randlee/coding-agent-hardening

Coding Agent Hardening

This skill is for editing existing agent prompts and orchestration instructions so they stop normalizing avoidable defects. It is not a general review skill. It is a prompt-hardening skill for agent behavior.

Scope

Use this skill when modifying prompts for:

• coding or implementation agents
• QA or review agents
• orchestration, triage, planning, or fix-routing agents

The goal is to eliminate default behaviors that allow issues to survive under labels like:

• minor
• non-blocking
• pre-existing
• technical debt
• needs direction, when no real decision exists
• changed code without validation
• regression reported without fix or root cause

Core Policy

Harden prompts toward a fix-first default:

• simple findings should be fixed, not escalated
• non-blocking findings should still be fixed when reasonable in the current pass
• pre-existing issues should not be ignored just because they predate the current change
• minor findings should not disappear into backlog language by default
• agents should exhaust reasonable alternatives before stopping for user direction
• technical debt should not be created as an escape hatch for small or moderate fixes
• code changes should be followed by relevant validation before the agent stops
• regressions should be fixed or root-caused before the agent stops

Human escalation is still allowed, but only for real decisions:

• major architectural trade-offs
• mutually exclusive product choices
• destructive or irreversible operations
• broad refactors whose blast radius materially changes schedule or scope

Agent Delegation

This skill operates directly in the current session. It does not delegate to background agents or Agent Runner.

References

• ./references/hardening-policy.md — normative rewrite policy for issue handling and escalation
• ./references/prompt-rewrite-patterns.md — concrete rewrite patterns to remove permissive wording
• ./references/agent-category-guidance.md — category-specific hardening rules for coding, QA, and orchestration agents
• ./references/repo-target-map.md — optional Synaptic Canvas repo target map; ignore outside this repository

Read ./references/hardening-policy.md first. Then read ./references/agent-category-guidance.md for the target agent type. Use ./references/prompt-rewrite-patterns.md while editing the actual prompt text. Read ./references/repo-target-map.md only when hardening Synaptic Canvas agents in this repository.

Workflow

When hardening an existing agent prompt:

1. Identify the agent category:

   • coding
   • QA/review
   • orchestration/planning

2. Read the applicable sections in:

   • ./references/hardening-policy.md
   • ./references/agent-category-guidance.md

3. Scan the target prompt for permissive patterns such as:

   • blocker-only language
   • confidence thresholds that suppress too many real findings
   • wording that asks before fixing obvious issues
   • "pre-existing" dismissal language
   • automatic deferral to technical debt or backlog
   • escalation on false decisions where a reasonable fix path exists

4. Rewrite the prompt so the agent:

   • fixes straightforward issues autonomously
   • reports all meaningful findings instead of filtering to a tiny subset
   • treats QA findings as remediation input, not merely triage metadata
   • keeps working until it hits a true decision or a real boundary
   • does not stop after code changes without running relevant validation
   • does not report regression without fixing it or root-causing it first

5. Preserve legitimate constraints:

   • safety boundaries
   • destructive-operation confirmations
   • scope limits that prevent major unplanned refactors

6. Before declaring the prompt hardened, verify that it no longer permits:

   • fake decisions or premature escalation
   • "minor" as an ignore/defer category
   • technical-debt escape hatches for reasonable fixes
   • pre-existing/not-worsened dismissal
   • stop-without-validation outcomes

Category Guidance

See ./references/agent-category-guidance.md for category-specific rewrite rules for coding, QA/review, and orchestration agents.

Output Expectations

When using this skill to update an agent prompt:

• state which permissive behaviors were removed
• state which replacement behaviors were added
• note any remaining escalation points and why they are real

Do not describe the prompt as hardened unless the old defer/ignore behaviors were explicitly removed.