Use this skill to package one inline prompt, readable text file, or complete plan package for a standing goal system without changing content, context, intent, or meaning. It prepares input. It does not execute the goal.

Return exactly six lines as one command. The first line points to the packaged goal. The next five begin with , , , , and . Keep the full command under 1900 characters. Do not use a code fence or collapse lines.

Write the package under the active workspace's goal-prompt directory. Save the resolved input to a source snapshot, compute SHA-256 with a tool, and verify saved bytes against the resolved input. Use a collision-safe timestamp and slug.

For ordinary input, assign IDs to every atomic rule. Record strength, source location, derived wording, proof target, and or . Require 100% coverage and no failed result. Preserve ambiguity under .

For plan input, read plan-package input, reuse existing IDs and completion contracts, and write a sorted plan manifest. The plan remains authoritative.

Include these sections:

Keep source meaning, outcome, constraints, boundaries, approvals, and verification fixed. When an expected path fails, choose the first safe source-defined fallback, workspace rule, project convention, first-party contract, observed runtime fact, or smallest reversible in-scope change. Break ties by fewer changed artifacts, no new dependency, then lexical path order. Record trigger, evidence, choice, and result.

Resolve retrievable facts with tools. Stop only when every safe path violates authority, security, source meaning, or approval, or when required access is unavailable.

The steering and task mutation policy must preserve message order, stable task IDs, narrow supersession, and nonconflicting instructions. Live steering may add, update, remove, reorder, or replace execution tasks without changing persisted completion criteria. Persisted subgoals may change only through their supported controls. Replacing original completion criteria requires a replacement goal package and explicit reset.

After every mutation, reapply adaptation order, tie breaks, approvals, the review gate, verification, and stop rules.

Apply the gate to every review-driven skill or memory mutation. Block them until five consecutive qualifying turns have completed. Each qualifying turn requires at least 20 top-level tool calls from the current foreground profile. Interrupted turns, intervening user turns, pause, resume, completion, clear, replacement, session change, profile change, and turns below the call threshold reset the streak.

Steering controls do not qualify, reset, or count. Background, judge, delegated, cron, and tool-internal calls do not count. Eligibility is rolling and never bypasses ownership, provenance, approval, or verification. One-turn-local learning remains forbidden.