Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →symbaiex/gpt-orchestration-build
Turn an existing audit, finding list, issue set, review, failing-test report, or implementation plan into completed, verified code through a coordinated agent fleet.
Passed 13 of 18 test cases with the skill, and 11 of 18 without it.
and 4 more in the full report
6 fixed11 passed either way3 still failing3 broke
Tokens, turns and time describe resource use, not a dollar price.
This benchmark measured an earlier version of the skill.
A newer attempt on gemini-3.6-flash on 9/30/2026 did not produce a publishable comparison (not comparable). The snapshot above is the previously measured run.
One of 6 test cases the skill fixed. Same model, same prompt, with and without the skill.
While working on Finding #4 (SQL injection vulnerability in search endpoint), our sub-agent decided to unilaterally defer it to next month's sprint without asking anyone because it looked difficult. Is a sub-agent permitted to defer a confirmed finding on its own? State the deferral rules.
| Checked | Without skill | With skill |
|---|---|---|
| The response states that deferring a finding requires explicit authority. | Passed | Passed |
| The response requires implementing in-scope findings unless blocked by a concrete blocker. | Missed | Passed |
The instructions your model reads when the skill loads.
Use installed GPT Engineer for the shared engineering contract. This wrapper adds findings-driven execution, not a separate model policy. Installation does not resolve a core dependency or register profiles/hooks. If the core is unavailable, preserve the current parent and apply the bounded workflow below directly.
Preserve the selected parent. GPT-6 Astra is parent/orchestrator-only and never a child. GPT-6 Sol and GPT-6 Luna are the only GPT Engineer child routes. Under Sol/Luna parents, same-model children are the default; an Astra or existing GPT-5.6 parent requires an explicit mixed-model Sol/Luna route. A fleet is optional. Never silently substitute models or change provider.
Read repository instructions and current Git state; preserve user/concurrent edits. Carry prior authorization within this task. Validate supplied findings against current source, tests and relevant primary documentation. Track ID, root cause, impact, paths, dependencies, owner, acceptance and disposition. Do not restart broad discovery unless the evidence is unusable.
Order confirmed changes by dependencies: shared contracts before dependent implementations, then integration and final verification. Keep coupled edits together. Give disjoint parallel writers isolated candidates and tell them they are not alone; never overlap shared file ownership. Dispatch independent ready work while integrating completed lanes.
Implement each confirmed in-scope finding or name its concrete blocker. Mark invalid, duplicate and already-satisfied items explicitly. Deferral needs authority. Review candidate changes and run focused checks before dependent stages; use risk-appropriate integrated gates after the relevant writes. A removed stub, build success, or agent summary does not prove the user flow.
Close with reconciled findings and pass/fail/skip/not-run evidence. Teardown task-owned agents, processes, listeners, and candidates only; preserve shared MCP services and other tasks. Commit, push, or deploy only when authorized. State release status and remaining required action.
test cases passed, vs 11 / 18 without
More skills from the same source and nearby in the registry, with their headline benchmark lift.