| 1 | --- |
| 2 | name: "weekly-learning-loop" |
| 3 | description: "Make analysis improvements stick by loading agent identity, persisting learnings, and reinjecting them into the next cycle." |
| 4 | domain: "analysis-operations" |
| 5 | confidence: "high" |
| 6 | source: "recurring learnings in Bender, Farnsworth, Hermes, and Leela histories" |
| 7 | --- |
| 8 | |
| 9 | ## Context |
| 10 | |
| 11 | A weekly AI workflow only improves if lessons survive the run that produced them. The loop is incomplete when learnings are written down but never loaded back into the next prompt. |
| 12 | |
| 13 | ## Patterns |
| 14 | |
| 15 | - Load the correct agent identity before analysis or reskill work begins. |
| 16 | - Persist durable outcomes in history, wisdom, or skill files as part of the same operating cycle. |
| 17 | - Inject shared wisdom and skills into the next prompt so the model can act on prior learning. |
| 18 | - Keep fallback paths aligned with the same prompt contract so learnings apply across execution modes. |
| 19 | |
| 20 | ## Examples |
| 21 | |
| 22 | - Good: call Copilot CLI with the registered agent name, not a file path. |
| 23 | - Good: store a repeatable lesson in `.squad/skills/` once it shows up across multiple agents. |
| 24 | - Good: render prompts with shared wisdom and skill context so new runs inherit the last run's conclusions. |
| 25 | |
| 26 | ## Anti-Patterns |
| 27 | |
| 28 | - Treating a post-run note as sufficient when the next run never reads it. |
| 29 | - Letting the fallback path drift to a different output contract. |
| 30 | - Keeping important learnings only in a single dated history entry. |