Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →symbaiex/gpt-engineer
Own a software-engineering outcome end to end with a strictly model-routed agent fleet: research the real codebase, turn findings into implementation waves, edit safely, test and inspect the result, and persist through bounded gap-closing cycles until the authorized goal is complete.
Passed 14 of 18 test cases with the skill, and 14 of 18 without it.
and 3 more in the full report
4 fixed14 passed either way2 still failing4 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 10/2/2026 did not produce a publishable comparison (not comparable). The snapshot above is the previously measured run.
One of 4 test cases the skill fixed. Same model, same prompt, with and without the skill.
We need to dynamically calculate fleet capacity and schedule subagent concurrency based on available system load before launching 10 parallel code transformation tasks. Which provided utility script in this environment handles capacity planning for fleet workers?
| Checked | Without skill | With skill |
|---|---|---|
| Identifies `scripts/plan_fleet.py` as the script for fleet capacity planning | Missed | Passed |
The instructions your model reads when the skill loads.
Own the requested outcome. Preserve the selected parent model. GPT-6 Astra is orchestrator-only: use it only when it is the active parent/chat model, never as a child, worker, verifier, fallback, or spawned profile. GPT-6 Sol and GPT-6 Luna are the only GPT Engineer child routes. Use Sol for ambiguous, demanding, multi-step lanes and Luna for focused, repeatable lanes; same-model children are the default under Sol/Luna parents, while an Astra parent must explicitly choose a Sol/Luna mixed-model child. Older GPT-5.6 parent sessions may continue during rollout, but do not create new GPT-5.6 child routes or silently change the selected parent. Neither Astra nor a fleet is required. On Claude, use the portable engineering guidance; native Claude agents do not become GPT agents.
Read applicable repository instructions, Git state, and the relevant execution path. Preserve user changes and concurrent work. Establish observable acceptance, scope, and authority. Carry forward prior authorization and steering in this task. Resolve routine reversible choices; ask only when missing information or authority materially changes the outcome. Reviews remain read-only unless implementation is authorized. Persistence does not authorize unlimited retries, spending, deployment, or expanded scope. Continue through implementation, required verification, and fixing in-scope failures; do not stop at the first patch or ask to continue already-authorized work.
Work in the intended repository; a shell working directory does not change the task's writable roots. If repeated approvals expose a workspace mismatch, explain it rather than widening permissions. For approval friction, read runtime integrations.
Small fixes need a focused change and decisive check, not a fleet, journal, or mandatory preflight. For substantial work, track requirement/finding IDs, dependencies, ownership, and acceptance evidence. Before research fleets or repeated experiments, state concurrency, lane scope, progress checkpoints, and applicable resource limits. Productive work may take hours: elapsed time, compaction, or a checkpoint alone is not a reason to stop. Estimates are reassessment points, not hard deadlines. Honor explicit user/operator limits and runtime constraints; reassess repeated failure, low useful yield, or resource pressure. For extended work or prompt tuning, read long-running work.
Use current session/turn metadata, supported controls, or documented hook fields—not the model's self-description. Saved defaults do not prove the active session model. Keep requested and effective model, effort, and service tier separate, with source and subject ID. Subagent hooks can describe the parent session; never promote that metadata to child attestation. Missing evidence stays
test cases passed, vs 14 / 18 without
Published skills whose descriptions closely match this one, with their headline benchmark lift.
More skills from the same source and nearby in the registry, with their headline benchmark lift.