Run a settled implementation goal as a bounded autonomous build loop — lock the goal from the conversation and any docs written this session, front-load every decision, map the terrain, write a completion-invariant contract before any code, then loop (building with explicit Agent-tool fan-out, committing at intervals, verifying at milestones with fresh-eyes against the contract) until the invariants hold, and close out. Gated on a `--confirm` flag asserting the user has manually set ultracode effort and auto-accept mode (a skill can set neither); without the flag it stops, gives the setup steps, and offers two paths — a managed `--confirm` run, or a copy-pasteable `/goal` command that hands the work to native goal + workflow orchestration. An optional `--commit` flag (default off) enables commit-at-intervals and push-at-milestones in both the managed run and the generated command; without it, nothing is committed. Use when a plan is settled and you want Claude to implement it end-to-end on its own — typically
Surface and confirm the load-bearing assumptions behind a task before a long or expensive run — goal, root, what may be edited, what is off-limits, what done means, and open questions — separating what can be cited from what is being guessed, so bad direction is caught before time is burned. Also pressure-tests the plan's acceptance criteria, flagging any that are weak or built on unproven assumptions, and asks the blocking uncertainties (scope boundaries, interpretations, soft "done" bars) as structured questions. Use at the start of a multi-step or high-cost task, when resuming ambiguous or handed-off work, or whenever the target is not crisply stated and you are about to commit to a long run.
BETA / under development. Mine the current project's Claude Code sessions for recurring corrections, then interrogate them — consolidate clusters into themes, ask a SET of open "why" questions per theme, self-answer the evidence-answerable ones and forward the developer-only ones to the user, record confirmed answers, and optionally export a CLAUDE.md of principles. Use when the user says "/debrief", "debrief my sessions", "what patterns do you see in how I work", or wants to refresh / interrogate their developer-context corpus. Drives the `debrief` CLI entirely via bash — NO manual paths, NO MCP registration required.
Fresh-context audit of a finished diff against the conversation's stated intent. A subagent with zero conversation history reviews the diff blind — reporting what it believes the change does, its quality, and any oversights — then the main context reconciles that blind read against what the work was actually meant to do. Use when a chunk of work is complete and you want fresh eyes to confirm nothing was missed, scope is fully covered, and no oversights or bugs slipped through, before moving on or shipping.
Audits a repository's Claude context-injection setup — CLAUDE.md, CONTEXT.md, docs/, .claude/agents/, and the per-project memory directory. Reports bloat, broken links, orphaned docs, security risks, missing rules in subagent prompts, and conflicts between memory and project instructions. Use when the user asks to audit their Claude setup, asks "what's wrong with my CLAUDE.md", wants to know if their agents/memory/docs are configured well, or wants to improve Claude's effectiveness in this repository.
Audit Claude's per-user memory for this project and produce a report. Asks at the start whether the user wants a non-technical summary or a full technical audit. Never edits memory — produces a report only. Usage: /memory-audit
Halt a drifting agent and re-anchor it to codebase evidence for the current task, clearing speculative context without a full compaction. Use when the agent has gone off course, hallucinated files/APIs/behavior, or started over-building beyond what was asked — or when the user types /reground or says "stop", "re-ground", "re-crown", "I don't see evidence for that", or "you're building something we don't need".