holycodex 0.16.9-dev.103.1 → 0.16.9-dev.108.1

package/dist/assets/plugin/skills/writing-instructions/SKILL.md

@@ -1,46 +1,37 @@
 ---
 name: writing-instructions
-description: Use when authoring or reviewing model-facing instructions.
+description: Use when model-facing instructions need authorship or review; make them coherent, scoped, and complete for their receiver.
 ---
 
 # Writing instructions
 
-Use this skill for model-facing instructions, including Root/session policy,
-specialist and Role.task contracts, skills, and conditional workflows. Keep
-shared behavior under one owner across model routes; account for model-specific
-differences only when supported by evidence.
-
-Before writing, identify the receiver's effective context: higher-priority and
-user instructions, repository instructions, Role/task policy, relevant skills,
-tools and capabilities, configuration, and current task context. Add only the
-missing semantic delta. Resolve conflicts against the governing authority;
-do not duplicate a meaning already supplied to the same receiver.
-
-Give each meaning one authoritative owner at the narrowest scope that applies:
-Root/session invariants in Root policy, repository conventions in AGENTS.md,
-shared specialist boundaries in the specialist baseline, route-specific behavior
-in Role.task, and genuinely conditional workflow or tool knowledge in a skill.
-Generated projections derive from or validate against that owner. Invocation
-metadata selects a branch; it does not become a second policy source.
-
-Define the required outcome and completion evidence, bounded authority, relevant
-constraints, material decisions or blockers that escalate, and stopping
-conditions. Keep required lifecycle and safety invariants explicit. Give the
-receiver room to choose routine safe, reversible, in-scope steps. A contract must
-carry authorized work through its requested terminal state, including required
-repair and proof; a first pass or pending gate is not completion.
-
-Remove older-model competence scaffolding, fixed recipes without a required
-ordering constraint, broad reading rituals, stale environment facts, duplicate
-policy, and unnecessary questionnaires. Preserve meaningful repository checks
-and review gates through their canonical owners instead of reproducing them in
-every skill. Use configuration for verbosity rather than generic style padding.
-
-When changing a skill's invocation boundary or conditional branches, follow
-[Skill mechanics](SKILL-MECHANICS.md).
-
-An instruction change is complete when the receiver can act within its authority,
-recognize completion and escalation boundaries, and return sufficient evidence;
-the effective context has no contradictory or duplicate rule for that meaning.
-Validate changed behavior with meaningful proof and repository-required gates,
-using the canonical testing policy rather than inventing extra test rituals.
+Identify the receiver's effective context: instruction hierarchy, repository
+rules, Root or Role.task policy, selected skills, tools, configuration, and
+Assignment facts. Add only the missing constraints; resolve contradictions
+against the governing authority. User instructions override skill guidelines.
+
+Give each meaning one owner: Root/session policy for orchestration and judgment,
+AGENTS.md for repository conventions, the specialist baseline for shared
+boundaries, Role.task for task authority, and skills for conditional workflows
+or tool knowledge. Generated projections derive from that owner. Metadata
+selects the workflow; it does not duplicate its policy.
+
+State the outcome, bounded authority, material constraints, completion evidence,
+and escalation conditions. Preserve known facts needed to act; point to sources
+for details the receiver can discover reliably. Routine safe in-scope choices
+should proceed without questions. Carry authorized work through repair and proof.
+
+Remove redundant policy, older-model scaffolding, broad reading rituals, rigid
+recipes without a real dependency, hypothetical gates, and repeated testing.
+Preserve lifecycle invariants and required checks. Use configuration for default
+verbosity; use prompts for task-specific evidence and output requirements.
+
+Keep skill descriptions short and situation-first. Audit triggers alongside
+nearby skills; avoid skills that merely restate a task contract. Keep coherent
+shared guidance together and link uncommon branch-specific mechanics only from
+the condition that needs them. Metadata default prompts invoke the named skill
+without embedding another instruction source. Preserve upstream attribution.
+
+Validate the effective projection, not exact prose: the receiver can act within
+its authority, recognize completion and blockers, and return sufficient evidence
+without contradictory rules. Reuse the canonical testing policy.

package/dist/assets/plugin/skills/writing-instructions/agents/openai.yaml

@@ -1,6 +1,6 @@
 interface:
   display_name: "Writing instructions"
   short_description: "Author and review model-facing instructions."
-  default_prompt: "Use writing-instructions for this instruction change."
+  default_prompt: "Use $writing-instructions for this task."
 policy:
   allow_implicit_invocation: true