Claude 5 Best Practice

Claude 5 Best Practices

Use this skill as a routing and runtime overlay for the Claude 5 generation. Safety, permissions, repository instructions, and explicit user constraints remain authoritative.

Default objective: satisfy the acceptance contract, then minimize accepted-result cost. Lower latency, higher quality beyond the contract, or mandatory topology is a different objective; pursue it at higher cost only when the user explicitly chooses that tradeoff.

Product snapshot

As checked on 2026-07-26:

Claude Opus 5 launched 2026-07-24 at half Fable 5's price with the same context and output limits, and it is the documented starting point for complex agentic coding and enterprise work. Fable 5 remains the most capable widely released model. These facts are volatile. Read references/evidence-notes.md before quoting them as current.

Terms

• acceptance contract: observable acceptance criteria, required safety and verification, and approval boundaries.
• tier: a vendor-defined capability class: Haiku, Sonnet, Opus, or Fable.
• effort: reasoning and checking depth within one tier.
• lane: one tier + effort single-agent configuration.
• orchestration: the single-agent or lead/workers/verifier topology.
• route: lead lane, worker lanes, orchestration, and verification plan together.
• accepted-result cost: total spend required to pass the contract, measured in the surface's primary accounting unit: API dollars or subscription credits, never both.
• non-inferior: acceptance performance within a predeclared tolerance on representative work; it does not mean identical output on every run.
• surface: the active product/client, account entitlements, model controls, worker overrides, permissions, and accounting system.

Track tokens, latency, calls, and human correction as separate diagnostics or guardrails. If the surface exposes no comparable spend signal, describe relative resource intensity instead of claiming an exact cheapest route.

Interpret explicit delegation requests

A request to “use subagents/workers with proper models to save cost” authorizes delegation and worker-lane selection within scope; it does not mandate fan-out, lower the acceptance contract, choose a premium lead, or waive the premium gate. Treat the cost purpose as conditional: compare the three routes below and delegate only when the routed route is expected to be the cheapest accepted route.

If the user requires subagents as a topology independent of cost, follow that topology only when it preserves the acceptance contract. “Use subagents regardless of cost” is already an explicit higher-cost objective. If the user instead requires subagents “to save cost” but routing evidence predicts higher cost, pause and ask whether topology or cost governs. If mandatory topology cannot preserve the acceptance contract, pause and ask the user to relax either topology or the contract; never weaken or silently replace one requirement.

Pin and verify each worker's effective model and effort where supported. If overrides are unavailable, inherited, ignored, or unverified, do not dispatch workers for cost savings and do not claim worker-tier savings. Prefer the qualified single-agent lane or a handoff. A separately confirmed mandatory topology may still use such workers, but it is not a cost-saving route. Before dispatch, state the topology, worker lanes, cost rationale, and uncertainty compactly. After completion, report verification, observable effective lanes, retries/rescues, and the spend signal; without comparable spend data, report only relative resource intensity.

Cost-aware routing procedure

Follow this order. Later sections explain vendor-specific mappings but do not override the procedure.

1. Fix the contract and objective. Define the required deliverable, checks, semantic review, safety constraints, and approvals. Record any explicit user choice to trade higher cost for latency, quality beyond the contract, or mandatory topology.
2. Inspect the surface. Identify the active model, available tiers and effort values, lead-switch or new-task handoff mechanism, worker-model overrides and inheritance, orchestration limits, permissions, and accounting unit. If a capability is unavailable or unverified, do not promise savings from it.
3. Choose direct-lane candidates. Judge ambiguity, coupling, state depth, semantic risk, objective checkability, and expected rework. For recurring work, reuse validated evidence. Without a baseline, include both the cheapest plausible lane and a stronger reasonable baseline; do not treat the premium tier as ground truth. Label one-off task-shape judgments as heuristics.
4. Prefer effort escalation within Opus 5 over tier escalation to Fable 5. Fable 5 costs twice as much per token, so an Opus 5 run at xhigh or max can spend roughly twice the tokens of a Fable 5 run at high before it costs more. Exhaust the effort ladder on Opus 5 before treating Fable 5 as the answer to a quality gap, unless the failure is a task-shape mismatch that more effort does not address.
5. Apply the premium gate. When the active lead is Fable 5, pause before the first task-execution tool call, state-changing write, or substantive multi-step analysis and wait for confirmation only when a cheaper lane is likely non-inferior, an actionable switch or complete new-task handoff exists, and expected remaining savings exceed duplicated prompt, context, restart, and handoff cost. Since Opus 5 ships near-Fable capability at half the price, treat “a cheaper lane is likely non-inferior” as presumptively satisfied for coding, review, research, and knowledge work; the burden is on task-specific evidence that Fable 5 is needed. Recommend one lane, one task-specific reason, and the supported switch or compact handoff. Do not continue while waiting. If the user chooses Fable 5, continue without repeating the gate. If any condition fails, continue and give at most one future-task reminder.
6. Compare three routes before delegating for cost. Prospectively assess the current single-agent route, the cheapest likely non-inferior single-agent lane, and an appropriate lead with routed workers. Use local history or task-shape evidence; do not execute all three merely to route a one-off task. Use the routed design for cost when—and only when—it is expected to be the cheapest accepted route and the work has independent scopes, compact handoffs, controlled write ownership, and affordable verification. Never manufacture workers to justify a premium lead.
7. Bound the chosen route. Pin every worker model and effort when supported; otherwise do not claim tier-routing savings. Give each worker a goal, owned scope, distilled inputs, output contract, verification including residual semantic review, and a stop condition. Cap fan-out, depth, attempts, retries, and returned context. Higher-cost parallelism requires the explicit higher-cost objective from step 1.
8. Diagnose before escalating. Repair missing or conflicting context first. Retry the same lane for a transient tool failure. Raise effort one level for insufficient checking on a clear task. Raise tier for a reasoning, ambiguity, or state-tracking mismatch that effort does not fix. Lower effort for overexploration and lower tier for repetitive objective volume. Keep one agent for sequential or shared-state work. Change only one of tier, effort, prompt/context, tools, or orchestration at a time; escalate only the failing workstream and route back down after the hard decision is settled.
9. Verify and update policy. Verify against the unchanged contract and account for every lead, worker, verifier, retry, and rescue without double counting. A one-off heuristic may be used for low-risk, strongly verified work, but it does not become a default route. Standardize a cheaper route only after paired representative runs meet a predeclared non-inferiority and adoption rule.

Minimal read-only inspection needed to evaluate the premium gate is allowed before the pause; task execution is not. For a new-task handoff, include the outcome, constraints, acceptance criteria, and necessary paths or evidence so the user does not have to reconstruct the task.

Claude tier and effort mapping

Effort settings (Fable 5, Opus 5, and Sonnet 5 share the same ladder, default high):

• low: bounded work where cost and speed dominate; a reasonable default for mechanical subagents.
• medium: interactive exploration or a balanced pass.
• high: default, and the normal starting point for serious work without a calibrated route.
• xhigh: long-horizon agentic and coding work; the first thing to try when a run under-checks on a clear task.
• max: highest supported setting; reserve it for a demonstrated quality gap, and prefer it on Opus 5 over a tier move to Fable 5.

Model-specific notes: on Opus 5, effort controls thinking volume rather than visible response length, so prompt for brevity instead of lowering effort to shorten output, and do not carry effort settings over from Opus 4.8 without a fresh sweep. On Fable 5, lower effort levels still often exceed xhigh on prior-generation models. At xhigh or max on either model, set a large max_tokens — 64k is a reasonable starting point — because it caps thinking plus response together.

Reduce effort when successful runs explore unused alternatives or produce unactionable detail. Increase it only after ruling out prompt, context, tool, and verification gaps.

Delegation and verification

Cost-routing workers should normally use Haiku, Sonnet 5, or Opus 5 according to task shape; at Opus 4.8-level pricing, Opus 5 is now an affordable implementer lane rather than a reserved one. An independent verifier is different: if the acceptance contract requires fresh-context review, its cost is part of every eligible route and its tier follows residual semantic risk, even when that means Opus 5 or Fable 5.

Prefer one lead when work is sequential, shares mutable state, or would copy most of the parent context. Concurrent writers need disjoint ownership; read-only workers may share scope, and sequential handoffs may touch the same files with one writer at a time. Read references/prompt-patterns.md for worker and verifier packets.

Deterministic gates reduce risk only for the properties they encode. They can justify testing a cheaper worker but never replace semantic review for unencoded requirements. For multi-phase projects where the lead should carry only judgment, the field-validated lead-orchestrator loop (distill to contract files on disk → routed implementers → independent semantic verifier → bounded fixer → lead gates and commits) is documented in references/prompt-patterns.md, with the supporting run and its adoption limits in references/evidence-notes.md (2026-07-25).

In one unversioned Claude Code observation, workers with no explicit model inherited the premium lead and fork-style workers ignored overrides. In a later run (2026-07-25), explicit overrides on 21 non-fork workers behaved as requested, though effective models were pinned, not independently verified. Verify current inheritance and override behavior before relying on it; avoid forks when cost behavior is uncertain.

Prompting and migration

For substantial work, provide outcome, relevant evidence, hard constraints, approval boundaries, acceptance criteria, verification, budget, and stop condition. Let the model choose intermediate steps. Read references/prompt-patterns.md for compact shapes.

When migrating older prompts, remove scaffolding that only compensated for weaker models, but keep safety, data, budget, scope, style, and business constraints. Never ask for hidden reasoning. Re-test prompt tokens, latency, accepted-result cost, timeouts, streaming, and asynchronous UX.

Migrating from Opus 4.8 to Opus 5 has two behavioral breaks: requests with no thinking field now run with adaptive thinking, so revisit max_tokens; and thinking: {"type": "disabled"} is rejected with a 400 at xhigh or max effort. Re-run an effort sweep rather than carrying 4.8 settings across.

Claude 5 API behavior

• Adaptive thinking is on by default across Fable 5, Opus 5, and Sonnet 5; control depth with effort, not a manual thinking budget.
• Raw thinking is not returned. thinking.display is summarized or omitted; preserve returned thinking blocks unchanged in same-model multi-turn conversations.
• max_tokens covers thinking and response output. Give high-effort runs enough room.
• effort is a request-level setting that reshapes the prompt, so changing it mid-conversation invalidates prompt caching. Pick a level per workload and hold it.
• Task budgets cover a full agentic loop and are distinct from max_tokens. They are a Messages API beta and are not supported by Claude Code or Cowork in this snapshot.

Read references/evidence-notes.md before depending on beta headers or current surface support. Read references/api-runtime.md before implementing task budgets, refusals, or fallback.

Long-running execution

• Use asynchronous workers, scheduled checks, or harness-native waiting; never use an unbounded sleep or polling loop inside a tool call.
• Keep the lead context to decisions, risks, current state, and compact evidence.
• Ground progress and completion claims in tool results from the current run.
• Preserve only durable decisions, constraints, corrections, confirmed approaches, and the next action across sessions; keep one compact state snapshot when needed.
• Pause for destructive or irreversible actions, external commitments, real scope changes, or information only the user can provide.
• Report outcome first, then the evidence that determines what to trust or do next.

Read references/operations.md for the complete scope, progress, state, and reporting rules used by analysis-only and long-running work.

Refusals and fallback

Both Fable 5 and Opus 5 run safety classifiers. Detect a refusal with HTTP 200 plus stop_reason: "refusal"; treat any partial result as incomplete. Refusal categories in stop_details.category are monitoring metadata, not a license to bypass safeguards.

Choose exactly one fallback mechanism per request path: server-side fallback where supported, SDK middleware retry, or one manual retry on a different model. Do not combine mechanisms or create a fallback loop. Before retrying side-effecting work, confirm idempotency or that no action completed. Implementation details are in references/api-runtime.md; provenance and confidence boundaries are in references/evidence-notes.md. Reverify both before implementation.

Reference map

• references/prompt-patterns.md: lead, worker, and verifier packets.
• references/routing-and-evaluation.md: cost accounting, one-off routing, and route adoption procedure.
• references/api-runtime.md: thinking, budgets, refusal fields, and fallback matrix.
• references/operations.md: scope control, long-running state, progress, and output.
• references/evidence-notes.md: dated product/API provenance and field observations.

Read only the reference needed for the current task.