On-Demand Context

What it measures

Do you have skill files for recurring workflows, built from successful runs rather than written speculatively?

The four diagnostic questions

1. 01You use skill files so the model pulls detail only when needed, rather than loading everything at session start.
2. 02Skills are built from successful runs — you walk the model through a workflow first, then ask it to write the skill.
3. 03Skill files include real failure cases and error-handling patterns, not just the happy path.
4. 04You iterate skills after failures: feed the error back, fix the workflow, update the skill before the next run.

What good looks like

Mature skills discipline. The next lever is recursive refinement — skills should get tighter with each iteration.

Common failure mode

You're likely loading everything at session start. Build skill files from successful runs — not templates. Context on demand, not context always-on.

Critique & references

Editor's notes — to be expanded

The skill-as-needed pattern aligns with token-optimization research: concise, targeted context components measurably improve task accuracy versus monolithic always-on instructions. Pull, don't push.

• → Anthropic — Claude Skills
• → Token-Budget-Aware LLM Reasoning (arXiv 2412.18547)