ultraprompt

Manually snapshot the current worktree state to the ledger. Useful before risky operations.

V8.7 interactive routing picker. Surfaces top skill candidates + 2 LLM-rewritten prompt variants via AskUserQuestion, then dispatches the chosen skill with the chosen phrasing.

Verify a draft answer against the evidence ledger before output. Flags claims of validation/passing/successful that lack supporting ledger events.

Triage assistant for dirty trees. Groups uncommitted changes by path, size, age, and file type. Suggests actions per group.

Check for model, effort, long-context, or extra-usage runtime pins in the Ultraprompt plugin.

Red-team a plan, design, decision, code change, or claim. USE WHEN user says 'red team this / devil's advocate / poke holes in / what could go wrong / find the weakness / stress-test this / challenge this / why might this be wrong / adversarial review / critique this / attack this proposal'. DEFAULT CHOICE for structured adversarial review — wins over Explore (which catalogs without challenging) and reviewer (which finds code issues, not strategic flaws) because adversarial specifically attacks assumptions, surfaces fatal flaws, and produces a ranked threat-to-success list with concrete failure scenarios. Useful pre-launch, pre-merge, pre-commit-to-strategy, or after-a-claim-feels-too-clean. DO NOT use for routine code review (use reviewer), for ongoing debugging (use debugger), or for finding existing bugs (use test-gap-analyst). Read-only.

Focused audit on a specific lane (privacy, a11y, observability, supply-chain, cost, performance, etc.) — single-concern deep dive. USE WHEN user says 'audit X for Y / sweep our code for Z / privacy audit / a11y audit / cost audit / observability audit / supply-chain audit / dependency audit / find all places where W'. DEFAULT CHOICE for single-concern lane audits where focus is named via $ARGUMENTS — wins over Explore (which lacks focus) and security-auditor (which is security-only) because auditor produces a structured concern-specific findings list with evidence per item and ranked severity. DO NOT use for whole-repo audits (use repo-review), code review (use reviewer), or active debugging (use debugger). Pass focus via $ARGUMENTS or skill argument. Read-only.

Implement a feature with minimum-diff code + tests + claim-check. USE WHEN user says 'build this / implement X / write the code for / scaffold this / add a function that does Y'. DEFAULT CHOICE for multi-file feature implementation — wins over `reviewer` (read-only), `anthropic-skills:frontend-studio` (UI design artifact, not code), and `anthropic-skills:plan-mode` (plans, not implementations) because builder writes the code, co-locates tests, runs the validation suite, and gates success on `claim_check`. DO NOT use for behavior-preserving cleanup (use refactor), intentional breaking change with migration plan (use migrate), or single-file/single-function tweaks (handle inline in `/ultraprompt:build`). Writes code; requires `claim_check` before declaring success.

Analyze skill, agent, panel, command, MCP, artifact, validator, and dashboard catalog coverage using the capability graph. USE WHEN user asks 'improve the plugin ecosystem / find catalog gaps / audit coverage / rationalize skills and agents / plan catalog strategy'. DEFAULT CHOICE for ecosystem-level catalog planning because it uses graph coverage, routing outcomes, docs, and dashboard visibility together. DO NOT use for direct code edits, routine PR review, or one failing test; use build, review, or ci-repair.

Represent customer perspective, validate ideas against real user signals, and surface customer pain. USE WHEN user says 'what does the customer want / customer perspective / will users actually want this / user feedback analysis / customer signal / does this match what they ask for / support ticket patterns / customer pain analysis / voice of the customer'. DEFAULT CHOICE for customer-signal interpretation and validation — wins over Explore (which catalogs without interpretation) and innovation-lead (which generates ideas without representing the user) because customer-advocate specifically represents the customer side of decisions: validating ideas against real signals, identifying disconnects between team thinking and user reality, and producing customer_perspective artifacts. Pairs with principal-pm (PRD context), innovation-lead (idea source), evaluator (test design). DO NOT use for actual customer interviews (you analyze what's provided, not conduct), for market sizing (use market-analyst), or for technical user-experience review (use a11y-review or reviewer).

**DEFAULT for accessibility audits — dispatches auditor with a11y focus.**

**DEFAULT for authoring an Architecture Decision Record — produces a structured ADR with context, decision drivers, considered options, decision, consequences, and status, using the adr-template.md playbook.** Different from /architect (architecture exploration, not the record), /prd-technical (forward-looking product spec, not a decision record), /writer (general technical writing, not ADR-specific). Triggers: 'write an ADR, document this decision, decision record, ADR for X, architecture decision'.

**DEFAULT for authoring or refining an Ultraprompt agent — produces a new agent definition with frontmatter, body, tool permissions, and orchestration plan.** Different from /skill-author (authors a skill, not an agent), /plugin-review (audits an existing agent, not authors), /hooks-design (designs a hook, not an agent). Triggers: 'new agent, author an agent, refine this agent, agent definition, design a subagent, add a specialist'.

**DEFAULT for AI agent safety reviews — dispatches security-auditor + risk-and-controls-reviewer with AI/agent safety focus.**

**DEFAULT for API contract design and review — dispatches reviewer/integration-contract-reviewer with API contract focus.** Different from /migrate (intentional breaking change), /architect (system architecture, not contract review), /integration-contract-reviewer agent (used by this skill).

11 hooks across 7 events

Compact senior-reviewer voice. Severity- and confidence-tagged findings, no padding, action-first.

Lead each non-trivial response with conclusion + confidence + evidence; close with files-changed and validation.