Prompt Engineering · ~7 min

The Re-read Protocol

Compaction frees context by summarising the conversation — and quietly paraphrases your instruction files in the process. The rules survive on disk; their precision doesn't. The fix is a targeted re-read.

When an agent compacts a long session, it summarises older turns to free context space. The summary preserves task state — what was done, what remains — but paraphrases the instruction-file references it carried, losing their precision. The agent works on, with degraded rule fidelity and no visible error signal (post-compaction-reread-protocol).

1 Survival is not compliance

The trap is assuming that because CLAUDE.md is reloaded after compaction, its rules still bind. Claude Code does reload it — the InstructionsLoaded hook fires with load_reason: "compact" — but reload alone does not reliably restore behavioural compliance (claude-code#14258).

2 Re-read, then confirm

The fix is to make the refresh explicit rather than trusting the reload. The minimal form is a prompt issued before resuming work:

The confirmation requirement is not decoration. Asking the agent to state the key rules it found forces the constraint set back into recent, high-attention context and gives you a visible signal that the re-read actually took — directive-only prompts restore less reliably than re-read-and-confirm.

3 Automate it on the compaction event

Relying on yourself to remember after every compaction is fragile. A SessionStart hook with matcher: "compact" fires only on compaction resumes, and its stdout is injected into the agent's context — so the re-read directive arrives automatically, exactly when the loss happened:

The mandatory, stop-and-confirm phrasing is intentional: a polite suggestion is treated as optional; an explicit directive is not. Note the right hook — PostCompact (added in v2.1.76) fires after compaction but has no context-injection capability, so it suits logging, not re-read. Use SessionStart.