Compounding Engineering

Triggers automatically when relevant; invoke directly with /compounding-engineering or ask, e.g. "turn this correction into a rule".

The system must improve between tasks, not just within them (Boris Cherny, creator of Claude Code). karpathy-guidelines governs the code you write; universal-loop governs one task's execution; this one governs what survives the task.

Core principle: a lesson that lives only in the conversation is discarded work. Write it into the system.

1. Compound Every Mistake

When the user corrects you — even mildly, even once — fix the issue, then propose a one-line rule for CLAUDE.md or the relevant skill. Same when you catch yourself repeating an error.

• "Mental note", "I'll remember", "self-enforce going forward" are forbidden — memory dies at /clear. A lesson not written into a file doesn't exist.
• One line, matching the file's style; mention the addition in one sentence, no ceremony.

Litmus: "Will this lesson survive /clear?"

2. Rule of Three → Automate

A manual step or correction recurring a 3rd time is a process gap, not a slip. Propose automation paired with a written rule:

• No-exceptions rule → deterministic guard (hook, lint, CI). Instructions are advisory; hooks are guaranteed.
• Judgment call → CLAUDE.md line, slash command, or skill.

Litmus: "Did I do this yesterday and the day before?"

3. Record Migration Boundaries

When old and new patterns coexist after your change (style, naming, API shape, framework), a verbal flag dies with the conversation. Record the boundary durably: a TODO in the project's tracking file, a CLAUDE.md note naming the canonical pattern, or an issue. Mixed states mislead humans and models. Scope stays surgical — record the boundary, don't expand the task.

Litmus: "Would a newcomer in a fresh session know which pattern is canonical and what remains to migrate?"

4. Fresh-Eyes Review

Self-review is biased toward code you just wrote. For significant changes (multi-file; auth, money, or data paths), after your own verification passes, have a fresh-context subagent review the diff against the stated criteria — it sees only the diff and the plan, not your reasoning. Tell it to report only correctness gaps, not style: chasing every finding produces over-engineering.

Litmus: "Has anything without my conversation history checked this?"

5. Context Economy

• Exploration → subagents; they read the files, you get the summary.
• After 2 fix attempts the user reported failed, the conversation is polluted with dead hypotheses. Don't patch a third time in place: restate the problem cleanly (fresh session or reset), carrying forward evidence, not hypotheses. Precedence: two-strike-debug's in-session re-diagnosis runs first; reset when re-diagnosis has also failed. (universal-loop's inner loop — iterating against your own checks — is different.)
• Every CLAUDE.md line must pass: "would removing it cause mistakes?" If not, cut it — bloat makes real rules get ignored.

Red Flags — stop and write it down

• "I'll remember that" / "mental note" / "self-enforce going forward"
• Third identical manual step without proposing automation
• Verbal-only mention of a mixed pattern state
• Declaring multi-file work done with no independent review
• Attempting a 3rd fix in a conversation full of failed diffs