working-style

claude-context-harness

Work all day in Claude Code without context rot — and make the harness better every session.

Tired of compacting? Start with The Shift — why long sessions degrade, why compaction is the wrong fix, and the workbench-not-filing-cabinet model that makes the problem disappear. The rest of this repo is the machinery; that page is the mental model that makes the machinery make sense.

This is a working harness built around two loops:

1. The memory loop — knowledge is saved to persistent file-based memory continuously at logical checkpoints (with pressure-triggered hooks as the backstop), so you /clear with confidence — clear, not compact — because everything important is already saved.
2. The improvement loop — end-of-session retrospectives ("how did we collaborate?") turn friction into feedback memories, hooks, and CLAUDE.md rules, so the next session is measurably better.

   work → hooks nudge delegation & watch context %
        → memory-capture (40%) / context-monitor (50%) prompt consolidation
        → /dream consolidates memories, /clear wipes the window
        → next session recalls from memory and starts clean
        ↘ session-retrospective → feedback memories / new hooks / new rules

Measured results

One month of real daily use (108 main sessions, ~3.6/day, mined from actual transcripts):

• 3.7% compaction rate — auto-compact fired in 4 of 108 sessions. The threshold arithmetic works: clearing at 50% means the ~90% compaction trigger is almost never reached.
• 28% of all transcript bytes diverted to subagents — 147.5 MB of reads, greps, and research ran in 801 disposable subagent windows instead of the main context. 56% of sessions delegated, ~7 dispatches per delegating session.
• Median session: 1.47 MB — short, sharp sessions instead of long degrading ones, because the memory system makes clearing free.
• Search-first reading — 1,598 grep/search commands vs 1,957 full-file reads, and 43.5% of reads bounded with limit/offset.
• The hooks do the reminding; the habits do the work — the context-pressure alert fires in 75% of sessions (its size-based fallback is deliberately conservative) and checkpointing happens inline in response; the formal checkpoint skill was never needed as a separate invocation. The local-knowledge and large-read guards fire in 45% and 42% of sessions. Meanwhile the delegation nudge fired in only 7 sessions and the hard agent-ceiling deny in 3 — those behaviors became habit, and the guards quietly retired into backstops.

Quickstart

Step 1 — Install the plugins (safe: touches nothing you have)

/plugin marketplace add mhardist/claude-context-harness
/plugin install context-discipline@claude-context-harness
/plugin install working-style@claude-context-harness

Plugin installs are purely additive. The skills, hooks, and commands live inside the plugin's own directory — your existing CLAUDE.md, settings.json, hooks, and skills are not modified. Don't like it? /plugin uninstall removes it cleanly. The hooks start nudging immediately; that's them working.

Step 2 — Adopt the templates (⚠️ merge, don't copy)

The files in templates/ are merge sources, not replacements. Never copy them over your existing files — you'd silently erase your own accumulated configuration. Back up first, then open the template side-by-side and pull in the sections you want:

# Back up before touching anything (yes, we practice what we preach)
cp ~/.claude/CLAUDE.md ~/.claude/CLAUDE.md.backup_$(date +%Y%m%d) 2>/dev/null
cp ~/.claude/settings.json ~/.claude/settings.json.backup_$(date +%Y%m%d) 2>/dev/null

• CLAUDE.md.example — append or weave sections into your existing ~/.claude/CLAUDE.md (or a project CLAUDE.md): the context-discipline rules, local-knowledge-first pattern, CLI-first preferences, learn-from-mistakes instructions. If you don't have a CLAUDE.md yet, copying the whole file is fine.
• settings.json.example — not a complete settings file. Merge individual keys (statusLine, permissions.deny, the extra hooks entries) into your existing settings.json by hand or with jq. If you already have hooks configured, append to the matching event arrays — don't replace them.
• mcp.json.example — reference for the env-var-placeholder pattern and the activate-on-demand philosophy; adapt to your servers rather than copying.

If you only do Step 1, you still get the full hook + skill machinery. Step 2 is where the standing instructions and statusline come from — worth doing, carefully.

What's inside

Plugin: context-discipline (the hero)