Context Engineering Fundamentals: Plugin Evaluation & Recommendations

[lsetiawan]

Context Engineering Fundamentals Evaluation: All Plugins

A comprehensive evaluation of all four plugins against context engineering fundamentals best practices. This issue documents findings and prioritized recommendations for improving context efficiency, progressive disclosure, and agent selection accuracy.

Evaluation Framework

Each plugin was evaluated against these core context engineering principles:

1. Context as Finite Resource — Smallest high-signal token set
2. Progressive Disclosure — Load information only when needed
3. Attention Positioning — Critical info at beginning/end
4. Clear Section Boundaries — Organized with structural clarity
5. Signal-to-Noise Ratio — Informativity over exhaustiveness
6. Redundancy — Avoid repeating the same information
7. Description Quality — Lightweight identifiers for agent selection

---

Plugin 1: scientific-python-development

Agents

scientific-python-expert.md (~325 lines)

Principle	Rating	Notes
Finite Resource	GOOD	Reasonable size, mostly actionable
Progressive Disclosure	GOOD	Uses 4 skills for deferred loading
Attention Positioning	GOOD	Purpose at top, completion criteria at bottom
Section Boundaries	GOOD	XML tags + Markdown headers
Signal-to-Noise	MODERATE	See findings below
Redundancy	MODERATE	See findings below
Description Quality	GOOD	4 <example> blocks with <commentary>

Findings:

• Lines 130-166 ("Scientific Python Process Principles"): ~36 lines of philosophical guidance ("Collaborate", "Don't Be Afraid to Refactor", "Prefer Wide Over Deep"). Low-signal for model decision-making — reads like a mission statement rather than actionable instructions. Recommendation: Condense to 5-8 bullet points of actionable principles, or move to a reference file.
• Lines 229-246 (Steps 5-7 of Response Approach): "Optimize for Reusability", "Document Thoroughly", "Enable Collaboration" largely repeat the principles and behavioral traits stated earlier. Recommendation: Remove or consolidate.
• Lines 315-323 ("Quality Assurance"): Repeats concepts from the self-review checklist at lines 203-227. Recommendation: Remove this section; the self-review checklist already serves this purpose.

scientific-docs-architect.md (~320 lines)

Principle	Rating	Notes
Finite Resource	GOOD	Well-scoped
Progressive Disclosure	MODERATE	No explicit skill references despite scientific-documentation skill existing
Attention Positioning	GOOD	"Remember" section at end is excellent placement
Section Boundaries	GOOD	Clean phase-based workflow
Signal-to-Noise	GOOD	Most content is actionable
Redundancy	MODERATE	Workflow Phases overlap with Workflow Patterns
Description Quality	GOOD	4 examples with commentary

Findings:

• No skills declared: This agent doesn't reference the scientific-documentation skill, meaning all documentation knowledge must live in the agent body. Recommendation: Add scientific-documentation as a skill so detailed reference content loads on demand.
• Lines 99-155 (Documentation Workflow Phases 1-4): Overlap with the "Workflow Patterns" section at lines 63-76. Recommendation: Keep one structure, not both.

Skills

Overall pattern is GOOD — Skills use progressive disclosure well with Quick Reference Cards at the top, references to references/ directories, and asset templates.

pixi-package-manager/SKILL.md (~333 lines):

• Lines 317-333 ("Summary"): 16-line paragraph repeats the core concepts and benefits already stated throughout the document. Ends with marketing-style call to action. Recommendation: Remove the Summary section entirely.

---

Plugin 2: project-management

Agents

project-onboarding-specialist.md (~194 lines)

Principle	Rating	Notes
Finite Resource	GOOD	Concise and focused
Progressive Disclosure	GOOD	Uses 2 skills
Attention Positioning	GOOD	Gap analysis early, completion criteria at end
Section Boundaries	GOOD	Clean XML + Markdown structure
Signal-to-Noise	GOOD	Almost entirely actionable
Redundancy	GOOD	Minimal repetition
Description Quality	POOR	No <example> blocks

documentation-validator.md (~160 lines)

Principle	Rating	Notes
Finite Resource	EXCELLENT	Most concise agent — everything is signal
Progressive Disclosure	GOOD	Uses 2 skills
Attention Positioning	GOOD	Priorities at beginning, completion at end
Section Boundaries	GOOD	Clean structure
Signal-to-Noise	EXCELLENT	No filler content
Redundancy	GOOD	No repetition
Description Quality	POOR	No <example> blocks

Critical Finding — Missing Examples in Descriptions: Both agents use plain-text descriptions without <example> blocks, which significantly reduces triggering signal for agent selection compared to agents in other plugins that include structured examples.

---

Plugin 3: scientific-domain-applications

Agent

astronomy-astrophysics-expert.md (~1,006 lines, est. ~15,000+ tokens)

Principle	Rating	Notes
Finite Resource	FAILS	Massively oversized — largest agent by 3x
Progressive Disclosure	FAILS	No skills referenced; all knowledge embedded
Attention Positioning	MODERATE	Good at top, critical framework buried in middle
Section Boundaries	GOOD	Well-organized headers and XML tags
Signal-to-Noise	LOW	60-70% of content is reference material, not instructions
Redundancy	MODERATE	Code examples repeat concepts from bullet lists
Description Quality	GOOD	4 examples with commentary

This is the most significant context engineering concern across all plugins.

Detailed findings:

1. Lines 114-350 ("Capabilities" section, ~236 lines): Exhaustive enumeration of astronomical knowledge domains. Most of these are knowledge the model already has — listing "Blackbody radiation and Wien's law" or "HR diagram interpretation" doesn't improve the model's ability to use that knowledge. This is the textbook example of exhaustiveness over informativity.

2. Lines 539-791 ("High-Precision Timing Systems", ~252 lines): Detailed reference tables, 7 code examples, and time system hierarchy documentation. This is reference material that should live in a references/ directory and be loaded only when timing-related tasks are encountered.

3. Lines 885-948 ("Common Astronomical Calculations"): Four code snippets that are reference material, not behavioral instructions.

4. Lines 504-835 ("Specialized Knowledge Areas"): More reference content that reads like a textbook rather than agent instructions.

5. No skills: Unlike every other agent in the collection, this agent declares no skills despite astropy-fundamentals and xarray-for-multidimensional-data existing in the same plugin.

Estimated token savings: Moving reference content to skills would reduce the agent from ~15,000 tokens to ~5,000 tokens while improving quality through progressive disclosure.

---

Plugin 4: ai-research-workflows

Agent

research-workflow-orchestrator.md (~430 lines)

Principle	Rating	Notes
Finite Resource	MODERATE	Slightly large but justified by scope
Progressive Disclosure	GOOD	Uses research-workflow-management skill
Attention Positioning	GOOD	"Remember" section at end
Section Boundaries	GOOD	Clear organization by workflow phase
Signal-to-Noise	MODERATE	Some redundancy
Redundancy	MODERATE	See findings
Description Quality	MODERATE	No examples in description frontmatter

Findings:

1. Lines 366-407 ("Common User Scenarios"): Repeats the decision patterns already described in lines 111-198 ("Core Decision-Making Framework"). Recommendation: Remove or merge unique patterns into the decision framework.
2. Description lacks examples: Examples at lines 10-44 are in the body, not the description frontmatter. For triggering accuracy, examples should also be in the description. Recommendation: Move examples into the description field.
3. Lines 204-269 ("Key Preferences"): "Iterative Approach" and "Evidence-Based Decisions" subsections are generic software engineering advice. Recommendation: Trim to workflow-specific preferences only.

---

Cross-Plugin Findings

1. Inconsistent Description Patterns (HIGH PRIORITY)

Agent	Examples in Description?	Pattern
scientific-python-expert	Yes	<example> blocks in multi-line description
scientific-docs-architect	Yes	<example> blocks in multi-line description
astronomy-astrophysics-expert	Yes	<example> blocks in multi-line description
project-onboarding-specialist	No	Plain text + metadata block
documentation-validator	No	Plain text + metadata block
research-workflow-orchestrator	No	Plain text, examples in body only

Impact: Agents without description examples are at a disadvantage for being selected at the right time. The description is the lightweight identifier that informs agent selection.

2. Progressive Disclosure Gaps

Plugin	Skills Used	Assessment
scientific-python-development	4 skills with references/	GOOD
project-management	2 skills with references/	GOOD
scientific-domain-applications	2 skills exist BUT agent doesn't reference them	POOR
ai-research-workflows	1 skill with assets/	GOOD

3. Token Budget Estimates

Agent	Est. Lines	Est. Tokens	Assessment
documentation-validator	160	~3,000	Excellent
project-onboarding-specialist	194	~3,500	Good
scientific-docs-architect	320	~5,500	Good
scientific-python-expert	325	~6,000	Good
research-workflow-orchestrator	430	~7,000	Acceptable
astronomy-astrophysics-expert	1,006	~15,000	Needs reduction

---

Priority Recommendations

P0 — Critical (Context Budget Violations)

• Refactor astronomy-astrophysics-expert.md: Move ~600 lines of reference content to skills. Add astropy-fundamentals (already exists) and create a new astronomical-timing skill. Reduce the agent to ~350 lines of behavioral instructions.

P1 — High (Progressive Disclosure Gaps)

• Add <example> blocks to project-onboarding-specialist description: 3-4 examples with user/assistant/commentary to improve triggering accuracy.
• Add <example> blocks to documentation-validator description: Same treatment.
• Add <example> blocks to research-workflow-orchestrator description: Move body examples into the description field.
• Link skills to astronomy-astrophysics-expert: Add astropy-fundamentals and xarray-for-multidimensional-data to the agent's skills list.
• Link scientific-documentation skill to scientific-docs-architect: Enable progressive disclosure of documentation reference content.

P2 — Moderate (Redundancy Reduction)

• Remove Summary section from pixi-package-manager/SKILL.md: Lines 317-333 repeat earlier content.
• Consolidate "Scientific Python Process Principles" in scientific-python-expert.md: Reduce lines 130-166 to 5-8 actionable bullets.
• Remove "Common User Scenarios" from research-workflow-orchestrator.md: Redundant with the decision-making framework.
• Remove "Quality Assurance" section from scientific-python-expert.md: Redundant with the self-review checklist.

P3 — Low (Polish)

• Deduplicate Response Approach steps 5-7 in scientific-python-expert.md: Consolidate with earlier sections.
• Trim generic preferences in research-workflow-orchestrator.md: "Iterative Approach" and "Evidence-Based Decisions" are not workflow-specific.

---

Summary Scorecard

Plugin	Context Efficiency	Progressive Disclosure	Signal Quality	Description Quality	Overall
scientific-python-development	B+	A	B+	A	B+
project-management	A	A	A	C (no examples)	B+
scientific-domain-applications	D	F	D	A	D+
ai-research-workflows	B	A	B	C (no examples)	B

The scientific-domain-applications plugin is the clear outlier needing immediate attention. The project-management and ai-research-workflows plugins have excellent content but need description examples for proper agent selection. The scientific-python-development plugin is the strongest overall but has moderate redundancy that could be trimmed.