Oma Skill Creation

Skill Creation - SSL-lite Skill Authoring

Scheduling

Goal

Create, revise, and validate OMA skills using the SSL-lite Markdown structure derived from Scheduling-Structural-Logical skill representation while keeping the result readable, executable, and maintainable.

Intent signature

• User asks to create a new OMA skill, agent skill, specialist skill, or .agents/skills/*/SKILL.md.
• User asks to convert or normalize a skill into the project skill format.
• User asks whether a skill has enough scheduling, structural, logical, reference, or canonical execution detail.
• User mentions SSL-lite, Scheduling / Structural / Logical, skill format, skill template, or skill creator.

When to use

• Creating a new skill under .agents/skills/{name}/SKILL.md
• Updating an existing OMA skill to the SSL-lite format
• Adding Canonical command path or Canonical workflow path to an execution-heavy skill
• Auditing whether a skill is usable by routing, execution, validation, and recovery logic
• Deciding whether detailed examples belong inline or in resources/

When NOT to use

• Installing third-party skills into $CODEX_HOME/skills -> use Codex's built-in skill installer (external; not an OMA skill)
• Creating a Codex plugin bundle -> use Codex's plugin tooling (external; not an OMA skill)
• Writing a general project plan unrelated to skill authoring -> use PM Agent
• Editing product code, infrastructure, frontend, backend, or mobile implementation directly -> use the matching specialist skill

Expected inputs

• skill_name: target directory name, preferably oma-{domain}
• skill_goal: capability the skill should add
• trigger_intents: user prompts or domains that should activate the skill
• boundaries: when the skill should not be used and which skill should handle those cases
• execution_model: whether the skill is command-heavy, workflow-heavy, judgment-heavy, or reference-heavy
• Existing files, tools, commands, resources, or standards the skill must preserve

Expected outputs

• A complete SKILL.md using the SSL-lite top-level sections
• Optional resources/*.md, config/*.yaml, scripts/*, or assets/* only when progressive disclosure or deterministic execution requires them
• A validation report with structure checks, routing checks, execution checks, and unresolved assumptions

Dependencies

• Existing OMA skill conventions in .agents/skills/*/SKILL.md
• SSL-lite template in resources/ssl-lite-template.md
• Validation checklist in resources/validation-checklist.md
• Local filesystem and search tools for reading adjacent skills and checking output

Control-flow features

• Branches by new skill vs existing skill update
• Branches by command-heavy vs workflow-heavy vs reference-heavy skill type
• Reads and writes local skill files
• Must preserve user edits and avoid unrelated rewrites

Structural Flow

Entry

1. Identify whether the task is creating a new skill, updating an existing skill, or auditing a skill.
2. Read nearby or analogous OMA skills before choosing wording and structure.
3. Determine the skill type: command-heavy, workflow-heavy, judgment-heavy, or reference-heavy.

Scenes

1. PREPARE: Collect skill purpose, trigger intents, boundaries, inputs, outputs, dependencies, and risk/resource constraints.
2. ACQUIRE: Read analogous skills, existing resources, project conventions, and any user-provided source material.
3. REASON: Decide what belongs inline in SKILL.md and what belongs in resources/, config/, scripts/, or assets/.
4. ACT: Create or update the skill using the SSL-lite template.
5. VERIFY: Run oma skills lint --skill {name}, then the remaining routing, execution, resource, utility-content, and formatting checks.
6. FINALIZE: Report created/changed files, validation result, and any remaining assumptions.

Transitions

• If the skill performs fragile or repeated commands, include ### Canonical command path inline.
• If the skill is mostly human judgment or investigation, include ### Canonical workflow path inline.
• If detailed examples are long, variant-specific, or optional, move them to resources/ and reference them explicitly.
• If a skill already exists, preserve working content and normalize only what is needed for the target format.
• If the trigger description is too broad, narrow it before writing the skill.

Failure and recovery

Exit

• Success: the skill has the four SSL-lite top-level sections, a canonical path, clear routing boundaries, resource constraints, and validation evidence.
• Partial success: the structure is valid but missing commands, tools, source material, or user decisions are reported.
• Failure: no skill file is written or the blocking ambiguity is explicit.

Logical Operations

Actions

Tools and instruments

• oma skills lint --skill <id> for automated smell detection (frontmatter, structure, canonical path, broken references, boundaries)
• rg, find, awk, sed, git diff --check
• The runtime's native file-edit tool for manual edits
• Existing OMA skills as examples
• resources/ssl-lite-template.md for the canonical section skeleton
• resources/validation-checklist.md for acceptance criteria

Canonical workflow path

1. Read 1-3 analogous existing skills before writing.
2. Classify the new skill as command-heavy, workflow-heavy, judgment-heavy, or reference-heavy.
3. Draft SKILL.md from resources/ssl-lite-template.md.
4. Add exactly one inline canonical path:
   • ### Canonical command path for fragile or repeatable commands
   • ### Canonical workflow path for decision, review, design, or research flow
5. Move long examples, provider-specific details, or optional protocols into resources/.
6. Validate with oma skills lint --skill <name> (frontmatter, headings, canonical path, broken references, boundaries), then git diff --check for whitespace. If the CLI is unavailable, fall back to the structural grep/awk checks in resources/validation-checklist.md.

Resource scope

Preconditions

• The target skill name and purpose are known or can be inferred safely.
• Adjacent skill overlap has been checked.
• The user has authorized changes under .agents/skills/ when project rules would otherwise protect .agents/.

Effects and side effects

• Creates or modifies files under .agents/skills/.
• May add resource files to support progressive disclosure.
• Does not stage or commit changes unless explicitly requested.

Guardrails

1. Keep the four top-level sections exactly: Scheduling, Structural Flow, Logical Operations, References.
2. Keep YAML frontmatter with clear name and description; routing depends on description quality. Run oma skills audit after editing description to confirm the new wording does not collide with adjacent skills (warn ≥ 60%, fail ≥ 75% TF-IDF cosine).
3. Include concrete When NOT to use boundaries and cross-routes to adjacent skills.
4. Include expected inputs, expected outputs, dependencies, and control-flow features. Prefer the structured outputs: block when artifacts can be globbed so oma verify can perform a closure check.
5. Include structural scenes using SSL scene vocabulary where practical: PREPARE, ACQUIRE, REASON, ACT, VERIFY, RECOVER, FINALIZE.
6. Include logical actions with SSL primitives such as READ, SELECT, VALIDATE, INFER, WRITE, CALL_TOOL, NOTIFY, and TERMINATE.
7. Include resource scope and resource targets for filesystem, codebase, process, credentials, network, user data, or memory.
8. Include effects and side effects, especially for commands, network calls, credentials, destructive actions, generated files, and long-running processes.
9. Add one canonical path inline so an agent can execute or reason without loading every resource file.
10. Put long variant-specific details in resources/, not in the main skill body. Keep the SKILL.md body under 500 lines; oma skills lint warns as body-too-long past that. 10a. Write examples only when the output format is a contract someone parses — a CLI's result block, a schema, a LAW-validated document. Worked input/output pairs that merely show "what a good report looks like" narrow the model's exploration space to the example's shape, so state the requirement instead and let the model choose the form. Per-task shape belongs in the task, not the skill. 10b. Do not add instructions to double-check, re-verify, or self-review before answering — current models already do this, and the instruction compounds into over-verification. Deterministic validators the skill can actually run (oma verify, a lint script, a schema check) are the exception and should stay: they are feedback loops, not re-reading.
11. Do not create extra README, changelog, or installation docs inside a skill.
12. Do not overwrite unrelated user edits.
13. Enforce the three utility-predictive content dimensions — failure mechanism encoding, actionable specificity, high-risk action blacklist — per the Utility Content Checks in resources/validation-checklist.md (SkillLens, arXiv:2605.23899).
14. Bulk-data scenes must run functions over data, not read data into context: route the processing through a deterministic tool/CLI stage (CALL_TOOL) and return only a summary plus the artifact path. Do not design scenes that stream raw harvested/parsed data through the model's context (see ../_shared/core/context-budget.md, Core Principle 5).

References

• SSL-lite template: resources/ssl-lite-template.md
• Validation checklist: resources/validation-checklist.md
• Shared context loading: ../_shared/core/context-loading.md
• Shared quality principles: ../_shared/core/quality-principles.md
• Skill utility eval: when creating a new skill, consider adding held-out task fixtures under .agents/eval/<skill>/ so oma skills eval can measure whether the skill improves task outcomes. See web/docs/guide/skill-eval.md for the fixture schema and checker types.
• Skill optimization: once eval fixtures exist, oma skills opt --skill <id> proposes SKILL.md edits and keeps only those that improve measured held-out utility lift — use it after authoring instead of hand-tuning prose.