debloat-contextlisted

# Debloat Context Audit the fixed per-session context cost of a Claude Code setup, report ranked findings, then apply approved trims one gated step at a time. Sized for a cheaper model (Sonnet tier, thinking off) — every judgment call below has an explicit rule so none is left to improvisation. **Model guidance:** the audit, report, settings, and incremental-trim steps run fine on Sonnet. The exception is a *first deep rewrite* of a large, dense CLAUDE.md (roughly 10 KB+ of interlocking rules): compression there risks silent semantic drift that a diff review can miss, so run that one step on a top-tier model (Opus class) — it is a one-off; later runs only trim increments. If staying on Sonnet anyway, trim in small sections across multiple gated steps, never one whole-file rewrite. **Core principle: measure, don't assume.** Numbers from a blog post, another machine, or a previous run are stale. Re-derive everything from this machine's files and transcripts, this session. ## 0. Baseline Ask the user to run `/context` and paste the output (it is a user command; the model cannot run it). Record the total and per-category breakdown — this is the "before" for step 4. If the user declines, proceed on file-size estimates and say the final delta will be estimated, not measured. Either way, resolve this **before the first edit in step 3** — a baseline captured after edits began is worthless, and the report must state which outcome happened. ## 1. Measure (read-only) **File sizes