Formulary claude

Formulary: Claude Prompting by Symptom, Remedy, and Dosage

This skill encodes Anthropic's official prompting guidance. Its goal is maximum output quality, not token savings. When quality and cost conflict, choose quality: higher effort up front usually reduces total turns and rework, so it is often cheaper end to end anyway.

When NOT to use this skill

Do not load this skill for: prompting non-Claude models (OpenAI GPT, Gemini, Llama, Mistral), Anthropic account, billing, or rate-limit questions, Claude Code configuration (settings, hooks, MCP servers), or coding tasks that involve no prompt writing. For SDK code and API integration mechanics beyond the parameters named here, prefer a dedicated API skill if one is available.

Workflow

Follow these steps in order whenever building or fixing a prompt:

1. Pick the model using the table below. Never guess model IDs.
2. Set the API parameters correctly for that model. Read references/models-and-params.md before writing any API call; it is the single source of truth for thinking config, effort, max_tokens, and the parameter combinations that return 400 errors.
3. Write the prompt using the universal techniques in references/core-techniques.md (clarity, motivation, examples, XML structure, role, long-context ordering).
4. Apply model-specific tuning from the per-model file. These are behavioral, not API-level: the same prompt lands differently on each model.
5. Fix symptoms with proven snippets. references/snippet-library.md maps observed problems to verbatim, Anthropic-tested prompt snippets, each with a number. Prefer these exact snippets over improvising your own wording.

For end-to-end demonstrations of this workflow, see the worked examples table below.

Model selection (quality-first)

Task	Model	ID
Hardest reasoning, multi-day agentic runs, end-to-end deliverables, deep debugging	Claude Fable 5	claude-fable-5
Default for everything serious: coding, agents, knowledge work, writing, review	Claude Opus 4.8	claude-opus-4-8
High-volume production, fast interactive coding, cost-sensitive agents	Claude Sonnet 5	claude-sonnet-5
Existing pipelines already tuned for it	Claude Sonnet 4.6	claude-sonnet-4-6
Simple classification, routing, subagent fan-out, latency-critical	Claude Haiku 4.5	claude-haiku-4-5

Rules of thumb:

• Default to claude-opus-4-8 unless there is a specific reason not to. Never silently downgrade for cost.
• Reach for claude-fable-5 when the task is above what prior models could do: long autonomous runs, first-shot builds of well-specified systems, navigating real ambiguity. Testing it only on easy workloads undersells it.
• Fable 5 is not intended for offensive cybersecurity or biology and life-sciences work (safety classifiers return stop_reason: "refusal"; configure fallback to Opus 4.8).
• Use exact ID strings as written. No date suffixes.

Quality-first defaults (summary)

The authoritative rules live in references/models-and-params.md; this is the one-line summary. Effort: high floor for intelligence-sensitive work, xhigh for coding and agents. Thinking: adaptive on current models, always-on for Fable 5, never budget_tokens. max_tokens: 64K+ at high effort, stream above ~16K. Sampling params and assistant prefill: rejected on the newest models (4.6 partially allows sampling; see references/models-and-params.md), steer with prompting instead.

Reference routing

Read the file matching your current need; do not load all of them by default.

Need	Read
Model IDs, thinking/effort/API params, hard 400 errors	references/models-and-params.md
Universal techniques: clarity, examples, XML, role, long context, formatting, tool use, agentic systems	references/core-techniques.md
Prompting Claude Fable 5 (long runs, memory, subagents, autonomy)	references/fable-5.md
Prompting Claude Opus 4.8 (verbosity, effort, subagents, frontend, code review)	references/opus-4-8.md
Prompting Claude Sonnet 5 or Sonnet 4.6	references/sonnet.md AND references/opus-4-8.md (Sonnet shares most Opus 4.8 patterns; read both)
Symptom-to-snippet lookup (verbatim Anthropic-tested prompt blocks, numbered 1-37)	references/snippet-library.md

Worked examples

Scenario	File
Coding agent narrates too much, summaries unreadable	examples/01-verbose-coding-agent.md
Choosing model and params for a code review pipeline	examples/02-code-review-pipeline.md
Support assistant answers from memory instead of searching	examples/03-assistant-wont-search.md

Non-negotiables

1. Use verbatim snippets when one exists. The snippets in this skill were tested by Anthropic; a paraphrase is an untested prompt. Cite them by number when recommending them.
2. State the goal and constraints, not the steps. Current models reason better than a hand-written plan. Over-prescriptive prompts written for older models actively reduce quality on Fable 5 and Opus 4.8; A/B with the scaffolding removed.
3. Give the reason behind the request (snippet 34). Context beats inference.
4. Full task spec in the first turn for agentic work. Well-specified single-turn kickoffs outperform progressive clarification across turns, on both quality and tokens.
5. Dial back aggressive language. "CRITICAL: You MUST use this tool" overtriggers on current models. Write "Use this tool when..." instead.
6. Positive instructions beat negative ones. "Write flowing prose paragraphs" beats "Do not use markdown".
7. Measure. After any prompt change, test against your success criteria; do not assume the direction of the effect.
8. Verify freshness. Model IDs, beta headers, and behavioral defaults change. This skill was built from docs fetched on the date in the frontmatter metadata; confirm live via the Models API or platform.claude.com when precision matters.