Context Management¶

Long agent sessions fill the context window, burying your instructions under noise. The model forgets your conventions, hallucinates file contents, and every reply costs more. This tutorial explains how to manage context effectively when running ClawBio skills.

What Is Context?¶

Every message re-sends the entire conversation to the API — your rules, all previous messages, every tool output. This bundle is the context window, measured in tokens (~0.75 words each).

What fills it:

System prompt / CLAUDE.md — fixed overhead every session
Chat history — grows ~1k tokens per turn
File reads — one 500-line file ≈ 3k tokens
Bash/test output — a single stack trace ≈ 5k–20k tokens (the silent killer)
Web search results — 5k–10k per search

Context Rot and Cost¶

Context rot happens past ~70% capacity: early instructions get buried, the model attends to recent noise over old rules, and hallucinations increase.

Cost scales linearly — every message re-sends the full window.

Model	$/MTok input	Context window	Cost at 50k	Cost at 150k
GPT-4o mini	$0.15	128k	$0.01	$0.02
Kimi K2.5	$0.60	256k	$0.03	$0.09
Claude Haiku 4.5	$1.00	200k	$0.05	$0.15
Claude Sonnet 4.6	$3.00	200k	$0.15	$0.45
Claude Opus 4.6	$5.00	200k	$0.25	$0.75

50 messages at 150k tokens on Sonnet = $22.50 in input costs alone. Context management is a billing issue, not just a quality issue.

For up-to-date pricing, see models.dev.

Why This Matters for ClawBio¶

ClawBio skills generate substantial output — genomic reports, variant tables, pipeline logs. A typical multi-skill session can fill 50% of the context window in 10–15 turns. If you're chaining skills (e.g. PharmGx → NutriGx → Profile Report), context management becomes essential.

The Solutions¶

`/clear` — Full Reset¶

Resets the conversation entirely. Zero history, zero tool outputs, back to just the system prompt. Use it between unrelated tasks or when a session has gone off track.

`/compact` — Summarise and Continue¶

Summarises the entire conversation into a structured digest (~5k tokens) and discards the raw history — typically a 97% reduction. Unlike /clear, it preserves continuity: decisions, file states, and completed work survive in the summary.

	`/clear`	`/compact`
History after	Gone	Summarised (~5k tokens)
When to use	Between tasks	Mid-task, keep continuity
Cost	Free	One summary call

Check Your Context¶

Use /context at any time to see how full the window is:

❯ /context

  Context Usage
  claude-sonnet-4-6 · 56k/200k tokens (28%)

  Estimated usage by category
  System prompt:   3.5k tokens  (1.7%)
  System tools:     21k tokens (10.5%)
  Messages:        28.2k tokens (14.1%)
  Free space:       140k tokens (69.8%)

Best Practices for ClawBio Sessions¶

One skill per session when possible — run PharmGx, compact, then run NutriGx
Compact after large outputs — if a skill dumps a 500-line report into the context, compact before continuing
Use --output flags — write results to files instead of printing them all to the terminal
Start fresh for unrelated tasks — /clear is free and instant
Watch for hallucinations — if the agent starts inventing gene names or mixing up variant data, it's probably context rot

Auto-compact is not reliable

The automatic trigger at 80% only saves titles and brief excerpts — not full content. Use /compact manually before you hit that threshold.

Further reading: Anthropic: Building Effective Agents