Repository navigation
Description
Context
Menu item #4 of the Boris-talk digest: every always-loaded skill description is a per-turn tax, so the marketplace's full-install listing surface is itself an unhobbling concern. This is the report-only analysis; the target decision (digest questions.md #4) is the operator's and gates any remediation.
Current numbers, measured on main (19a3f7b3) with skill-quality's check-listing-budget.sh pooled over plugins/*/skills:
- 134 listing-eligible skills across 61 plugin roots (skills with
disable-model-invocation: trueexcluded — they spend no budget). - Aggregate: 86,499 chars ≈ 10.8x the 8,000-char documented default budget (
SLASH_COMMAND_TOOL_CHAR_BUDGETfallback, ~1% of context window). Live, Claude Code drops the least-invoked skills' descriptions to name-only when this overflows — silently degrading auto-invocation for the tail. - Advisory only: the live figure depends on the consumer's model window,
skillListingBudgetFraction, andskillOverrides;/doctorin a live session is the authority. The estimate is an upper bound for full-install consumers.
Top contributors (entry chars):
| chars | skill | root |
|---|---|---|
| 1468 | babysit-loop | plugins/source-control/skills |
| 1197 | audit-instructions | plugins/claude-config/skills |
| 1161 | audit-derivability | plugins/docs-hygiene/skills |
| 1116 | running-retro | plugins/session-flow/skills |
| 1072 | reconcile | plugins/session-flow/skills |
| 1058 | clarify | plugins/adhd/skills |
| 1057 | improve | plugins/architecture/skills |
| 1054 | script-the-deterministic-work | plugins/discipline/skills |
| 1028 | audit-answers | plugins/planning/skills |
| 1022 | reuse-or-replace | plugins/discipline/skills |
Context for the decision: a full install is ~86k chars ≈ 21–22k tokens of always-loaded listing text before any work starts. But installs are per-plugin selective — no consumer is forced to take all 61 roots — and per-skill entries are already capped at 1,536 chars by the skill-quality per-skill gate (none currently exceed it).
Proposed work — the operator picks one (questions.md #4)
- Accept-as-is with rationale: installs are selective; the per-skill 1,536 cap plus the advisory aggregate report is enough. Record the rationale here and close.
- Tighten the worst descriptions: mechanical pass over the top-N contributors (e.g. everything >1,000 chars — 10 skills) trimming toward the trigger phrases that earn their keep. No structural change.
- Curated starter bundles / per-plugin caps: define install sets or a hard per-plugin budget in doctrine. Structural; needs its own design pass.
A quarterly report-only "listing-budget watch" recurring routine is being wired separately (self-maintenance pilot) so drift is observed regardless of which option is chosen.
Acceptance criteria
- Operator has chosen among the three options above (or named a fourth) and the choice is recorded on this issue.
- If option 2/3: follow-up work items exist for the remediation; if option 1: rationale recorded and this issue closed.
References
- knowledge-corpus digest slice
boris-cherny-we-cut-80-of-claude-code-s-qyPCVqFUyDo—recommendations/menu.mdci: onboard to the CI platform (ci-workflows + standards) #4,recommendations/questions.mdci: onboard to the CI platform (ci-workflows + standards) #4,research/findings/repo-instruction-surface.md(baseline: 184 skills / ~28k tokens; live count now 134 listing-eligible afterdisable-model-invocationexclusions) - plugins/skill-quality/skills/check —
listing-budgetaction (the measurement tool; advisory by design) - docs/PLUGIN-PHILOSOPHY.md — "Instruction economy"
Activity
chore: pilot two report-only self-maintenance routines (#2025)
kyle-sexton commentedon Aug 9, 2026
Operator decision recorded (2026-08-08 decision round): option 2 executed now, option 3 pursued as follow-on — "make sure all adhere, focus on the worst first, maybe also do #3."
This issue closes when #2054 merges (acceptance criteria: decision recorded ✓, follow-up work items exist ✓).
refactor: tighten the 10 worst skill-listing descriptions toward the …
kyle-sexton commentedon Aug 9, 2026
Closing: acceptance criteria met. Option decision recorded above (option 2 + option 3); option 2 executed and merged (#2054 — ten entries trimmed, trigger phrases preserved, post-change aggregate 85,650 chars / ~10.7x advisory); option 3 tracked in #2034; the standing watch (#2023) discharged its first cycle with these numbers and observes drift going forward.
fix(ci): close the reverse direction of the changelog-parity gate (#2154