divecode

Strip AI tells from a markdown file (Korean or English) while preserving meaning. Single-pass workflow: scan with bin/divecode-humanize-scan, apply targeted rewrites per checklists/humanize-{ko,en}.md taxonomy, re-scan to verify, output a before/after report. Hard guard: change rate > 30% warns, > 50% halts. Use when asked to "humanize this", "AI 티 없애줘", "drop the AI tone", "윤문해줘", or to post-process drafted text before publishing. Korean taxonomy is a focused subset of im-not-ai (https://github.com/epoko77-ai/im-not-ai). English taxonomy is divecode's own collection. Use im-not-ai directly for deeper Korean analysis with academic-cited patterns.

PRD interrogation entry point — the product wedge of divecoding. Takes a rough PRD (path or pasted text), fires pattern-pack triggers, surfaces per-pack failure modes as a risk-map, and drives an interrogation loop that asks the 8-12 questions the PRD didn't answer. Produces design.md (Spec §2 populated from PRD), risk-map.md, and open-questions.md before any code is generated. Use when starting a new bolt with a written PRD. For cold-start (no PRD), fall through to /divecode-spec instead.

The always-on Genie pause for ad-hoc commands. Lighter than /divecode-prd — no PRD file required. Takes any user request ("build me a thing", "fix this bug", "add caching here") and runs three Genie questions before any code is generated: 1. What you literally asked for is X — confirm. 2. Things you didn't specify but X depends on: A, B, C. 3. Specify those now, or grant X literally and accept the consequences? Pattern packs supply A, B, C — same trigger matcher as /divecode-prd, just fed the user's ad-hoc request instead of a PRD file. Use when the user drops a one-line or paragraph-long command and you're tempted to just execute it. Voice triggers: "wish", "genie", "before you build", "are you sure".

Inception sub-phase for in-progress projects. Reads existing source + .md docs + sibling repos + relevant OSS ecosystem to surface SILENT DECISIONS (those made by action or omission without recorded rationale). Produces divecode/audit-<feature>.md. Auto-invoked from /divecode entry when in-progress detected (feature/* branch or existing source on main). Use when starting divecode on an existing project, or asked to "audit this feature", "what decisions are silent", "review the in-flight work".

Final phase of the bolt. Delete the working branch + worktree (if any), sync the integration branch, archive the bolt's marker so subsequent runs do not resume it, and prompt the user to write a lore entry if the bolt produced architectural decisions worth preserving. Use after /divecode-merge.

Group bolt changes into logical commits per profile.commit_convention. Honors style (conventional | freeform) and co_author (include | omit). Use after /divecode-fix-loop (or /divecode-review if no must-fix), before /divecode-push-pr.

Address review findings from final-review.md. Max 3 rounds; escalates to user if must-fix items remain after round 3. Tests stay green after each round. Use after /divecode-review when must-fix findings exist; before /divecode-commit.

Implementation phase of divecode. Reads requirements, design, and ARCHITECTURE, then writes code in small slices — each slice paused for human review. Iron Law: no slice exceeds what a human can sanely review in one sitting (~150 lines or one logical unit). Surfaces drift from the spec ("the architecture said X but doing it strictly creates Y problem — should we revise the architecture or accept the trade-off?"). Use after /divecode-arch.

Merge the bolt PR per profile.pr.merge_strategy (squash | rebase | merge). Use after /divecode-pr-watch returns status: green (or after a final fix-loop has cleared must-fix). Before /divecode-cleanup.

Watch a PR — auto-respond to comments and CI failures. Uses bin/divecode-pr-watch which polls via gh and emits a routing decision (status + route). 6-status / 7-route routing per agent-flow: green→merge, has_comments→fix-loop, ci_failed→fix-loop, pending→block, closed→block, merged→cleanup, error→block. Use after /divecode-push-pr, before /divecode-merge.

Push the bolt branch and open a PR against profile.pr.target_branch. PR body references design.md sections, slice list, verification results, open risks. Degrades gracefully when gh CLI is missing (prints push instructions instead). Use after /divecode-commit, before /divecode-pr-watch.

Final review phase — spawn multi-reviewer in parallel: generalist + architecture-design specialist (mandatory) + any angles declared in profile.review_angles. Aggregate findings with severity (must-fix / should-fix / notes) and source angle. Cites paths as path/to/file:line. Use after /divecode-implement, before /divecode-fix-loop.

Entry point for divecode — the deliberate, detail-first counterpart to vibe coding. Built on AWS AI-DLC macro flow (Inception → Construction → Operations) with agent-flow phase-level guardrails. Auto-detects a project profile (light/standard/strict) on first invocation, then routes to the appropriate sub-skill. Use when asked to "divecode", "start divecode", "do this properly", or when the user wants to design+build a feature with full human-in-the-loop interrogation rather than letting the agent run free. Voice triggers: "dive code", "dive in", "do it properly", "no vibes".

Decompose design.md (or requirements.md + ARCHITECTURE.md for light) into TDD-ready slices. Each slice ≤ 50% of the agent's context window and produces a reviewable artifact. Profile-conditional fields: light requires goal + files + tests only; standard adds layer + aggregates + verification; strict additionally requires test case names drawn from acceptance criteria. Produces divecode/slice-plan.md. PAUSES at end for human review of design + slice-plan together. Use after /divecode-spec (or /divecode-arch for light) and before /divecode-worktree.

Requirements interrogation — the spec phase of divecode. Asks the user about domain, users, access patterns, data shape, performance targets, failure modes, consistency requirements, security, and compliance. Surfaces niche knowledge (Redis cache stampede, SQL isolation, N+1, eventual consistency, etc.) from divecode/checklists/. Iron Law: never assume — always ask. Produces divecode/requirements.md. Use when starting a divecode session or when the user wants to nail down requirements before building.

Create the working branch (and optionally a git worktree) for the current bolt. Honors profile.branching: naming.prefix, slug_style, slug_source, max_slug_length, worktree (required | optional | never). Default fallback: feature/<kebab-summary>. Refuses to reuse manual namespaces (fix/, claude/) — those are out of divecode scope. Use after /divecode-slice-plan, before /divecode-implement.

Architecture interrogation phase of divecode. Reads requirements + design and pins down DTOs, layer boundaries, where transactions start/end, who owns cache invalidation, module/package structure, error type hierarchy, and dependency direction. The point is to be annoyingly detailed about things vibe coding skips — because that's where production bugs come from. Produces divecode/ARCHITECTURE.md. Use after /divecode-design, before /divecode-implement.

UI/UX design phase of divecode. Generates lightweight HTML mockups for each major screen/state, then runs a question loop ("did you think about empty state? loading? error? offline? 10,000 items? accessibility?"). Surfaces UX niche knowledge from checklists/ux-hig.md (Apple HIG, Material, a11y). If open-design is installed, delegates the design-system work to it. Produces divecode/design/*.html and updates requirements.md with UX decisions. Use after /divecode-spec, before /divecode-arch.

Usage limit awareness for divecode sessions. Checks Claude/Codex/Gemini quota via llm-usage-cli (preferred) or ccusage (fallback), then advises whether to proceed with a planned operation, split it into chunks, or wait. The point: divecode sessions can be long (multi-phase interrogation, mockup generation, guided implementation) — getting cut off mid-phase is worse than splitting upfront. Use before starting a long divecode stage or when the user asks about quota.