Skill optimization

Skill Content Optimization

Core Philosophy

1. Finding-Based: Every change resolves a recorded issue or follows a named project-specific source
2. Concrete: Each pattern provides detection criteria and transform methods
3. Structure-Focused: Optimizes expression and organization; domain knowledge remains unchanged
4. Intent-Preserving: Records the original requirements before changing structure, wording, constraints, context, or examples
5. Traceable: Connects every applied change to a finding or named project source
6. Self-Contained: Keeps every pure skill executable when loaded alone; duplication across independently loaded pure skills is valid when each copy is required for standalone execution

Content Optimization Patterns

P1: Critical (Must Fix)

Issues that directly reduce LLM execution accuracy when consuming the skill.

BP-001: Negative Instructions → Positive Form

Detection	Transform
"don't", "do not", "never", "avoid" in skill instructions	State the desired action or allowed state first. Preserve an explicit prohibition only when the violation is an irreversible operational action, the caller cannot normally recover it, and a positive-only rewrite would blur the boundary. Pair the prohibition with the safe alternative and the condition that authorizes crossing the boundary. Rewrite reviewable quality policies in positive form.

Exception boundary examples:

• Permitted: "Move obsolete records to the recoverable archive. Do not permanently delete them unless the user explicitly authorizes permanent deletion."
• Rewrite in positive form: "Do not invent issues" → "Base every issue on BP patterns or 9 principles", "Do not skip P1 issues" → "Evaluate all P1 issues in every review mode", "Do not give grade A when P1 exists" → "Assign grade A only when P1 count is zero"

Quality policies, role boundaries, scoring criteria, and general work rules always use positive form. Outputs that the caller validates, overwrites, or discards are never irreversible.

Skill example:

• Before: "Don't use generic variable names"
• After: "Use descriptive variable names that reflect purpose (e.g., userId not x)"

Why critical for skills: A prohibition alone leaves the executable target state unspecified.

BP-002: Vague Instructions → Specific Criteria

Detection	Transform
Vague term ("appropriate", "good", "proper", "best", "should be clear") that leaves a decision the intended outcome requires, where plausible interpretations would materially change execution or verification	Resolve it with the least-restrictive sufficient criterion, following the resolution steps below
Unspecified format, length, scope, tone, or success criteria whose plausible interpretations satisfy the intended outcome equally well	Treat as acceptable flexibility; add a constraint only when one interpretation is required (for a format a downstream consumer requires, see BP-003)

Resolution steps (first-row findings):

1. Choose the least-restrictive sufficient criterion — the measurable if-then rule or threshold that supplies the required precision while excluding the fewest valid behaviors.
2. Record its precision contribution: the observable output difference it improves for the intended outcome.
3. Record its constraint cost: the valid solutions allowed by the original intent that it excludes.
4. Apply it only when the precision contribution is identifiable and the constraint cost preserves the original intent.
5. When input or project context cannot determine the decision, record the required source instead of guessing.

Skill exception: Expressions that the LLM can resolve unambiguously from input context (e.g., "where the user left gaps" when the user's prompt is available for comparison) are not vague — they describe a deterministic operation, not a subjective judgment.

Skill example:

• Before: "Handle errors appropriately"
• After (criteria derived from a named source): "Follow the project error-handling policy (docs/error-handling.md): wrap external API calls, file I/O, and JSON.parse in try-catch; log error.name, error.stack, and timestamp; re-throw with context when the caller must handle it."
• After (no source available): "Record 'error-handling policy' as the required source instead of inventing try-catch targets, log fields, or thresholds."

Why critical for skills: A vague instruction forces the model to choose an outcome-relevant behavior without a supplied criterion.

BP-003: Missing Output Format → Structured Output

Detection	Transform
Skill describes what to do but not the expected deliverable format	Add an output section defining the structure, fields, and ordering required by the output consumer (parsing, routing, comparison, verification), rather than selecting a format by convention

For a skill review, the output contract contains BP-001 through BP-008 coverage, unique finding IDs, severity, location, quoted evidence, one resolution per finding, preservation requirements, unresolved inputs, and the final grade. For skill creation, the output is the complete SKILL.md content plus any required same-directory references or scripts.

Skill example:

• Before: "Analyze the code for issues"
• After (format required by the review-report consumer): "Emit ## Issues Found as a table the report renderer parses: | Severity | Location | Description | Suggested Fix |"

Why critical for skills: Structured output constraints reduce hallucination and make skill results consistent.

P2: High Impact (Should Fix)

Issues that reduce skill effectiveness when addressed.

BP-004: Unstructured Content → Organized Format

Detection	Transform
Wall of text without headings	Apply standard section order (see below)
Multiple topics mixed in one section	Split into distinct headed sections
No tables for reference data	Convert lists of criteria/patterns to tables

Standard skill section order:

1. Context/Prerequisites
2. Core concepts (definitions, patterns)
3. Process/Methodology (step-by-step)
4. Output format/Examples
5. Quality checklist
6. References

Conditional: Skip restructuring if skill is under 30 lines and covers a single topic.

BP-005: Missing or Excess Context → Necessary and Sufficient Context

Detection	Transform
Skill assumes knowledge not stated	Add Prerequisites section listing required context
Domain terms used without definition	Add definitions inline or in a glossary table. Skill exception: Terms within the LLM's baseline knowledge (widely-used technical terminology, standard domain vocabulary) require no definition. Only project-specific terms, internal naming conventions, or domain jargon outside common LLM training data need explicit definition.
No "when to use" guidance	Add trigger conditions with concrete scenarios
Duplicated, distracting, or unactionable context with no downstream effect	Condense repeated facts into one operative statement; keep raw background behind a path or reference when only an extracted fact is needed; name the source for project-specific facts

Skill example:

• Before: "Apply the strangler pattern for migration"
• After: "Prerequisite: Existing monolith with identifiable module boundaries. When to use: Replacing legacy module while maintaining production traffic."

BP-006: Complex Content → Decomposed Steps

Detection	Transform
3+ objectives in one instruction	Break into numbered steps; each step names its output evidence and the transition condition that permits the next step
Sequential dependencies not explicit	Make each step's transition condition depend on the prior step's output evidence
Multiple dependent actions presented as one step	Split so each produces observable completion evidence before the next begins

Conditional: Skip decomposition for simple reference tables or single-criteria rules.

Key insight: Goal is externally visible state progression — each step produces evidence that controls whether the next step is valid, not decomposition for its own sake.

For creation and comprehensive review, use three gates in order:

1. Analysis gate: Original requirements are recorded, BP-001 through BP-008 are covered, every issue has evidence, and no unresolved input blocks faithful work.
2. Optimization gate: Every finding has one applied/skipped resolution, each change is traceable, and all preservation requirements remain represented.
3. Balance gate: Intent preservation, decision sufficiency, information density, constraint necessity, and traceability pass before the result is final.

P3: Enhancement (Could Fix)

Incremental improvements for specific contexts.

BP-007: Unnecessary or Biased Examples → Minimal Necessary Examples

Detection	Transform
Examples restate behavior already known to the LLM	Replace with a concise rule or consumer-required output shape, and remove the examples
Examples encode a domain-, product-, or organization-specific mapping, non-obvious exception, or boundary a rule cannot express	Keep the smallest set that covers those mappings; map each example to the ambiguity it removes
Multiple examples remove the same ambiguity, or all share the same surface pattern	Reduce to the smallest covering set; add a different case only when it removes a distinct ambiguity

BP-008: No Uncertainty Permission → Explicit Escalation

Detection	Transform
Skill demands definitive answers always	Classify claims as observed, inferred, or unknown; add escalation criteria for ambiguous cases
No "when to stop" guidance	When an unknown blocks the next step, stop at that gate and name the exact evidence or user decision required to continue

Skill example:

• Before: "Determine the root cause"
• After: "Classify the root cause as observed, inferred, or unknown. When missing evidence blocks the next step, stop at the current gate and name the exact evidence or user decision required to continue."

9 Skill Editing Principles

Measurable quality criteria for skill content. Each principle includes a pass/fail test.

#	Principle	Pass Criteria	Fail Example
1	Context efficiency	Every sentence contributes to LLM decision-making. No filler.	"This is an important skill that helps with..."
2	Deduplication	No concept is explained twice at the same abstraction level within one skill. Duplication across independently loaded pure skills is valid when each copy is required for standalone execution; evaluate those copies for semantic consistency rather than replacing them with sibling-skill references	The same rule appears twice in one skill without adding a distinct execution role
3	Grouping	Related criteria in single section (minimize read operations)	Scattered error handling rules across 4 sections
4	Measurability	Criteria name observable evidence, deterministic decision rules, or justified thresholds	"Write clean code" without an observable condition
5	Positive form	Instructions state what to do (BP-001 applied)	"Don't use any" instead of "Use only X"
6	Consistent notation	Uniform heading levels, list styles, table formats	Mix of -, *, 1. in same context
7	Explicit prerequisites	Project-specific and non-baseline prerequisites are stated or linked; baseline technical knowledge is left concise	Uses "DI" without defining Dependency Injection
8	Priority ordering	Most important items first, exceptions last	Edge cases before common patterns
9	Scope boundaries	Explicitly state what the skill covers and the conditions that activate conditional content. A pure skill contains the context required for standalone execution. Cross-skill references are reserved for skills whose role is orchestration or skill selection	A pure skill omits an operative rule because another independently loaded skill also contains it

References

• Creating skills: See references/creation-guide.md for generation flow and description guidelines
• Reviewing skills: See references/review-criteria.md for evaluation flow and grading