v1.1.0 — Battle-tested Install + High-Resolution Skills

compound-engineering v1.1.0

Install once. Your AI coding agent automatically learns from every session — plans, records, remembers, reviews. Code quality compounds over time, on autopilot.

What's New in v1.1.0

One-line install & uninstall (tested end-to-end)

Install (Claude Code + Codex, one shot):

curl -fsSL https://raw.githubusercontent.com/cklxx/compound-engineering/main/install.sh | bash

Uninstall (clean removal, project data untouched):

curl -fsSL https://raw.githubusercontent.com/cklxx/compound-engineering/main/uninstall.sh | bash

Cross-platform: macOS, Linux, Windows (WSL/MSYS/Cygwin).

High-resolution skills

All 5 skills rewritten with sharp boundaries and concrete examples:

Skill	Role	Key improvement
writing-plans	THINK before coding	Filled rate-limiter plan example, precise trigger/don't-trigger table
recording-practices	LOG raw events	Real nil-map-panic error entry + table-driven-tests good entry
managing-memory	PERSIST distilled knowledge	Pass/fail qualification table, verified+durable+actionable gate
running-retros	COMPRESS entries into rules	Pattern clustering algorithm, full retro report with real data
using-compound-engineering	ORCHESTRATE the loop	Skill interaction flowchart, "Does NOT do" column for each skill

Clear skill boundaries

recording-practices = raw event log     (high volume, low curation)
managing-memory     = distilled knowledge (low volume, high curation)
running-retros      = compression step   (entries → patterns → rules)

The Compound Loop

  Plan ──→ Code ──→ Record ──→ Remember ──→ Review
   ↑                                          │
   └──────── better plans, fewer mistakes ────┘

Session  1: memory = 0  → records 5 entries, saves 2 to memory
Session  5: memory = 8  → avoids 3 known pitfalls, discovers new patterns
Session 10: memory = 17 → retro distills 8 rules, prunes stale items
Session 50: memory = 40 → deep project expertise, rare mistakes, fast fixes

v1.0.0 — Compound Engineering

compound-engineering v1.0.0

Install once. Your AI coding agent automatically learns from every session — plans, records, remembers, reviews. Code quality compounds over time, on autopilot.

One-Line Install (Claude Code + Codex)

curl -fsSL https://raw.githubusercontent.com/cklxx/compound-engineering/main/install.sh | bash

One-Line Uninstall

curl -fsSL https://raw.githubusercontent.com/cklxx/compound-engineering/main/uninstall.sh | bash

Your project data (docs/) is never touched.

How It Works

5 auto-triggered skills that form a compound learning loop:

  Plan ──→ Code ──→ Record ──→ Remember ──→ Review
   ↑                                          │
   └──────── better plans, fewer mistakes ────┘

Skill	Role	Auto-triggers when
writing-plans	THINK before coding	Non-trivial work (3+ files, design choice, new feature)
recording-practices	LOG raw events	Bug fixed, pitfall hit, effective pattern discovered
managing-memory	PERSIST distilled knowledge	Verified + durable + actionable insight emerges
running-retros	COMPRESS entries into rules	10+ entries, milestone done, 7+ days elapsed
using-compound-engineering	ORCHESTRATE the loop	Every session start

The key distinction:

• recording-practices = raw event log (high volume, low curation) — "nil map panic in config loader"
• managing-memory = distilled knowledge (low volume, high curation) — "Always initialize maps in constructors"
• running-retros = the compression step — finds 3 entries with same root cause → creates a rule

The compound effect:

Session  1: memory = 0  → records 5 entries, saves 2 to memory
Session  5: memory = 8  → avoids 3 known pitfalls, discovers new patterns
Session 10: memory = 17 → retro distills 8 rules, prunes stale items
Session 50: memory = 40 → deep project expertise, rare mistakes, fast fixes

Each session reads from all prior sessions. Each retro compresses raw experience into rules. Rules make future sessions better. Compound interest applied to engineering.