| name | write-skill |
| description | Use when authoring, revising, or validating a Claude Code skill through its create, update, and verify actions: define a new reusable capability, align existing skills with repository policy, or check structural, trigger, and behavior compliance before deployment. Invoke as write-skill create, update, or verify. |
| model | opus |
| context | fork |
| argument-hint | <create|update|verify> [target] [--changes=...] [--mode=...] [--runtime] |
Write Skill
Author, revise, and validate Claude Code skills. One skill, three actions:
create a new skill, update one or more existing skills, or verify a
skill's structural, policy, and trigger compliance. The first argument selects
the action; route to its reference below and follow that workflow.
Actions
create — Add a new reusable skill that teaches a missing behavior with
clear ownership and triggers. See
references/create.md.
update — Revise existing skill behavior, wording, or triggers, or align
skills with current policy, without creating a competing skill. See
references/update.md.
verify — Validate structural, repository-policy, and (when behavior or
discovery changed) trigger and behavior compliance, optionally exercising
isolated runtime prompts. See references/verify.md.
If the action is missing or ambiguous, ask which action is intended rather than
guessing. create requires that no suitable owner exists; when one does,
switch to update. create and update call the verify action for
functional and trigger evaluation.
Shared policy
Follow
${CLAUDE_SKILL_DIR}/../../constitution/references/authoring-invariants.md for
all three actions. Claude Code's validator remains authoritative for manifest
and frontmatter schema; repository policy checks never duplicate the evolving
Claude schema.
Shared thought experiment and blindspot test
Whenever an action changes a trigger or behavior, conduct a paper-only thought
experiment and blindspot test over positive and near-miss prompts before and
after the change. If written notes help, keep them as a temporary Markdown
scratch document in an OS temp folder (for example ${TMPDIR:-/tmp}/check.md)
using ${CLAUDE_SKILL_DIR}/../../constitution/references/check.md as the table
format with :white_check_mark:/:x: status markers. Delete the scratch
document before staging; these notes are reasoning aids, not deliverables, and
must not be committed. Do not claim runtime trigger behavior was exercised
unless an executable evaluation actually ran.
Verification
claude plugin validate --strict <plugin-path>
python3 "${CLAUDE_SKILL_DIR}/scripts/quick_validate.py" <skill-or-plugin-path>
Run both commands after every fix iteration. When a check fails, change only
the reported cause and re-run that check; loop fix and re-verify at most 3
times, then report partial completion with the remaining issues.
Completion
Report the action taken, affected skill paths, ownership boundaries or changes,
thought-experiment and blindspot coverage, validation results, runtime
evaluation status (exercised, not requested, or blocked), and any
intentionally deferred cases or unresolved ambiguity.
Confirm any temporary Markdown thought-experiment notes were deleted before
commit. Never claim a bulk update without listing its targets.