| name | seed-evaluation |
| description | Use when asked whether to adopt a tool, library, plugin, skill collection, agent definition, hook library, or any external project as part of the Claude Code harness. Applies the foundation/03 two-stage methodology (pre-filter then deep-eval), produces a binary integrate / integrate-with-constraints / reject decision, and prevents evaluation theater by rejecting rubric scoring. |
Seed Evaluation
When this fires
The skill activates when the conversation proposes adopting any external project into the harness:
- "Should we add X library / plugin / skill?"
- "I found this repo, can we use it?"
- "Let's evaluate Y as a replacement for Z."
- "Has anyone integrated A with Claude Code?"
The skill does NOT fire for routine tool use of already-adopted dependencies, nor for one-off invocations that do not produce a persistent harness change.
Two stages
Stage 1: Pre-filter (target 30 seconds per candidate)
Three questions. Any "no" rejects.
- License: Compatible with this repo's MIT and with the candidate's intended use?
- Architecture support: Works on Mac (Apple Silicon), Jetson AGX Orin (ARM64 Linux), and Windows (x86_64)? Where it does not, is the gap acceptable for this component, or is a per-platform equivalent required?
- Maintainership: Commit in the last 90 days, and issue tracker shows real responses?
Output: one sentence of rationale per rejected candidate. Survivors move to Stage 2.
The 30-second target is real. If pre-filter on a candidate takes 10 minutes, the question framing is failing. Tighten the question, not the time budget.
Stage 2: Deep evaluation (integration, not scoring)
Survivors get wired into a sandboxed session and exercised against three workloads:
- Nominal task: A task the candidate is supposed to do well. Measures expected-case quality.
- Edge case: A task that exercises the failure mode the threat model worries about. Measures resilience.
- No-op interaction: The cost of having the candidate installed and idle. Measures cache footprint, startup latency, tool pool inflation.
Output per candidate in mac/evaluations/deep-eval.md:
### <candidate>
Stage 2 entry: Phase <N>
Date evaluated: YYYY-MM-DD
Decision: integrate / integrate-with-constraints / reject
Nominal task: <one paragraph>
Edge case: <one paragraph>
No-op cost: <one paragraph>
Constraints (if integrate-with-constraints):
- <constraint, with the hook or deny rule that enforces it>
Rationale: <one paragraph naming the failure mode prevented and the alternatives rejected>
Drift trigger: <upstream major version | security advisory | periodic review>
Version pin: <semver>
What is not in the methodology
No rubric scoring. Rubric scoring on harness seeds is mostly noise. The qualities that matter (cache behavior, interaction with the permission system, prompt injection resilience, drift over upgrades) do not score cleanly on a 1-to-5 scale.
No exhaustive matrix. A 15-column matrix comparing 20 seeds across every dimension is the visible artifact of evaluation theater. The output is a short narrative per survivor.
No vendor pitch trust. A vendor's claim that their product "works with Claude Code" is data, not evidence. Stage 2 integration is the evidence.
When the decision is ambiguous
Prefer rejection. The post-launch revision cadence is the appeal mechanism for rejected candidates. A rejected candidate can come back when the signals change. An adopted candidate that turns out wrong costs revisions to remove.
Adoption produces three artifacts (same commit)
- The wiring change (config, MCP server registration, skill copy, hook script).
- The rationale: commit message under "Why" or in the phase output. Names the failure mode the seed prevents and the alternative seeds rejected.
- The drift trigger: line in
mac/ARCHITECTURE.md (and equivalents) recording the seed's version pin and the next re-evaluation trigger.
Adoption is the start of a maintenance commitment, not the end of an evaluation.
Related artifacts
- Foundation:
foundation/03-seed-evaluation-methodology.md (the authoritative methodology)
- Foundation:
foundation/02-architectural-principles.md Principle 4 (stress every component against the current model)
- Worksheet:
mac/evaluations/pre-filter.md (Stage 1 log)
- Worksheet:
mac/evaluations/deep-eval.md (Stage 2 log)
- Related skill:
mcp-server-pre-trust-audit (the same discipline applied specifically to MCP servers)
- Owner: harness-engineering (Phase 4, 2026-05-11)