More Codex skills can mean less useful skill context.

The catalog is bounded before a skill is invoked.

Current Codex source uses a configured token budget when present, otherwise two percent of the known model context window. When the window is unavailable it falls back to an 8,000-character budget. A configured value is capped at 10,000 tokens. Each catalog description is first capped at 1,024 characters.

If full lines do not fit, Codex distributes description space across entries. If even name-and-location lines exceed the budget, later entries can be omitted from the model-visible list. This is catalog visibility, not the size of a selected skill's full SKILL.md.

Find pressure before deciding what is relevant.

1. Audit both user roots and repository scopes

   instructree skills checks the shared ~/.agents/skills root, Codex's deprecated default ~/.codex/skills compatibility root, and repository-local scopes. It keeps their provenance distinct and excludes hidden system descendants.

2. Inspect the largest configured contributors

   The concise report ranks up to five enabled candidates by estimated fallback-character cost. It applies supported user configuration first, caps descriptions like current Codex, and retains logical paths so the report is share-safe.

3. Resolve duplicate names before adding more

   A same-name user and repository skill can create ambiguous expectations even when runtime precedence is deterministic. Inspect both sources and keep the intended copy; Instructree never chooses or deletes one for you.

4. Check a deliberately small starter stack

   Run instructree starter to compare six source-grounded engineering companions with local state. It covers discovery, testing, diagnosis, review, research, and verification, and prints install commands only for missing companions.

The ranking makes aggregate pressure reviewable.

signals
duplicate names: 2 · metadata failures: 0 · metadata warnings: 1
skill-list estimate: exceeds-unknown-window-reference · 12640/8000 chars

largest configured contributors · top 5 of 34
- research-suite · 1112 chars · description capped for estimate · ~/.agents/skills/research-suite/SKILL.md
- deployment-review · 884 chars · ~/.agents/skills/deployment-review/SKILL.md

High estimated cost is not proof that a skill is bad or unused. It is a place to start a relevance review. Use --json for deterministic structured data and inspect the actual SKILL.md before changing configuration.

Install for tasks, not for collection size.

The starter stack is intentionally small. Popularity is only a discovery filter; it does not establish quality, security, or fit. Review every external skill before installation because skill instructions influence agent decisions.

$ instructree starter .
focused stack · 6/6 ready
✓ find-skills
✓ tdd
✓ diagnosing-bugs
✓ code-review
✓ research
✓ verification-before-completion

See the Codex setup and selection rationale for the pinned sources and install commands.

The estimate follows current source while stating what it cannot know.

• Codex skill renderer at 9b4a0f8 — default and configured budgets, per-description cap, truncation, omission, and path aliasing.
• Codex skill configuration at 9b4a0f8 — user-facing budget and enablement fields.
• Official Codex skills documentation — skill discovery, installation, and configuration.
• openai/codex#19679 and #24299 — user reports requesting configurable or clearer skill metadata budgets.
• openai/codex#25324 — user demand for clearer same-name skill precedence.