lean-agent-skills-complete

Design module boundaries, data ownership, interfaces, domain models, migrations, refactors, and architecture decision records from real constraints. Use for structural choices with lasting change cost.

Build or run authorized browser automation, real-user QA, data entry, or extraction with stable locators, explicit state, bounded retries, and evidence of the user-visible result.

Design or review a command-line interface that humans and agents can run reliably. Use for headless automation, flags, help, output contracts, exit codes, pipelines, retries, dry-run, and safe state changes.

Diagnose a hard defect or performance regression through a tight reproducible feedback loop, evidence-first investigation, falsifiable competing hypotheses, targeted instrumentation, and verified regression coverage.

Design a controlled product, performance, or engineering experiment—or a disposable prototype—with a falsifiable hypothesis, comparable baseline, guardrails, and stopping rule.

Run a bounded adversarial quality loop with independent critics, frozen benchmarks, repair, and retest over a real artifact. Use only when hidden-defect risk makes one direct check insufficient.

Take ownership of a complex, multi-session, or long-running task until verified completion. Use when the user wants durable execution, program-scale orchestration, and evidence rather than advice or a plan alone.

Interview the user one decision at a time to stress-test a plan or requirement. Use only when an important choice is unresolved and the answer is not already available.

Create a concise status recap or durable handoff. Use after substantial work, before interruption, or when another session needs the exact current state and next action.

Implement a bounded change from a clear request, spec, or ticket. Use when the work fits a focused session; use get-it-done for long-horizon ownership and gauntlet-loop for costly adversarial acceptance.

Resolve merge, rebase, or cherry-pick conflicts by preserving intended behavior from both sides and verifying the integrated result. Use when conflict markers or semantic integration failures exist.

Create, edit, inspect, convert, or repair DOCX, PDF, PPTX, and spreadsheet files while preserving structure and validating the final artifact with the available file tools.

Turn evidence and resolved decisions into an executable proposal, spec, ticket set, workflow, refactor plan, or comparison of competing plans. Use for planning artifacts, not implementation unless also requested.

Create or update durable project context; map or explain verified repository structure and flows; record approved lessons, AI asset cards, or retrospectives. Use only when the user explicitly requests one of these artifacts.

Deliver a user-authorised working slice fast, with one cheap reality check and visible deferrals. Use only when the user explicitly requests Quick Mode, a rough prototype, or an equivalent fast-path scope; not for quick questions or short replies.

Prepare and verify a software or artifact release, including version, changelog, build, package, checksums, migration notes, staged rollout, operations evidence, and rollback. Publish only with explicit authorization.

Investigate a question using current primary sources and produce a source-backed briefing. Use for documentation, factual checks, literature or product research, and source-faithful transcript or video summaries.

Independently review code, interfaces, writing, data, architecture, repository agent-operability, or another agent’s work against the real contract and artifact. Default to read-only; repair only when explicitly requested.

Create, refactor, evaluate, package, import, or route portable agent skills, workflow instructions, and skill stacks. Use for SKILL.md, plugin compatibility, benchmark design, third-party audits, trigger quality, executable workflow review, or reducing skill-set bloat.

Teach a concept or skill through a plain explanation, worked example, active recall, and bounded practice. Use when the user asks to learn, understand, practise, study, or be quizzed—not for a quick reference answer alone.

Design or improve automated tests and test-first feedback loops. Use for TDD, regression coverage, characterization, integration, end-to-end, property, fuzz, concurrency, compatibility, performance, or agent-trajectory testing.

Triage a bug, request, alert, work item, or live incident by verifying evidence, classifying impact, stabilizing risk, and naming the next owner or action. Use when the report or operational state is uncertain.

Re-pitch a confusing, dense, or context-poor response in friendly outcome-first ASD-STE100-inspired prose. Use when the user explicitly asks for a clearer restatement.

Draft or edit instructions, UI text, errors, help, emails, documentation, reports, and other prose for purpose, evidence, clarity, audience, and task fit while preserving the requested voice.