Doctor report: https://github.com/anthropics/claude-for-legal/tree/main/commercial-legal/skills/nda-review

[conikeec]

Public GitHub skill URL

https://github.com/anthropics/claude-for-legal/tree/main/commercial-legal/skills/nda-review

Notes

A sample run of claude's legal plugin based nds review

[github-actions]

SkillSpec Doctor report

Target: https://github.com/anthropics/claude-for-legal/tree/main/commercial-legal/skills/nda-review

Shape: simple_skill - one atomic SKILL.md package; detailed doctor analysis can run

Status: full

Remote: https://github.com/anthropics/claude-for-legal.git

Primary skill: SKILL.md

What This Measures

Doctor estimates how likely an agent is to miss, reorder, improvise, use the wrong surface, or finish without proof because of the current skill shape.

This is a baseline of the current skill shape at doctor time. It is not a grade of domain knowledge, human usefulness, author effort, or legal/medical correctness.

Higher risk means a higher chance the agent skips, reorders, or improvises load-bearing instructions.

Current Skill Baseline

• Verdict: critical reliability debt
• Follow-through readiness: very low
• Agent follow-through risk: critical (100/100, single-skill drift)
• Discovery risk: low (0/100)
• Raw activation load risk: critical (100/100 before contract mitigation)

Surface

• Activation body: 304 line(s), 19974 byte(s), approximately 4994 token(s)
• Deferred resources: 0 file(s), 0 byte(s)
• Frontmatter: 8 line(s), 314 byte(s)
• Activation-loaded surface: 98%
• Unmapped package files: 0
• Instruction signals: 12 modal obligation(s), 5 late obligation(s), 5 numbered step(s), 3 dependency mention(s)
• Code in active skill: 4 block(s), 1 unlabeled

Findings

1. HIGH: Most package text loads at activation

Location: https://github.com/anthropics/claude-for-legal/tree/main/commercial-legal/skills/nda-review/SKILL.md

Evidence: 98% of loaded text surface is in SKILL.md activation body; little is deferred behind task-specific references.

Basis: reliability_gap_metadata_context_pressure, reliability_gap_instruction_density

Fix: Move long references, examples, and code into referenced files and keep the activation body as a compact router.

2. HIGH: Dense load-bearing prose

Location: https://github.com/anthropics/claude-for-legal/tree/main/commercial-legal/skills/nda-review/SKILL.md

Evidence: Found 12 modal obligation spans and 5 numbered steps in the activation body.

Basis: reliability_gap_instruction_density, contract_trace_activation_adherence_enforcement

Fix: Promote route choices, forbids, elicitations, dependencies, and tests into skill.spec.yml so they can be checked instead of remembered.

3. HIGH: Operational prose lacks a structured execution contract

Location: https://github.com/anthropics/claude-for-legal/tree/main/commercial-legal/skills/nda-review/SKILL.md

Evidence: The activation body tells the model to use/run/create/fetch/click/install or similar, but there is no SkillSpec route, tool boundary, command template, or trace vocabulary.

Basis: contract_trace_activation_adherence_enforcement, contract_trace_unproven_verdict

Fix: Add a SkillSpec contract with routes, phase tool boundaries, commands, scenario tests, and trace requirements.

4. HIGH: Dependencies are implicit

Location: https://github.com/anthropics/claude-for-legal/tree/main/commercial-legal/skills/nda-review/

Evidence: Detected 3 dependency mention(s), 0 code file(s), 0 manifest file(s), and 4 code block(s), but no deps.toml ledger.

Basis: reliability_gap_implicit_environment_contract

Fix: Create deps.toml and preserve dependency authority, local status, install risk, and degraded proof impact before proof or install.

5. HIGH: No machine-checkable behavior contract

Location: https://github.com/anthropics/claude-for-legal/tree/main/commercial-legal/skills/nda-review/

Evidence: No skill.spec.yml was found, so route choices, forbids, tool boundaries, dependency checks, scenario tests, and trace expectations are not falsifiable.

Basis: reliability_gap_unfilled_requirement, contract_trace_behavioral_contract

Fix: Run skillspec source map, then skillspec import-skill, and complete the generated contract before install or proof.

6. MEDIUM: Late load-bearing instructions are exposed to primacy bias

Location: https://github.com/anthropics/claude-for-legal/tree/main/commercial-legal/skills/nda-review/SKILL.md

Evidence: 5 modal obligation span(s) appear after the first 60% of the activation body.

Basis: reliability_gap_instruction_density

Fix: Move late obligations into earlier route/rule summaries or structured checks; do not rely on a model remembering buried instructions.

7. MEDIUM: Code is mixed into the activation instruction body

Location: https://github.com/anthropics/claude-for-legal/tree/main/commercial-legal/skills/nda-review/SKILL.md

Evidence: Found 4 fenced code block(s) in SKILL.md, accounting for about 15% of activation bytes.

Basis: reliability_gap_no_execution_guarantees, contract_trace_static_well_formedness

Fix: Move executable code into scripts/resources or structured code entries and state whether snippets are executable, examples, or reference material.

8. MEDIUM: Code fence language is ambiguous

Location: https://github.com/anthropics/claude-for-legal/tree/main/commercial-legal/skills/nda-review/SKILL.md

Evidence: 1 code fence(s) in SKILL.md omit a language label.

Basis: reliability_gap_no_execution_guarantees

Fix: Label code fences and classify each one as executable code, command example, or non-executable reference.

1 more finding(s) omitted. Use --json or --html for the full report.

Next Actions

Recommended next action

/skillspec import https://github.com/anthropics/claude-for-legal/tree/main/commercial-legal/skills/nda-review/, compile it for <target>, install it, and prove it

• Capture this baseline before changing the skill: skillspec doctor https://github.com/anthropics/claude-for-legal/tree/main/commercial-legal/skills/nda-review/ --markdown > docs/skillspec-doctor-baseline.md.
• Install the skillspec skill into your harness if it is not already installed.
• From the harness, ask /skillspec import https://github.com/anthropics/claude-for-legal/tree/main/commercial-legal/skills/nda-review/, compile it for <target>, verify it, test it, and prove it. Print the alignment summary.
• Read the alignment summary before trusting the port: decision replay should pass, required steps should be proven, missing proof should be explicit, and forbidden actions should show no violations.
• Optionally publish the baseline doctor report, generated skill.spec.yml, compiled loader, and alignment report with the source repo or PR.
• Restart the harness after install, then try the SkillSpec-backed skill normally on a real task.
• For a CLI-only port, run skillspec source map https://github.com/anthropics/claude-for-legal/tree/main/commercial-legal/skills/nda-review/ --out <draft-dir>/.skillspec/source-map, then skillspec import-skill https://github.com/anthropics/claude-for-legal/tree/main/commercial-legal/skills/nda-review/ --out <draft-dir>/skill.spec.yml --source-map <draft-dir>/.skillspec/source-map/source-map.json.
• Create or complete deps.toml; preserve dependency authority, local status, install risk, and degraded proof impact.
• Promote operational prose into routes, rules, forbids, command templates, scenario tests, and trace/progress proof obligations.

Research Basis

17 cited basis item(s) are attached to this report.

• reliability_gap_instruction_density - Large, dense activation bodies increase dropped-step risk and later instructions suffer primacy bias. (SkillSpec Skills Reliability Gap)
• reliability_gap_metadata_context_pressure - Skill frontmatter metadata and activated bodies consume context; broad surfaces degrade routing and adherence. (SkillSpec Skills Reliability Gap)
• reliability_gap_implicit_environment_contract - Scripts and snippets carry dependency contracts whether or not the skill declares them. (SkillSpec Skills Reliability Gap)
• reliability_gap_no_execution_guarantees - Prose instructions and embedded snippets are guidance unless moved into checkable or enforced surfaces. (SkillSpec Skills Reliability Gap)
• reliability_gap_unfilled_requirement - Reliable skills need a portable, machine-checkable account of intended behavior, dependencies, and proof. (SkillSpec Skills Reliability Gap)
• contract_trace_activation_adherence_enforcement - Activation, adherence, and enforcement are separate gates; static prose mostly influences adherence, not enforcement. (SkillSpec Contract Trace Methodology)
• 11 more basis item(s) omitted. Use --json for the full basis registry.

Use --html for a shareable review page or --json for machine evidence.

Full artifacts: https://github.com/modiqo/skillspec/actions/runs/28296856919/artifacts/7926718939

The artifact contains:

• doctor-report.txt
• doctor-report.md
• doctor-report.html
• doctor-report.json
• target.txt