Pular para o conteúdo
TakoAPI
Voltar às skills

cut-your-tokens-97percent-savings-on-session-transcripts-via-observation-extraction

por aeromomo

Coding Agents & IDEs
1.1kDownloads informados pelo clawskills.sh
0Visualizações
0 curtidaEntre para curtir esta skill
—Ainda sem avaliações
Sua avaliação

Resumo

Claw Compactor v6.0 — 50%+ savings through rule-based compression, dictionary encoding, session observation.

Descrição

O que esta skill faz

Compresses AI agent workspace files, including memory, session transcripts, and sub-agent context, using 5 layered techniques: rule-based dedup, dictionary encoding, observation extraction, RLE patterns, and format optimization. All compression is deterministic and runs without LLM calls. Session transcripts compress ~97%; first-run savings typically reach 50%+.

All compression runs locally without LLM calls, so there is no additional API cost to reduce token spend.

Quando usá-la

  • Reducing token costs on long-running Claude agent sessions
  • Compressing accumulated session transcripts before starting a new project
  • Running weekly maintenance on AI agent memory files via cron
  • Shrinking sub-agent context payloads before handoff
  • Auditing workspace token usage before a billing cycle

Exemplo de fluxo de trabalho

ENTRADA

User asks: compress my Claude agent workspace to cut token costs

AGENTE
  1. 1

    Run benchmark command to preview savings without modifying any files

  2. 2

    Execute full pipeline to apply all 5 compression layers in sequence

  3. 3

    Convert session JSONL transcripts to structured observation summaries

  4. 4

    Generate L0/L1/L2 tiered summaries for progressive context loading

  5. 5

    Save codebook to memory/.codebook.json for future decompression

SAÍDA

Workspace compressed 50%+ overall; session transcripts reduced ~97%; tiered summary files written to memory/

Instalação

clawhub install aeromomo/cut-your-tokens-97percent-savings-on-session-transcripts-via-observation-extraction

Execute o comando da CLI ou cole o prompt no Claude Code, Cursor, Codex ou no agente de código que você usa.

Links