claude-swe-workflows

Autonomous technical lead. Takes commander's intent and drives a project to completion through an OODA loop over implementation, refactoring, review, and bug-hunting skills. Has broad authority — creates tickets, commits, invokes any skill — and only escalates via andon cord for irreversible actions, public releases, or genuinely blocking decisions. User acts as product owner; skill acts as tech lead.

Cuts a project release. Discovers the project's release procedure (Makefile target, RELEASING.md, CI workflow, etc.) and offers to capture it durably if missing. Always invokes /review-release as preflight, proposes a version bump from CHANGELOG, then presents an exact command plan for operator confirmation before executing step-by-step. Halts on first failure rather than attempting rollback.

Pre-release readiness review. Scans for debug artifacts, version mismatches, changelog gaps, git hygiene issues, breaking changes, and license compliance. Runs tests and build verification. Presents consolidated findings for human review before release.

Proactive bug-hunting workflow. Assesses codebase risk through complexity, coverage, and structural analysis, then spawns focused investigators that write reproducing tests to validate suspected bugs. Thoroughness over speed. Advisory only — produces findings and proposes tickets; does not implement fixes.

Autonomous bug-elimination loop. Iteratively invokes /bug-hunt and /implement-batch until findings converge below an operator-specified severity floor. At termination, runs /review-test scoped to the run's new reproducing tests and fixes quality issues above the floor. Auto-approves /bug-hunt and /review-test ticket proposals; pulls the andon cord on contested findings, repeated implementation failure, hard-cap exhaustion, or breaking-change proposals. Optional /refactor finisher.

Autonomous comprehensive refactoring. Three-phase pipeline — Phase 1 /refactor (tactical, loops internally to convergence) → Phase 2 loop of /review-arch + /implement-batch until architectural findings converge below a severity floor (default HIGH+) → Phase 3 /refactor (catches tactical issues introduced by restructuring). Auto-approves /review-arch ticket proposals; pulls the andon cord on contested findings, breaking-change proposals, repeated implementation failure, or Phase-2 hard-cap exhaustion (5 architectural iterations).

Autonomous comprehensive review. Once-through pipeline over /review-health, /review-arch, /review-security, /review-perf, /review-a11y, /review-test, /review-release. Operator-configurable ticket creation at startup — when ON, auto-approves sub-skill ticket proposals per the orchestrator-family contract; when OFF, uniformly declines and surfaces findings in the consolidated report. Auto-detects sub-skills that do not apply (no web content → skip /review-a11y; no tests → skip /review-test).

Architectural review workflow. Analyzes codebase organization via noun analysis and produces a target blueprint. Advisory only — does not implement changes. Always offers to cut tickets for the planned work; orchestrators receive the proposal and apply their own autonomy judgment to approve / edit / decline.

White-box security audit. Blue-teamer and lead red-teamer run in parallel isolation for an independent first pass — neither sees the other's output during reconnaissance. A synthesis step categorizes findings into four prescriptive buckets (anchoring-suppressed, convergent, blue-flagged-unverified, divergent), producing a unified target list. Focused red-teamers then deep-dive each target. Iterates when exploit chains are discovered. Heavy and thorough by design. Advisory only — produces findings and proposes tickets; does not implement remediation.

Comprehensive test suite review across five phases. Surveys unit coverage, integration coverage, E2E (browser) coverage for webapps, fuzz coverage, and test quality — in that order. Advisory only — produces findings and proposes tickets; does not implement test changes.

Mechanical documentation hygiene. Spawns a doc-maintainer agent to audit all project documentation and fix correctness, completeness, freshness, and consistency issues autonomously within its authority. Surfaces anything requiring user judgment for approval.

Multi-ticket batch workflow. Takes a batch of tickets, plans execution order, implements each via /implement in autonomous mode, runs cross-cutting quality passes, and presents results for final review.

Full-lifecycle project workflow. Takes batched tickets, implements via /implement-batch, runs smoke tests, then executes a comprehensive quality pipeline (refactor, review-arch, review-test, tidy-docs, review-release). Maximizes autonomy with andon cord escape.

Autonomous iterative refactoring workflow. Scans for tactical code quality improvements (DRY, dead code, naming, complexity), implements through SMEs, verifies with QA, commits atomically, and loops until no improvements remain.

First-pass strategic-orientation review of a repository. Produces an evidence-cited map (not a grade) calibrated to a reference class, helping the user decide where to engage, where to tread carefully, and where to leave alone. Advisory only — no changes made.

Mechanical git repo hygiene. Prunes stale remote-tracking refs, deletes merged local branches (with confirmation), and reports stashes, untracked files, and unpushed work. Never touches remote state. Safe by default — destructive operations require explicit operator approval.

Adversarial project planning workflow. Explores the problem space, drafts tickets in batches, then runs sequential adversarial loops — a mandatory UX loop ("should we build this?"), zero or more discretionary specialist loops (security, performance) for projects with architectural implications in those domains, and a mandatory implementer loop ("could we build this?") — to find gaps before implementation. Creates well-specified, batch-tagged tickets upstream only after every applicable loop signs off.

Exploratory workflow for understanding problem spaces and drafting feature proposals, refactorings, or complex bug-fixes. Creates detailed tickets in your issue tracker without doing implementation work.

Mutation testing workflow. Systematically mutates source code to verify tests actually catch bugs. Multi-session with progress tracking.

Bug-fixing workflow that coordinates diagnosis, test-driven reproduction, root-cause analysis, and targeted fixes. Use when the user wants to fix a bug with thorough investigation and regression testing.

Iterative development workflow that coordinates implementation, refactoring, QA, and documentation agents to complete features systematically. Use when the user wants a full development workflow with quality checks.

Pre-compaction housekeeping. Walks a checklist (persistent memory updates, git hygiene, trash cleanup) plus an open-judgment audit, produces an SBAR with a go/no-go recommendation, and emits a copy-pasteable resume prompt for the post-compaction agent if work remains. Run this immediately before /compact.

Abductive reasoning about causes. Takes a phenomenon, separates observations from interpretations, then spawns diagnosticians applying different reasoning lenses (technical, human-factors, process, incentive-structure, environmental, temporal, measurement-artifact, statistical) in isolation. Orchestrator evaluates candidate causes for fit-against-evidence, calibrates confidence honestly, and reports leading candidate(s) with distinguishing evidence needed. Produces feedback only — no code, no tickets, no artifacts.

Analysis of Competing Hypotheses. Operationalizes Richards Heuer's CIA-tradition technique for systematically narrowing among multiple hypotheses against evidence. Builds an explicit hypothesis-vs-evidence matrix and focuses on disconfirmation rather than confirmation — the central insight: hypotheses cannot be proven, only disconfirmed; the surviving hypothesis is the one with the least disconfirming evidence. Spawns parallel hypothesizers (in isolation, across hypothesis-generation angles) and parallel evidence-gatherers (across evidence classes), then synthesizes into a matrix, diagnosticity analysis, sensitivity analysis, and falsification milestones. Produces feedback only — no code, no tickets, no artifacts.

Prospective failure imagination for a plan, decision, or running system. Operates in two modes — **plan mode** (a plan has not yet been committed to; imagine its catastrophic failure broadly) and **scenario mode** (a specific catastrophic scenario is posed against an existing system; investigate causes against the actual code and architecture). Spawns parallel pre-mortemers in isolation across three categories of lens: standard failure classes (technical, operational, estimation, scope, adoption, dependency, team, incentive, detection, reversibility, adversarial), 0-3 ad-hoc target-specific lenses the orchestrator names when warranted, and a first-principles lens that always runs to catch what the prescribed taxonomy misses. Synthesizes into a prioritized risk register with early-warning signals. Produces feedback only — no code, no tickets, no artifacts.

Accessibility audit. Dispatches an accessibility auditor to identify WCAG conformance gaps, prioritize by user impact, and recommend fixes. Advisory only — no changes made.

Performance review. Detects project type and dispatches compute and/or web performance reviewers. Advisory only — no changes made.

Retrospective learning from a completed experience. Takes a project, incident, decision that played out, or time period; gathers ground truth (observations) separately from recollections; spawns reflectors applying different lenses (what-worked-vs-got-lucky, what-didn't, what-surprised, system-rewards-vs-intent, decisions-that-aged, what-to-tell-past-self, patterns-that-recur) in isolation; synthesizes into updated mental models as first-class output. Produces feedback only — no code, no tickets, no artifacts.

Problem redefinition before problem solving. Takes a stated problem, extracts its premises, then spawns reframers applying different lenses (problem-vs-symptom, scope-shift, stakeholder-shift, level-of-abstraction, time-horizon, inversion, category-shift, constraints-shift) in isolation, and synthesizes the alternatives into a report with an orchestrator recommendation. Produces feedback only — no code, no tickets, no artifacts.

Good-faith divergent brainstorming for achieving a goal. Validates assumptions, then spawns brainstormers running different techniques (first-principles, working-backwards, lateral, analogical, constraints-shift, etc.) in isolation, and synthesizes the pool into a report of standouts, reasonable ideas, and hybrid combinations. Produces feedback only — no code, no tickets, no artifacts.

Adversarial deliberation process for making decisions. Spawns advocate agents for each option who argue their cases before a judge (you) who renders a verdict.

Good-faith adversarial scrutiny of an idea or plan. Spawns critical skeptics across relevant lenses, pairs them with an advocate defending the idea, and synthesizes the exchange into a report. Produces feedback only — no code, no tickets, no artifacts.