improve-agent-native

Improve Agent Native

Produce a file-evidence scorecard for whether AI coding agents can follow repo guidance, use relevant product or workflow capabilities, run the right checks, recover from failure, and leave useful proof.

When To Use

• The user asks if a repo, agent-facing product surface, MCP server, autonomous workflow, Claude/Copilot/Codex setup, or AI-native app is ready for agent-native work.
• The user asks to audit repo agent compatibility, AGENTS.md quality, docs quality, missing test/build commands, proof loops, command evidence, action parity, tool design, dynamic context injection, or outcome testing.
• A repo needs a scored gap list, specific fixes, or a keep/move/delete guidance review.
• Agents keep drifting, skipping required evidence, using the wrong workflow, or needing the same correction.

Do not use it for broad architecture rewrites, enterprise process design, or implementation work unless the user explicitly asks to patch the repo after the audit.

Inputs

• Target repository path or diff.
• Whether the user wants a scorecard, recommendations only, or patch work after the audit.
• Repo-local guidance and validation entrypoints when present.

Resolve the target, expected artifact, and edit authority from the user request, current working directory, repo-local instructions, and existing task context before asking. Ask one plain-language question only when the missing answer cannot be discovered safely and would materially change the audit.

Outputs

Preserve the user's requested comparison or reporting lanes. When the user does not specify an artifact shape, return the following scorecard schema. Whether the result is prose, a table, or YAML, preserve the target, evidence-backed strengths, severity-ranked gaps, smallest durable next moves, exact validation evidence, and residual risk.

schema_version: 1
target_repo: <path or name>
score: <0-100 or no-score with reason>
working:
  - dimension: <context_routing|command_discovery|durable_repo_knowledge|autonomous_execution_loop|capability_parity_and_tool_design|mechanical_guardrails|proof_of_work|recovery_and_safety|feedback_to_harness_compounding>
    finding: <repo, workflow, or product strength>
    evidence: <file path, command, or blocker>
gaps:
  - severity: high|medium|low
    dimension: <same dimension enum>
    failure_category: <missing_validation|claim_boundary|proof_gap|scope_control|context_routing|safety_boundary|not_applicable>
    finding: <agent-readiness gap>
    evidence: <file path, command, or blocker>
    next_move: <smallest durable guardrail>
validation_evidence:
  - command: <exact command or not-run reason>
    outcome: pass|fail|blocked
    attempts: <optional ordered list when the command was retried or reshaped>
    diagnostic: <optional failure class and why the final command worked>
residual_risk:
  - <what the audit does not prove>