skillWriter

A skill encodes what cannot be derived from first principles. Before writing, understand what you're encoding.

Track Progress

Use your todo tool to track these checkpoints:

• • Discovery sufficient to attempt zone assessment
• • Zone assessment completed (100 points distributed)
• • writing-documents skill invoked before writing each document
• • Self-check passed
• • Testing completed

Skill Structure

skill-name/
├── SKILL.md           # Gestalt and cross-cutting concepts
├── meta/
│   ├── DISCOVERY.md       # Research notes, domain understanding
│   └── CREATION_NOTES.md  # Feedback on authoring experience
└── references/
    └── ...            # Detailed guidance by topic

Capsule: Discovery

Invariant You cannot write an effective skill without understanding the subject in excruciating detail first.

Example Before writing a mermaid skill: read all mermaid documentation, create dozens of diagrams, discover what breaks, learn which features are experimental, understand why certain patterns fail. //BOUNDARY: Zone assessment comes after discovery, not before. You can't assess what you don't understand.

Depth

• •Sources: the user (ask questions), the internet (research), existing code (what actually works), your intuition (you have good instincts)
• •Document findings in meta/DISCOVERY.md—this is your working memory
• •Identify what leads to great outcomes vs what causes failure
• •Only then can you assess which zones matter and what to encode
• •Discovery documents are messy and complete; skills are clean and selective

The Four Zones

Skills contain different kinds of content. Most skills blend these, but the blend matters.

Capsule: Knowledge

Invariant Inject context that cannot be derived from first principles.

Example API schemas defining valid request shapes. Company naming conventions. Business rules for edge cases. //BOUNDARY: If Claude could figure it out by reasoning, it belongs elsewhere.

Depth

• •Distinction: Knowledge is facts Claude lacks; explanation is facts Claude has but needs prompted.
• •NotThis: Generic best practices Claude already knows.
• •SeeAlso: references/knowledge.md

Capsule: Process

Invariant Coordinate sequences where order matters and steps cannot be safely skipped.

Example Deployment: build → test → stage → approve → deploy. Skipping "test" or reordering causes harm. //BOUNDARY: If steps can be reordered freely, use a checklist instead.

Depth

• •Distinction: Process enforces order; checklists verify completion.
• •Trade-off: More structure reduces errors but limits flexibility.
• •SeeAlso: references/process.md

Capsule: Constraint

Invariant Enforce rules where deviation is the failure mode.

Example Security policy: Never log credentials. Never commit secrets. Always validate external input. //BOUNDARY: If deviation is acceptable in some contexts, it's guidance.

Depth

• •Distinction: Constraints are non-negotiable; guidance permits judgment.
• •Trade-off: Stricter constraints reduce risk but may block valid approaches.
• •SeeAlso: references/constraint.md

Capsule: Wisdom

Invariant Introduce ways of thinking that generalize across contexts.

Example "Write for intelligent agents who need structure, not scripts." This reshapes approach to documentation, skills, communication—everything. //BOUNDARY: If it only applies to one domain, it's knowledge.

Depth

• •Distinction: Wisdom shapes thinking; knowledge provides facts.
• •NotThis: Templates, procedures, or domain-specific rules.
• •SeeAlso: references/wisdom.md

Zone Assessment

After discovery, before writing, distribute 100 points across the zones.

This is not optional. The distribution clarifies what you're building and guides every subsequent decision.

Calibration Examples

Interpreting Your Distribution

Primary Zone

Where did most points go? That zone's reference file is your primary guide.

Agency Level

Your distribution suggests how much to trust the agent versus prescribe behavior.

Note: High constraint doesn't automatically mean low agency. "These outcomes must be achieved" (high agency) differs from "Follow these exact steps" (low agency). Both enforce constraints.

Content Style

Before Writing

Invoke the writing-documents skill now. Skills are documents. Their effectiveness depends on writing quality. Know how to write well before writing skill files.

With your distribution clear, ask:

On purpose:

• •What would Claude get wrong without this skill?
• •What cannot be derived from first principles?

On scope:

• •What does this skill NOT cover?
• •Where should Claude go for adjacent needs?
• •Is this skill focused at the right level? Are we likely to have many variations of it?

On form:

• •Does the content style match your primary zone?
• •Are you prescribing where you should trust?
• •Are you trusting where you should constrain?

Reference Files

Consult based on your distribution:

• •references/knowledge.md — Structuring facts for retrieval, when high agency fits
• •references/process.md — When sequences matter, balancing flexibility with structure
• •references/constraint.md — Rules without justification, when prescription is appropriate
• •references/wisdom.md — Introducing ideas that change thinking, not templates

Skills with balanced distributions: consult multiple references, weight by allocation.

Cross-Cutting

• •references/scripts.md — When to use scripts, best practices, platform considerations. Scripts serve Knowledge (fetch current truth), Process (automate steps), and Constraint (verify compliance).

Self-Check

Before shipping, verify:

• • Zone assessment completed and recorded in skill.md yaml front matter
• • Content style matches primary zone
• • Agency level appropriate for the blend
• • Boundaries explicit where non-obvious
• • Description triggers for intended queries
• • Every section earns its token cost
• • Tested with a subagent

Testing

Launch a subagent to use the skill for real work:

1. •Subagent works — uses the skill for its intended purpose
2. •Subagent self-checks — verifies output follows the skill's rules
3. •Human reviews — qualitatively assesses the result
4. •Subagent reports — documents experience via FeedbackLoop

The subagent catches rule violations (linear sequences in a mermaid diagram, missing error handling in a process). The human catches qualitative issues (is this actually useful? does it feel right?). The feedback captures what the skill got wrong.

Capsule: PolymorphicSkill

Invariant A polymorphic skill presents shared concepts upfront, then branches to specific variations via progressive disclosure.

Example This skill: SKILL.md defines four zones and assessment. Each zone has a reference file with specific guidance. You learn the shared model first, then dive into the variant you need. //BOUNDARY: If variations share nothing, they're separate skills, not a polymorphic one.

Depth

• •Shared concepts reduce duplication and establish common vocabulary
• •Progressive disclosure keeps the gestalt readable
• •Variations inherit context from the shared layer
• •The agent reads only the branch they need, not all branches
• •Zone inheritance: references inherit the parent's zones as a starting point; adjust per reference based on what needs to be presented
• •SeeAlso: ProductiveAbsence (forces reading the right branch)

Capsule: ProductiveAbsence

Invariant Omit enough that the reader cannot succeed without engaging the referenced material.

Example The zone capsules above tell you what each zone IS, not how to write one well. You cannot write a good wisdom skill from this page alone—you must read references/wisdom.md. //BOUNDARY: Absence is productive only when the path forward is clear. Missing information with no pointer is just incomplete.

Depth

• •Distinction: Productive absence drives engagement; incompleteness causes failure.
• •The gap is a forcing function. Enough to orient, not enough to fake competence.
• •This document applies the principle to itself.

Capsule: FeedbackLoop

Invariant Each skill includes meta/CREATION_NOTES.md documenting the authoring experience, rewritten by each maintainer.

Example

1. Create or update the skill using skillWriter
2. Read existing meta/CREATION_NOTES.md (if exists)
3. Rewrite with your experience synthesized (replacement, not append)

//BOUNDARY: Notes are about creating/maintaining the skill, not using it. Read only when rewriting.

Depth

• •Work first, read notes only when rewriting—fresh perspective during creation
• •Rewrite entirely—synthesize prior experience with yours
• •Focus on: what helped, what confused, what's missing, what would improve skillWriter
• •Aggregating CREATION_NOTES.md across skills periodically improves skillWriter

You can't teach what you don't understand.