greenfield
Find the best technical solution by designing as if the repo's established rules, patterns, and conventions didn't exist, then audit only the rules the winning design collides with — updating outdated
SKILL.mdskills/greenfield/SKILL.md
GitHubEstablished rules harden into "hard truths" that design the project into corners long after their reasons died. This skill inverts the usual order: design first as if the rules didn't exist, then put only the colliding rules on trial. The best design is the primary deliverable; the rule audit is a light byproduct — never let it dominate the run's time or tokens.
Two failure modes to hold off simultaneously: anchoring (a session immersed in repo patterns produces the status quo with a new coat of paint — the greenfield frame exists to fight this) and laundering (the designer wants its design to win, so every inconvenient rule becomes "outdated" — the origin test exists to fight this).
Quick tier (default) — inline, no subagent
Run this yourself, in-session; it should work as a short detour mid-task.
- Ground. If the session doesn't already hold the problem's context, get what it needs first — read the relevant docs/code rather than designing from guesses. Proportionate to the question; grounding is not the deliverable.
- Greenfield design. Reframe the problem as if building fresh — same product goal, data shapes, scale, and security invariants; no inherited patterns, conventions, or "how we do it here". Sketch up to 2 genuinely distinct approaches — or 1 plus a sentence on why nothing else is close (forced second options are filler). Pick the winner on the merits.
- Collision report. List every established rule, pattern, or convention the winner breaks. You know them; report honestly — a hidden collision defeats the whole exercise.
- Origin test per collision (~30s each): "this rule prevents Y; Y is no longer real because Z" — the verdict must be about the rule's job, never the design's convenience. Verdicts: outdated / still load-bearing (bend the design, note the cost) / uncertain. Uncertain never stalls the run: log it as an open question and proceed on best judgment.
- Edits. Rules judged outdated get edited where they live, in the current working tree, disclosed in the final message (and PR body if one is in flight) — but only ordinary-convention-class rules. Escalation classes (below) always stop for the operator's approval — ask the user.
Deep tier (--deep) — for architecture-level stakes
When status-quo anchoring would poison the answer, buy structural blindness (~10–15 min):
- Sanitize. Distill a repo-agnostic problem brief — product goal, data shapes, scale, latency/cost envelope, security invariants, plus minimal code excerpts — into a pack in the session scratchpad. Banned in the brief: naming any current pattern, library choice, or architecture as a requirement; if it's load-bearing, state the underlying constraint instead.
- Blind design. Spawn a subagent on your strongest model whose context is ONLY the pack — it never opens the repo or its instruction files. It returns 2–3 distinct candidates with trade-offs and its pick.
- Independent adjudication. Spawn a fresh judgment agent with the winning design + the repo's rule surfaces: map collisions, and for each do real archaeology — the rule's recorded rationale, the git history that born it, whether the guarded failure still exists.
- Propose, never auto-edit. Deliver the verdict table; every rule change waits for the user's approval, then lands where the rule lives.
Rule surfaces and escalation (repo-adaptive)
Discover at runtime: the repo's agent instruction files (CLAUDE.md / AGENTS.md and what they cite), any rules directory, and any decision registry. Respect the repo's own severity scheme where one exists (e.g. Quirk's [security]/[infra] tags and its locked-decisions registry). Always-escalate, no tier exception: security- or infra-class rules, entries in a locked/decision registry (never edit these — surface them as challenges), and anything recorded as an operator decision. In repos without a severity scheme, treat rules touching auth, secrets, data exposure, or money as escalation-class by default.
Report (turn-final, both tiers)
- Design: the winner and why; runner-up in one line.
- Collisions: table — rule → verdict → origin-test sentence.
- Edits made (quick) or proposed (deep), each with its file.
- Open questions: uncertain collisions and anything escalated.