
/optimize-tokens, "оптимизирай токените", "намали разхода"python3 ~/.claude/skills/token-optimizer/scripts/audit.py
This script measures ALL auto-loaded files and outputs a structured report with:
DO NOT read large files manually to count lines — the script does it.
Based on the audit report, apply these fixes IN ORDER of impact:
| Priority | Action | Trigger |
|---|---|---|
| 1 | Archive processed learnings | .ai/learnings.md > 50 lines |
| 2 | Condense CLAUDE.md files | Any CLAUDE.md > 100 lines |
| 3 | Deduplicate memory vs CLAUDE.md | Overlapping content detected |
| 4 | Split large reference files | Any memory/*.md > 80 lines |
| 5 | Clean stale patterns | .ai/patterns.md entries with 0 evidence > 30 days |
After fixes, re-run the audit script to confirm savings:
python3 ~/.claude/skills/token-optimizer/scripts/audit.py
Report the delta to the user: "Спестени X токена/сесия (Y% намаление)"
To analyze token cost of a specific skill:
PYTHONIOENCODING=utf-8 python3 ~/.claude/skills/token-optimizer/scripts/skill_cost.py <skill-name>
To see all skills ranked by cost:
PYTHONIOENCODING=utf-8 python3 ~/.claude/skills/token-optimizer/scripts/skill_cost.py --all
To see only metadata cost (always-loaded overhead):
PYTHONIOENCODING=utf-8 python3 ~/.claude/skills/token-optimizer/scripts/skill_cost.py --all --metadata
The script shows 3-layer breakdown: Metadata (always) → SKILL.md (on trigger) → references/ (on demand).
~/.claude/skills/token-optimizer/references/last-audit.json before running. If last audit < 24h ago AND no files changed, skip with "Вече оптимизирано."