| name | defending-code-customize-plan |
| description | Plan safe harness customization only. |
| version | 1.0.0 |
| author | Anthropic, ported by Hermes Agent |
| license | Apache-2.0 |
| platforms | ["linux","macos","windows"] |
| metadata | {"hermes":{"category":"security","tags":["security","harness-customization","planning"]}} |
Defending Code Customize Plan
When to Use
Use this skill when the user wants to adapt the reference harness idea to another language, target shape, vulnerability class, or detection signal, but only wants a Phase 1 planning artifact.
Prerequisites
- User authorization and a clear target domain.
- Optional local reference checkout or design notes.
delegate_task available for read-only source or documentation review with toolsets=["repo-read"].
- Parent file-writing capability for planning artifacts only.
How to Run
Interview the user, inspect only local docs or source with read-only children, and write a customization plan. Do not validate the plan by executing target code.
Quick Reference
- Output:
CUSTOMIZE-PLAN.md
- read-only review:
delegate_task(..., toolsets=["repo-read"], role="leaf")
- Planning axes: target shape, input modality, detection signal, isolation boundary, dedupe signature, report format, patch-verification signal
- Execution handoff: Phase 2 operator workflow or a separately approved implementation task
Procedure
- Ask what the user wants to find, who will run the workflow, and what output matters.
- Ask targeted follow-ups for missing axes:
- Vulnerability class.
- Target shape.
- Detection signal.
- Proof-of-concept shape.
- Isolation boundary.
- Dedupe signature.
- Report structure.
- Patch verification signal.
- If a local reference checkout or existing target exists, use a read-only child to inspect it:
delegate_task(
goal="Map which local harness or target files appear domain-specific and which are generic.",
context="Path: <path>. Static planning only. Do not modify files or execute code.",
toolsets=["repo-read"],
role="leaf",
)
- Create
CUSTOMIZE-PLAN.md with:
- Goal and scope.
- Domain assumptions.
- Required inputs and target layout.
- Detection and grading oracle.
- Reporting schema.
- Files or prompts likely to change.
- Safety requirements.
- Unknowns and validation plan for a later phase.
- Stop at the plan. If the user asks for implementation or validation, explain that it is outside Phase 1.
Pitfalls
- Do not rewrite harness code from this skill.
- Do not validate by running sample targets.
- Do not weaken sandbox assumptions. If execution will happen later, preserve the requirement for a real isolation boundary and constrained egress.
- Do not promise that a static-only plan proves the customized workflow works.
Verification
CUSTOMIZE-PLAN.md exists and is explicitly labeled planning-only.
- Any local inspection uses
toolsets=["repo-read"].
- The plan lists unresolved validation work instead of claiming it has been performed.