Skills

Chooses reasoning effort levels per workload for models that expose adaptive thinking or reasoning-effort controls. Segments traffic by task difficulty, measures quality, latency, and cost at each effort level, and produces a routing table with defaults, escalation rules, and monitoring so teams stop paying for maximum effort on easy requests.

Writes reusable agent skills: packaged instructions that a coding agent loads on demand for a specific kind of task. Defines a precise trigger description, scopes what the skill covers, structures step-by-step instructions with supporting scripts and references, and adds test prompts to confirm the skill triggers and performs as intended.

Writes API reference documentation from an OpenAPI spec, route definitions, or source code. Documents each endpoint with purpose, authentication, parameters, request and response schemas, error codes, and runnable examples in multiple languages, using consistent terminology and noting pagination, rate limits, and versioning.

Reviews an HTTP API against the common API risk classes: broken object and function level authorization, weak authentication, excessive data exposure, missing rate and resource limits, mass assignment, injection, and misconfiguration. Maps each finding to the affected endpoint, rates severity, and proposes concrete fixes and tests.

Plans large offline LLM workloads using batch APIs or self-managed queues. Decides which jobs can tolerate delayed results, sizes batches and concurrency against rate limits, designs idempotent job records, retries, and partial-failure handling, and compares batch discounts against synchronous cost.

Interprets published LLM benchmark results critically. Explains what each benchmark measures, checks comparability across reported settings such as effort level, tools, and scaffolding, flags vendor-reported versus independent numbers and possible contamination, and translates scores into what they do and do not imply for a specific use case.

Plans and executes a migration of an Anthropic integration from Claude Opus 5 or earlier to Claude Opus 5.5. Inventories model IDs and request parameters, replaces manual extended-thinking budgets with adaptive thinking and effort levels, reviews context- window and output-limit assumptions, recalculates cost with the new pricing and cache- read rates, and defines an eval and rollout plan.

Designs input and output guardrails for an LLM product. Defines the policy to enforce, places checks at the right points such as input filtering, retrieval and tool-result screening, and output validation, chooses between rules, classifiers, and model-based checks, and specifies fallback behavior, logging, and evaluation of false positives and negatives.

Reviews and rewrites the tool names, descriptions, and input schemas of an MCP server so models pick the right tool and fill arguments correctly. Removes overlap between tools, clarifies when to use each one, tightens parameter types and enums, trims token-heavy descriptions, and proposes test prompts to verify tool selection.

Builds an evaluation plan for swapping one LLM for another in production. Assembles a representative test set from real traffic, defines graded criteria and automated checks, compares the two models on quality, latency, cost, refusals, and tool-call correctness, and sets pass thresholds that gate the rollout.

Designs a prompt caching strategy for LLM applications. Restructures prompts so stable content such as system instructions, tool definitions, and reference documents forms a cacheable prefix, places cache breakpoints, estimates break-even from cache-write and cache-read prices and hit rates, and defines monitoring for cache effectiveness.

Writes an engineering RFC or design document for a proposed change. Frames the problem and goals, states non-goals, describes the proposed design with diagrams in text, compares alternatives with trade-offs, and covers rollout, migration, testing, risks, and open questions so reviewers can make a decision.

Produces a technical due diligence brief on a vendor, product, or codebase. Covers architecture and scalability, security and compliance posture, engineering practices, dependency and licensing risk, team and bus-factor signals, and cost structure, then summarizes red flags, open questions, and a recommendation with stated confidence.

Writes and revises interface microcopy: button labels, form hints, empty states, confirmations, onboarding steps, tooltips, and notifications. Keeps wording short, action-oriented, and consistent with the product voice, and provides alternatives sized to the available space with notes for localization.

Writes scripts for short and long-form video. Opens with a hook, structures the narrative into timed segments, pairs narration with on-screen visuals and b-roll suggestions, adapts tone and pacing to the platform and audience, and ends with a clear call to action.

Designs agent workflows that let a model continue useful reasoning while external tools run asynchronously. Covers tool contracts, pending-call state, idempotency, timeout handling, wait-tool patterns, result reconciliation, and user-visible progress for long-running business or engineering tasks.

Migrates existing OpenAI model integrations to GPT-6 Astra with attention to Responses API usage, reasoning effort, tool calling, unsupported parameter cleanup, long-context behavior, cost controls, and rollout validation. Useful when moving complex reasoning, coding, research, or computer-use workloads from GPT-5.6 or earlier models to Astra.

Structures long-running agent work so new user instructions can be incorporated mid-turn without losing the original goal. Defines steering precedence, conflict handling, safe interruption points, unchanged-work preservation, and final-response reconciliation for complex coding, research, and operations tasks.

Produces documents, spreadsheets, presentations, and structured deliverables that follow an existing template, style, and business context instead of generating generic output. Covers source selection, required context checks, layout constraints, citation boundaries, validation passes, and final artifact review.

Reviews Amazon Bedrock AgentCore runtime, memory, identity, gateway, policy, browser, and code-interpreter operations before deployment. Separates read-only, billable, mutating, and destructive actions and produces scoped enablement, observability, rollback, and cost-control recommendations.

Plans governed data-processing workflows across AWS Glue, EMR, EMR Serverless, Athena, S3, and IAM. Produces catalog, compute, orchestration, data-quality, encryption, cost, and observability decisions with read-only discovery first and explicit approval before resource changes.