Reduce context overhead: conductor/router skills instead of 254 always-loaded skills

[erikpr1994]

Problem

Every SKILL.md description loads into the agent's context at session start. PostHog/skills currently ships 254 skills (skills/posthog 157, skills/omnibus 85, skills/team 12). Even namespaced, this is a very large fixed token cost imposed on every session that installs the plugins — it measurably crowds out other tools and the user's own working context before any PostHog work happens.

Proposal

Collapse each cluster behind a single conductor/router skill (e.g. one posthog entry, one omnibus entry) whose description covers "what this can do," and which dispatches to the 100+ underlying procedures as on-demand reference files loaded only when a path is chosen — rather than 157 always-resident peers.

Existing examples of this conductor pattern:

• materialize — one conductor, ~20 on-demand modes from reference/<mode>/: https://github.com/erikpr1994/materialize
• improve — one entry skill, sub-capabilities pulled in only when invoked: https://github.com/shadcn/improve
• impeccable — same conductor approach: https://github.com/pbakaus/impeccable

Benefit

Always-loaded surface drops from ~254 descriptions to a handful of routers; full procedures remain, just lazily loaded. Cheaper, less crowding, same coverage.

---

Evidence (measured)

Claude Code's own per-plugin token accounting (~/.claude/plugins/plugin-catalog-cache.json, Opus-4-7) ranks the packaged posthog plugin (v1.1.32, 84 skills + 3 commands) as the most expensive of all 218 catalog plugins by always-on cost:

• always_on = 15,925 tokens — paid every session before any PostHog work. Add PostHog Debugger skill #1 / 218 plugins, 25.7× the median plugin (619 tokens), 1.8× the update to skills repo #2 (twilio, 8,752).
• on_invoke = 349,645 tokens (all skills/commands resolved).

A single conductor/router skill (e.g. materialize, improve, impeccable) costs ~170–300 always-on tokens for equivalent coverage, loading its procedures on demand — i.e. ~1–2% of the current standing cost. Full breakdown and a reproduction snippet in the comments.

[erikpr1994]

Hard numbers backing the always-on cost

Claude Code keeps a per-plugin token accounting in ~/.claude/plugins/plugin-catalog-cache.json. For the packaged posthog@claude-plugins-official plugin (v1.1.32, 84 skills + 3 commands), measured for claude-opus-4-7:

metric	posthog	context
always_on — loaded into every session before any PostHog work	15,925 tokens	#1 of 218 plugins in the catalog
vs. median plugin (619 tokens)	25.7×	mean across plugins: 1,285
vs. #2 plugin (twilio, 8,752)	1.8×
on_invoke — all skills/commands resolved	349,645 tokens

So enabling the plugin spends ~16k tokens of standing context per session — the single largest always-on cost of any plugin in the marketplace, ~26× a typical plugin. A conductor/router skill (materialize / improve / impeccable) costs ~170–300 tokens always-on for one entry point, loading its 80+ procedures on demand — the same coverage at ~1–2% of the standing cost.

Live cross-check from a session /context: the Skills block measured 29.9k tokens (7.5% of a 400k window), with PostHog the dominant contributor. The catalog also exposes a double-listing artifact — each skill appears both bare and posthog:-namespaced — inflating the apparent count (~150) vs. the real 84.

Reproduce locally:

python3 - <<'EOF'
import json, os
d = json.load(open(os.path.expanduser('~/.claude/plugins/plugin-catalog-cache.json')))
def walk(o):
    if isinstance(o, dict):
        if o.get('plugin') == 'posthog':
            print('tokens (opus-4-7):', o['tokens']['claude-opus-4-7'])
            print('skills:', len(o['components']['skills']), 'commands:', len(o['components']['commands']))
        for v in o.values(): walk(v)
    elif isinstance(o, list):
        [walk(v) for v in o]
walk(d)
EOF

(Figures are Opus-4-7; the same cache reports Sonnet-4 always_on at 10,828.)