| name | icl |
| description | In-Context Learning — select and inject worked examples (demos) before executing recurring task types. Load when a task matches a known pattern family. |
Skill: In-Context Learning
Injecting 1–2 matched demos before a task measurably beats zero-shot on recurring patterns. The pool lives in .berserqir/skills-resources/icl/demos/ (hub, on-demand regime).
Demo resolutions
| Resolution | Content | Use for |
|---|
| task-level | full trajectory of a completed similar task (goal → steps → outcome) | complex multi-step tasks |
| step-level | snippet of one recurring step (e.g., "update memory-medium after task") | protocol steps agents get wrong |
Selection (before EXECUTE phase)
- Match by frontmatter tags (area, pattern family) + task similarity.
- Inject 1–2 demos maximum — more dilutes; budget lives in
core/protocols/context-budget.md.
- No match ≥ threshold? Inject nothing. A wrong demo is worse than zero-shot.
Demo format
---
tags: [area, pattern-family]
outcome: success | anti-example
source: manual | promoted-instinct | eval-failure
---
# Goal
# Trajectory (condensed ReAct: thought → action → observation)
# Outcome & verification
Anti-examples
Eval failures and recurring errors become outcome: anti-example demos — injected alongside a success demo when the task family has a known failure mode ("this looks right but fails because…").
Provenance
Demos enter the pool by: manual curation · instinct promotion (confidence ≥ threshold + eval pass — see plan F4) · eval failure capture. Every demo carries source.
