1 / 41
KineticKeeper, 2026.06
Thinking → Action → Observation → repeat
As the loop runs, observations pile up in the context
Use lots of tokens -> good results (X): stuffing in more information doesn't make it smarter
Good results -> use lots of tokens (O): you need enough information to get good results
Context = information, and context use => directly drives token usage
Using more tokens means paying more money
Managing context efficiently is how you cut token waste
Cut the waste and spend tokens where they matter
/compact [summary instructions]
CLAUDE.md
~/.claude/CLAUDE.md
{project path}/CLAUDE.md
{project path}/CLAUDE.local.md
/init
.claude/rules/
~/.claude/projects/{project path}/memory/MEMORY.md
MEMORY.md
mistakes, fixes, lessons, newly learned facts
/btw
/fork
/branch
/resume
/rewind
/compact
/context all
▐▛███▜▌ Claude Code v2.1.152 ▝▜█████▛▘ Opus 4.7 (1M context) with max effort | Claude Max ▘▘ ▝▝ ~/DEV ❯ /context ⎿ Context Usage ● ● ● ● ● ● □ □ □ □ □ □ □ □ □ □ □ □ □ □ Opus 4.7 (1M context) □ □ □ □ □ □ □ □ □ □ □ □ □ □ □ □ □ □ □ □ claude-opus-4-7[1m] □ □ □ □ □ □ □ □ □ □ □ □ □ □ □ □ □ □ □ □ 17.3k/1m tokens (2%) □ □ □ □ □ □ □ □ □ □ □ □ □ □ □ □ □ □ □ □ □ □ □ □ □ □ □ □ □ □ □ □ □ □ □ □ □ □ □ □ Estimated usage by category □ □ □ □ □ □ □ □ □ □ □ □ □ □ □ □ □ □ □ □ ● System prompt: 2.3k tokens (0.2%) □ □ □ □ □ □ □ □ □ □ □ □ □ □ □ □ □ □ □ □ ● System tools: 6.6k tokens (0.7%) □ □ □ □ □ □ □ □ □ □ □ □ □ □ □ □ □ □ □ □ ● Custom agents: 494 tokens (0.0%) □ □ □ □ □ □ □ □ □ □ □ □ □ □ □ □ □ □ □ □ ● Memory files: 5k tokens (0.5%) □ □ □ □ □ □ □ □ □ □ □ □ □ □ □ □ □ □ □ □ ● Skills: 2.8k tokens (0.3%) ● Messages: 41 tokens (0.0%) □ Free space: 982.7k (98.3%) Custom agents | /agents Memory files | /memory Skills | /skills └ 5 agents | 494 tokens └ 3 files | 5k tokens └ 40 skills | 2.8k tokens /context all to expand
allow / ask / deny
rm -rf
--dangerously-skip-permissions
.meta
PreToolUse
❯ ~/dev/my-game/Assets/Game/Common/Resources/dummy.txt.meta Change the GUID to 0 Read 1 file ● Update(my-game/Assets/Game/Common/Resources/dummy.txt.meta) ⎿ Error: PreToolUse:Edit hook error: ["$HOME/.claude/hooks/block-meta-edit.sh"]: No stderr output ● The .meta edit was blocked. The blocking hook ~/.claude/hooks/block-meta-edit.sh stopped it (enforcing the "no manual .meta edits" rule in CLAUDE.md).
dotnet build | grep "error CS"
Why do you think so?
superpowers
grill-me
ouroboros
SPEC-xxx.md
DESIGN-xxx.md
PLAN-xxx.md
/simplify
/code-review
ultracode
dynamic workflows
/deep-research
atlassian-page-cli
backtrace-tracker
reading-bytes-tables
LLM Wiki
.claude/rules/*.md
paths
ReAct Loop
Design an environment where the AI can produce good results, and keep maintaining it