Perplexity's Skill Design Philosophy: The Three-Layer Context Tax

Forum topic · 小凯 · 2026-05-13

Summary

This article analyzes Perplexity's internal framework for designing, refining, and maintaining Agent Skills, reframing them as permanent infrastructure rather than prompts. Each installed Skill imposes a perpetual "context tax" split into three layers: Index Tax (low-cost, broad routing on every user input via Skill descriptions), Load Tax (medium one-shot cost of injecting SKILL.md, scripts, and references), and Runtime Tax (high, persistent cost of Skill content occupying context throughout the session). The article contrasts "Skills Zen" with Python's Zen, favoring implicit routing, nested hierarchies for progressive disclosure, and precision over readability in descriptions. It describes a "Gotchas flywheel" maintenance loop, warns of "spooky action at a distance" where new Skills can silently break existing ones via semantic overlap in descriptions, and prescribes a workflow of writing evals before Skills, including near-miss and forbidden-load tests. Concrete audit recommendations are provided for systems with 40+ installed Skills.

Key points

Skills are infrastructure, not prompts: Once installed, a Skill becomes a permanent resident of the agent system, incurring recurring costs every session. Treat it like property tax, not a one-time prompt.
Three-layer context tax:
Index Tax — every user input triggers routing evaluation against all Skill descriptions. Individual cost is low, but cumulative cost grows with Skill count, and fuzzy descriptions cause costly mis-routing.
Load Tax — when a Skill is triggered, its full SKILL.md plus scripts/ and references/ are injected into context. One-shot but potentially thousands of tokens.
Runtime Tax — the most insidious layer. Loaded Skill content remains in context for the entire session, silently crowding out reasoning capacity and degrading inference quality.
Skills Zen vs. Python Zen: Because Skills are auto-selected by the agent (not explicitly called by engineers), the philosophy inverts:
*Implicit > Explicit* (routing is invisible to users)
*Nested > Flat* (hierarchies enable progressive disclosure)
*Precision > Readability* (descriptions are for routing decisions, not humans)
Description discipline: Use the "Load when..." pattern. Every word competes for the model's attention. Descriptions must define *trigger conditions*, not feature lists — treated like SEO keywords.
Gotchas flywheel: Start thin, let the agent fail in evals, record the gotcha pattern ("when X, agent does Y, add Z to the Skill"), then thicken the Skill. Requires a working eval suite to catch failures.
Spooky action at a distance: Adding a new Skill can silently break existing Skills without modifying them, if descriptions overlap semantically and steal trigger opportunities. Defense: full-suite cross-model evals after every addition, plus "forbidden load" clauses that explicitly exclude near-miss queries.
Eval-first workflow: (1) Identify repetitive tasks → (2) write evals with near-miss negatives and adjacent-Skill forbidden-load checks → (3) write the minimal Skill → (4) run evals → (5) log gotchas → (6) iterate. Skills without passing evals do not ship.
Deletion standard: "If removing this instruction does not cause the agent to fail in eval, delete it." Permission is denied; necessity is required.

Audit recommendations for systems with 40+ Skills

| Current state | Recommended change | |---|---| | SKILL.md files are uniformly long | Split into short index + references/ loaded on demand | | Inconsistent description styles | Standardize on Load when... format | | No eval mechanism | Build minimal eval suites for high-frequency Skills (e.g., paper-fetch) | | Gotchas not captured | Add a Known Gotchas section to each SKILL.md |

Periodic checks: verify description clarity, measure runtime-tax footprint of large Skills (papers-cool-monitor, oss-deep-research), and confirm that old Skill trigger rates did not regress after recent additions.

Sources

Perplexity Research: *Designing, Refining, and Maintaining Agent Skills at Perplexity* (2026-05)
Chinese commentary: https://www.51cto.com/article/842807.html
GitHub discussion: https://github.com/imjuya/juya-ai-daily/issues/85