Conciseness

The body is efficient — a routing table, exact commands, and no explanation of concepts Claude already knows — but minor over-explanation could be trimmed: the "apply" internals ("content-addressed backup", "atomic same-directory replace") describe script behavior Claude does not need in order to act, and the opener "without pretending every model supports one million tokens" is rhetorical. It is not 3 because there is no genuinely unnecessary explanation, and not 5 because those few sentences do not strictly earn their place.

Actionability

Copy-paste-ready "uv run --no-project python "$SKILL_DIR/scripts/codex_context_window.py" doctor/apply/verify" blocks for both macOS/Linux and Windows PowerShell, an intent-to-action routing table, an explicit path-resolution procedure, and an exact report checklist. It is not 4 because the guidance fully covers the common cases with executable commands rather than hints.

Workflow Clarity

The sequence is explicit with validation checkpoints and feedback loops: "doctor" runs first and is read-only, "Run only after `doctor` succeeds", "a failure restores the exact prior bytes" on strict-config diagnostics, and "verify" exits non-zero on mismatch so drift is caught. This is a config-modifying operation, but validation is present, so the destructive-operation cap does not apply. It is not 4 because every checkpoint — pre-check, rollback, post-write read-back, and re-verification — is explicit.

Progressive Disclosure

The body is a clear overview (routing, commands, reporting, scope refusals) with one well-signaled, one-level-deep reference — "Read [references/context_window_contract.md] … when diagnosing model caps, the 95% usable-window display, or compaction semantics" — and both referenced bundle paths (references/context_window_contract.md, scripts/codex_context_window.py) exist as real files. Deep contract details and source links are appropriately split into the reference rather than inlined. It is not 4 because the split is clean and the reference is clearly signaled with its when-to-read condition.

Specificity

"Configures and verifies an expanded, model-aware context window" plus the named keys "model_context_window/model_auto_compact_token_limit" give two concrete actions with named targets — base 4 — but "you want 500K or 1M tokens" is second person, so the rubric's third-person penalty reduces it to 3. It is not 2 because the domain and actions are concrete, not generic; it does not reach 4/5 because the second-person penalty applies and the action list (two verbs) is not comprehensive.

Completeness

It explicitly answers what ("Configures and verifies an expanded, model-aware context window for OpenAI Codex CLI and Desktop") and when ("Use when Codex shows ~258K context, you want 500K or 1M tokens, auto-compaction fires too often, or model_context_window/model_auto_compact_token_limit needs repair") with concrete trigger phrases, plus a negative-scope clause. It is not 4 because both halves are explicit and specific, not merely present.

Trigger Term Quality

"Codex shows ~258K context", "500K or 1M tokens", "auto-compaction fires too often", and the raw key names cover the exact natural phrases a user would say for this niche, matching the comprehensive-synonyms anchor. It is not 4 because no common variation of this request (size symptom, desired amount, compaction behavior, key repair) is missing.

Distinctiveness Conflict Risk

"Not for changing the selected model or other Codex settings" plus the narrow Codex context-window focus give a clear niche with distinct triggers and minimal conflict with other skills. It is not 4 because the exclusion clause removes even the minor overlap with adjacent Codex-settings skills.