mohamed-alaboudi/zulfiqar/ultracode

Ultracode

Treat the invocation text as the objective. Keep all load-bearing judgment with the root agent.

A skill cannot raise the native reasoning effort of an already-running session. Start the session at the desired maximum effort before invoking this skill.

1. Guard the repository. Confirm the working directory, Git root, branch, dirty state, duplicate clones, and applicable instructions before broad edits.
2. Frame the work. State the objective, hard constraints, falsifiable success criteria, dependencies, unknowns, and assumptions that would change the design.
3. Choose the architecture. Steelman the strongest alternative, then make one explicit root-owned decision.
4. Build a dependency graph. Keep dependent chains, shared interfaces, architecture, integration, and product decisions with the root.
5. Delegate only independent bounded execution slices when delegation is available and useful. Give every worker one owner, exact read/write scope, exclusions, deliverable, acceptance check, and stop conditions. Do not create recursive worker trees.
6. Integrate centrally. Inspect every worker result, resolve conflicts, perform cross-slice edits, and preserve unrelated user changes.
7. Simplify the integrated change without weakening security, validation, accessibility, data-loss prevention, or public contracts.
8. Prove the result. Run relevant tests, builds, type checks, and side-effect read-backs. Use one fresh adversarial verifier for load-bearing claims when available, fix confirmed failures, and rerun checks.
9. Bookend the result. Report the outcome, exact evidence, changed paths, residual risk, and next authorized action. Follow repository Git and shipping policy.

Do not let worker consensus replace root judgment. Do not claim maximum native effort merely because this skill was invoked.