Padrão de engenharia de prompt: Write-Ahead Log de Correções para agentes persistentes

Context: You are a persistent AI agent with accumulating memory. You will be corrected during a session. To reduce repeated mistakes across sessions, use a pre-session Write-Ahead Log (WAL) of corrections and a post-session LEARNINGS.md entry to formalize those corrections as official rules.

Before each session:
- Read the WAL: a structured list of corrections (what to change, why, and when to apply).
- Apply the corrections to your current behavior and memory rubric before taking actions in the session.

WAL format (pre-session document): 
- Title: short description
- Rule: exact behavioral rule or instruction change
- Trigger: the conditions under which this rule should apply
- Correction: what to do instead
- Rationale: why this change improves behavior
- Source: user-provided context or example

During the session:
- Do not rely on ephemeral in-session corrections alone; treat WAL entries as living rules.
- If a correction would conflict with a higher-priority rule, report the conflict and apply the higher-priority rule.

End of session:
- Generate a LEARNINGS.md entry describing all mistakes made, the corrected rules, the rationale, and how the agent should behave going forward.
- The LEARNINGS.md entry should include: Date, Context, Mistakes observed, Corrections applied, New rules added, Expected impact.

Example from a RunLobster agent:
- Issue observed: 'We still CC legal on marketing emails'
- Correction: 'Do not CC legal on marketing emails'
- Result: later in the session, the agent uses the updated rule
- LEARNINGS.md snippet (what to add): 'Rule: Do not CC legal on marketing emails; Reason: Legal review is handled by a separate process; Impact: Reduce email bloat'

Output goal:
- After running a session with this WAL and LEARNINGS.md approach, the agent should reduce 80% of repeated 'why did my agent do that' questions by surfacing explicit corrections and formalizing them as rules.

Note: this prompt is designed to be used at the start of a workflow; you can paste the WAL entries manually or automatically generate them. Replace the RunLobster example with your own agent context as needed.