Claude consulting for teams who picked the wrong Claude.

Claude consulting is the engineering practice of getting Claude-backed features into production with the right model tier, the right prompt rig, and the eval harness to know whether either is working. The honest version of the category builds and runs the systems it sells. The dressed-up version sells prompt-tweaking by the hour. We are the first kind. The proof is Sentinel.

• Opus 4.7 for high-stakes reasoning, multi-step agent loops, and the calls where being right matters more than being fast.
• Sonnet 4.6 balances cost and quality - it wins most evals. Most production calls live here as the daily driver.
• Haiku 4.5 for high-volume classification, routing, and the calls where p95 latency (the slowest 5% of requests) under 500ms is the constraint.

The rule is: write the eval first, run all three at the cheapest setting, and let the eval pick. Most teams arriving at us are using Opus where Sonnet would win and Sonnet where Haiku would win.

XML tags, not Markdown headers. Claude was trained to follow XML structure (, , , ) more reliably than free-form prose, and the lift from the same prompt restructured into XML is usually two to five eval points without any other change. We also use thinking blocks for hard reasoning calls, system prompts for invariants, and put the most important instruction at the end of the user message - Claude weights recency more than the middle of long contexts. The system prompt is for the role; the user message is for the task; the assistant prefix is the underrated lever.

Prompt caching breaks even at roughly two cache hits within the five-minute TTL, since the write is 1.25x and the hit is 0.1x of base cost. That makes caching a near-universal win for any prompt with a stable system message, a long retrieved context, or a multi-turn conversation. We wire caching at the call-site by default and set the cache_control (the API flag that activates prompt caching) breakpoint at the largest stable prefix - usually the system prompt plus retrieved documents. Teams who haven't enabled it are typically overspending by 40-70% on Sonnet and Opus calls.

Treat the 200K window as a budget, not a license. Claude's recall degrades on long contexts the same way every model's does - middle-of-context information gets lost more than start or end. We retrieve narrowly, place the most important information at the top and the task at the bottom, and use prompt caching to make long stable contexts cheap. Published benchmarks for long-context recall look better than real-world performance. The production reality is to design for the first 32K tokens and let caching pay for the rest.

Yes. Tool use is the right primitive for any feature that needs to fetch, calculate, or write. We define tools with strict JSON schemas, validate every tool result, and run a retry-with-correction loop when a tool returns an error. Thinking blocks (extended thinking on Opus and Sonnet) are reserved for genuinely hard reasoning calls where the answer benefits from chain-of-thought - most production prompts do not need them and pay for them anyway when teams enable thinking globally. Eval first, then decide.

Anthropic consulting is the enterprise procurement lens - vendor selection, contract terms, security review, deployment options across Anthropic API, AWS Bedrock, and Google Vertex. Claude consulting is the product-engineering lens for teams who already picked Claude and need the model tier, prompt rig, and eval harness right. OpenAI consulting is the same engineering work but on the GPT side. We do all three. This page is for teams shopping the product side.

Strategy sprints run $25–45k. Production proofs-of-concept run $50–150k. Full implementations start at $150k and scale with integration depth. Embedded team augmentation is a monthly retainer. Pricing is the same regardless of industry - we charge by engagement shape, not by model provider.

Yes, mutual NDA before any technical conversation. We do not work for clients with conflicting active engagements in the same competitive set during a quarter - a rule we enforce on ourselves more strictly than most clients ask us to.