Implement Agent Reflection

Common Contract (Load First)

When this skill is activated, first load skill-common and apply the shared conventions it defines: canonical section order, frontmatter rules, anti-pattern guidance, and the Review In Mind loop in review-in-mind. This skill adds only domain-specific rules below.

Who

Architects, principal developers, or coding agents designing self-correcting workflows for autonomous systems.

What

Implement an Evaluator-Generator pattern where an initial output is systematically critiqued by a distinct reflection node, yielding actionable feedback that is fed back into the generator for iterative refinement.

When

• Triggered by requests like: "add reflection to this agent", "make this prompt self-correcting", or "implement an evaluator-generator loop".
• Do not use for simple deterministic tasks (like basic data formatting) where a single API call is sufficient.

Where

Applies to LangChain/LangGraph orchestrations, Google ADK topologies, or custom Python agent frameworks.

Why

Large Language Models excel at critiquing content. By explicitly asking an agent (or a separate model) to critique an initial draft against strict acceptance criteria, the system can self-correct logical errors, hallucinations, and formatting mistakes before the user ever sees them.

Inputs

• Base Generator Prompt: The original instructions for generating the artifact.
• Evaluation Criteria: The strict rubric or rules the critique node must enforce.
• Max Iterations Limit: A hard limit (e.g., 3) to prevent infinite looping.

Output (Logical Evidence)

• Reflective Orchestration Code: LangGraph or ADK logic containing distinct generate and reflect nodes.
• State Schema: A schema including a critique field and a retry_count integer.

Optimization Readiness

• Failure Signals: Reflection loops become open-ended, critique prompts collapse into generic feedback, generator and reflector share the same role framing, or revisions do not measurably improve output quality.
• Evidence To Collect: Critique logs, retry counts, before/after drafts, rubric results, and examples where reflection caught or missed specific defects.
• Safe Mutation Boundaries: Refine rubric prompts, retry policy wording, state-schema details, and validation examples without changing the core generator-plus-reflector loop.
• Acceptance Criteria: Accept revisions only if the workflow enforces a hard iteration limit, uses distinct prompts, and produces drafts that improve against explicit evaluation criteria.
• Rejected Revision Handling: Record vague critique styles, ineffective retry behaviors, and rubric gaps so they are not reused blindly.
• Transfer Check: Confirm the workflow still works for both formatting corrections and deeper factual or logical critique scenarios.
• Stop Rule: If the evaluation rubric or retry ceiling is missing, stop and ask before building a reflection loop.

Constraints (Logical Boundaries)

• Hard Iteration Limit: The workflow MUST include a strictly enforced loop limit. If the retry_count hits the limit, the loop must break and return the best effort.
• Separate Prompts: The generation and reflection phases must use explicitly different system prompts.

One More Thing

If the user has not provided the explicit Evaluation Criteria (the "rubric"), stop and ask them what the reflection node should be looking for.

How (Structural Workflow)

1. Define the State Schema

• Define a typed state dictionary (or Pydantic model) that holds the draft_output, the critique_feedback, and a retry_count.

2. Implement the Generator Node

• Write the node that takes the original prompt and any existing critique_feedback.
• If critique_feedback exists, the generator must explicitly address the critique in its new draft.

3. Implement the Reflection Node

• Write a distinct node that acts as a harsh, objective reviewer.
• Provide the Reflection node with the strict Evaluation Criteria.
• The Reflection node must output two things:
  1. A boolean flag is_acceptable.
  2. A string critique_feedback containing specific, actionable directives for fixing flaws.

4. Wire the Conditional Edge

• Route the Generator to the Reflection node.
• Route the Reflection node:
  • If is_acceptable == true, route to __END__.
  • If retry_count >= MAX_RETRIES, route to __END__ (or a fallback handler).
  • Otherwise, increment retry_count and route back to the Generator.

5. Validate the Loop

• Ensure the state schema is correctly passed and mutated at each step.
• Verify that the iteration limit is not off-by-one.

Review In Mind (ReviewInMindGenie)

Execute the common review loop in review-in-mind before delivering.

Review lens for this skill:

• Is the evaluator distinct from the generator, with specific and actionable critique feedback?
• Are conditional edges and retry limits correct (no off-by-one, no infinite loop)?
• Would a genuinely bad output fail the evaluator and produce usable feedback?