Menu
Analyse phase — interrogate the request for missing requirements, unstated assumptions, and unhandled cases before any code is written.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
Start an Agent Smith ticket the right way — read the ticket, its dependencies, the PRD Decision Log, and the affected architecture/docs before editing. Use when picking up an AS-NNN ticket, deciding what to work on next, or before touching code for a backlog item. Triggers: "work on AS-", "start a ticket", "implement the ticket", "what should I work on next", "pick a ticket".
Agent Smith pull-request lifecycle — open a PR, subscribe to activity, reply to and resolve every review thread, handle the Gemini/Copilot review order, and auto-merge when clean. Use when work on a branch is committed and pushed, when a PR exists for the branch, or when handling review comments or CI on a PR. Triggers: "open a PR", "PR review", "review comment", "CI failed on the PR", "merge the PR", "is the PR ready".
Reflect phase — produce the three success artifacts (a measurable success metric, an instrumentation diff, a check-back ticket draft) without ever reading shipped-app runtime data.
Analyse phase — trace what else the change touches: callers, shared state, persisted formats, and cross-package contracts.
Plan phase — pressure-test the proposed plan: smallest viable change, ordering, and the test that will prove it.
Reflect phase — capture what was learned as durable, grounded notes: surprises, follow-on tickets, and method tweaks.
Based on SOC occupation classification
| name | grill-gaps |
| description | Analyse phase — interrogate the request for missing requirements, unstated assumptions, and unhandled cases before any code is written. |
You are in the analyse phase of Coding Mode. Your job is to grill the feature for gaps before a line is written — surface what the request leaves unsaid so the plan does not bake in a wrong assumption.
Work through, and report only what you actually find:
Every gap you raise must cite the concrete thing — a file, function, missing test, type, or ticket. A gap with no anchor is noise; drop it. Never emit generic advice like "consider edge cases" or "follow best practices".
Good (grounded):
internal/session/store.go Load() never handles a truncated log file; there
is no test for a partial write — add one before relying on resume.Config.Window is always set, but routing.Default()
leaves it zero — AS-042 only fills it when a policy is configured.Bad (rejected — no anchor):