PromptingCraft.com

Sources

Official recommendations, with their receipts.

This page is generated from a validated snapshot of PromptingCraft.com’s canonical knowledge ledger — not from separate marketing copy. Each entry links to the provider’s documentation and to the exact line of the ledger at a pinned commit.

Ledger file
skills/promptcraft/references/knowledge.json
Pinned commit
d0babf8f
Curated on
2026-10-12
Snapshot taken
2026-10-12
File SHA-256
68d7e4c3b32e123a2a7940a413e7e9b0b0cc7588ed1a6e007fa4090c88de4249
Records
17 official · 1 methodology

Dated, not live. Dated retrieval snapshot; revalidate applicable sources during use. Retrieval is not publication date. Retrieval dates are when the source was read, not when the provider published it. When your agent can fetch the web, the skill revalidates relevant sources during use.

What the labels mean

OFFICIAL

Primary provider recommendation, scoped and conditional.

MODEL_SPECIFIC

Applies only to the named model family/version/interface.

GENERAL

Provider describes a broader technique; still not a universal mandate.

INFERRED

PromptingCraft.com adaptation, not an official provider claim.

EXPERIMENTAL

Effectiveness unestablished outside the stated evaluation.

Only documentation from these official hosts can justify model-specific prompting advice: Anthropic — code.claude.com, platform.claude.com, www.anthropic.com; OpenAI — developers.openai.com, learn.chatgpt.com.

Anthropic

9 records
  1. OFFICIALGENERAL
    claude-clear

    State output and constraints clearly; request above-and-beyond behavior explicitly rather than relying on inference.

    Applies to
    Current Claude families covered by the living page
    Limitation
    Examples and XML are techniques, not required formats for every task.
    How PromptingCraft.com uses it
    User requirements set scope; a recommendation adds no mandatory deliverable.
    Apply when
    Claude target; apply only relevant techniques. Not mandatory for every prompt.
    Known tension
    Generic completeness and speculative edge cases cannot extend the task.

    Anthropic: Prompting best practices· retrieved 2026-10-12· content hash 057fbadeb4e8…ledger line 166 ↗

  2. OFFICIALGENERAL
    claude-long-context

    For large document inputs, place long documents first and the query later; identify document sources and ground answers in relevant passages.

    Applies to
    Current Claude families covered by the living page
    Limitation
    Reported gains are provider test results for long/multidocument tasks, not PromptingCraft.com measurements.
    How PromptingCraft.com uses it
    Preserve source identity in long handoffs; use only relevant context.
    Apply when
    Large supplied documents; host context limits permit it. Not mandatory for every prompt.
    Known tension
    Do not require XML, quotations or a long-context layout for a small fix.

    Anthropic: Prompting best practices· retrieved 2026-10-12· content hash 057fbadeb4e8…ledger line 185 ↗

  3. OFFICIALMODEL_SPECIFIC
    claude-sonnet-scope

    If unsolicited additions occur, instruct the model to finish requested work and checks, then report helpful extras as suggestions.

    Applies to
    Claude Sonnet 5.5 only
    Limitation
    Initiative varies with effort; supplied system-prompt examples do not replace risk/authority rules.
    How PromptingCraft.com uses it
    A documented model-specific reason to separate optional extras from required scope.
    Apply when
    Executor model established as Sonnet 5.5 and the observed behavior matches. Not mandatory for every prompt.
    Known tension
    Do not assume the executor is Sonnet 5.5 or prohibit user-requested reviews/tests.

    Anthropic: Prompting Claude Sonnet 5.5· retrieved 2026-10-12· content hash 6886cdd2625b…ledger line 204 ↗

  4. OFFICIALMODEL_SPECIFIC
    claude-json-reasoning

    For the documented JSON reasoning failure at low/medium effort, test the guide's think-before-answer instruction or adaptive higher effort.

    Applies to
    Claude Sonnet 5.5 API reasoning tasks with JSON output
    Limitation
    This is a specific accuracy intervention; effort/thinking controls belong to API configuration and affect latency/cost.
    How PromptingCraft.com uses it
    Record the conflict rather than making either provider's reasoning recipe universal.
    Apply when
    This exact model/interface/failure is evidenced; not an unknown coding-agent handoff. Not mandatory for every prompt.
    Known tension
    Differs from generic OpenAI reasoning advice; neither justifies demanding hidden chain-of-thought.

    Anthropic: Prompting Claude Sonnet 5.5· retrieved 2026-10-12· content hash 6886cdd2625b…ledger line 223 ↗

  5. OFFICIALGENERAL
    claude-progressive

    Keep skill instructions concise, choose degrees of freedom by task fragility, and load supporting references progressively.

    Applies to
    Claude Agent Skills
    Limitation
    Metadata, instructions and references have different loading costs; body guidance alone is not behavior proof.
    How PromptingCraft.com uses it
    One shared core with narrowly loaded knowledge/provider notes.
    Apply when
    Authoring/maintaining the skill. Not mandatory for every prompt.
    Known tension
    Do not load a full provider manual or turn illustrative workflows into fixed requirements.

    Anthropic: Skill authoring best practices· retrieved 2026-10-12· content hash 727f3fa1168e…ledger line 242 ↗

  6. OFFICIALGENERAL
    claude-tool-descriptions

    Explain what each tool does, when to use it, inputs, outputs and limitations.

    Applies to
    Claude API tool-capable models
    Limitation
    Detailed API tool descriptions are distinct from general user-task verbosity.
    How PromptingCraft.com uses it
    Specify only confirmed available tools when useful; preserve action permissions.
    Apply when
    Tool-building tasks or confirmed executor capabilities. Not mandatory for every prompt.
    Known tension
    No requirement to provision tools for an ordinary optimized prompt.

    Anthropic: Define tools· retrieved 2026-10-12· content hash bc9b98eec352…ledger line 261 ↗

  7. OFFICIALGENERAL
    claude-evaluations

    Define specific, measurable, achievable and relevant criteria; use task-specific cases and multiple dimensions, including context and privacy.

    Applies to
    Claude applications and prompt iterations
    Limitation
    Example thresholds are examples, not user targets. Graders require validation; volume does not cure mislabeled requirements.
    How PromptingCraft.com uses it
    Compare requirement fidelity, clarity, readiness, suitability and length separately.
    Apply when
    Evaluating a prompt change; use existing infrastructure. Not mandatory for every prompt.
    Known tension
    Do not import example numeric quality thresholds or use self-awarded scores.

    Anthropic: Define success criteria and build evaluations· retrieved 2026-10-12· content hash not recordedledger line 280 ↗

  8. OFFICIALGENERAL
    claude-code-authoring

    Use a specific discovery description and SKILL.md with supporting files loaded as needed; distinguish portable fields from Claude Code extensions.

    Applies to
    Claude Code skills, independent of executor model
    Limitation
    Native invocation controls and subagent options are host-specific.
    How PromptingCraft.com uses it
    Keep standalone Codex/manual ChatGPT behavior on the same contract.
    Apply when
    Packaging/testing Claude Code support. Not mandatory for every prompt.
    Known tension
    Do not force Claude-only frontmatter or separate engines onto all surfaces.

    Anthropic: Extend Claude with skills· retrieved 2026-10-12· content hash 35fd2c11f870…ledger line 298 ↗

  9. OFFICIALGENERAL
    claude-workflow-complexity

    Choose the simplest workflow that solves the task and add orchestration only when its quality benefit justifies latency and cost.

    Applies to
    General agent/workflow design; historical 2024 engineering article
    Limitation
    Article explicitly says its tooling landscape changed since 2024; use its design distinction, not old setup commands.
    How PromptingCraft.com uses it
    One composition/revision/contract-review core; no mandatory multi-agent pipeline.
    Apply when
    Designing a workflow; not a directive to change user architecture. Not mandatory for every prompt.
    Known tension
    Do not add paid services or delegation merely to improve a prompt.

    Anthropic: Building effective agents· retrieved 2026-10-12· content hash not recordedledger line 317 ↗

OpenAI

8 records
  1. OFFICIALGENERAL
    oai-clear-context

    State the desired result and relevant supplied context; separate instructions from examples/reference data.

    Applies to
    OpenAI text and reasoning models
    Limitation
    Model snapshots differ; evaluate changes on the actual target.
    How PromptingCraft.com uses it
    Keep the shared contract legible without prescribing unchosen implementation details.
    Apply when
    Any prompt; adapt structure to complexity. Not mandatory for every prompt.
    Known tension
    Broad acceptance checklists can introduce requirements.

    OpenAI: Prompt engineering· retrieved 2026-10-12· content hash 3a077358aaa7…ledger line 14 ↗

  2. OFFICIALGENERAL
    oai-instruction-roles

    Respect the API instruction hierarchy and distinguish developer instructions from user input.

    Applies to
    OpenAI API message roles
    Limitation
    A copied user prompt cannot replace host/system/developer authority.
    How PromptingCraft.com uses it
    Preserve host controls; user contract overrides this skill's preferences within them.
    Apply when
    API authoring or explaining authority; no API configuration in ordinary CLI tasks. Not mandatory for every prompt.
    Known tension
    The skill priority statement must not imply user text overrides host security rules.

    OpenAI: Prompt engineering· retrieved 2026-10-12· content hash 3a077358aaa7…ledger line 33 ↗

  3. OFFICIALMODEL_SPECIFIC
    oai-reasoning

    Use simple direct outcome instructions and explicit constraints; try zero-shot before adding examples; omit generic step-by-step reasoning requests.

    Applies to
    Documented OpenAI reasoning models; historical o-series particulars stay model-specific
    Limitation
    Historical formatting/developer-role details are version-specific; examples can help or hurt.
    How PromptingCraft.com uses it
    Do not impose a reasoning script on an unknown executor.
    Apply when
    Known reasoning target and relevant interface. Not mandatory for every prompt.
    Known tension
    No universal reasoning-before-output rule; no mandatory examples or verbose process.

    OpenAI: Reasoning best practices· retrieved 2026-10-12· content hash 00d3f8ee705e…ledger line 52 ↗

  4. OFFICIALGENERAL
    oai-evaluation

    Use task-specific representative, edge and adversarial cases; compare versions against explicit criteria and calibrate automated judgments with human feedback.

    Applies to
    OpenAI models and LLM applications
    Limitation
    Scores alone are insufficient; synthetic cases and model judging do not establish production improvement.
    How PromptingCraft.com uses it
    Report contract fidelity separately from length and package checks.
    Apply when
    Changing prompt/skill behavior; existing offline harness is sufficient infrastructure. Not mandatory for every prompt.
    Known tension
    Existing three-case sample and self-review cannot justify behavioral validation.

    OpenAI: Evaluation best practices· retrieved 2026-10-12· content hash 93aeca2e3b1f…ledger line 71 ↗

  5. OFFICIALGENERAL
    oai-structured

    Use a schema when the application needs structured responses, and function calling when connecting tools; handle refusals/incomplete output.

    Applies to
    OpenAI models/interfaces supporting Structured Outputs
    Limitation
    Schema adherence does not prevent mistakes or hallucinated contents.
    How PromptingCraft.com uses it
    Optional contract annotations get local validation and semantic review.
    Apply when
    Machine-consumed output is requested or needed; ordinary prompts need no JSON. Not mandatory for every prompt.
    Known tension
    A JSON-shaped contract or checker exit zero is not a semantic preservation verdict.

    OpenAI: Structured model outputs· retrieved 2026-10-12· content hash 09aa3814251c…ledger line 90 ↗

  6. OFFICIALGENERAL
    oai-tools

    Describe actual tool purpose, parameters, outputs and when to use it; prefer predictable interfaces and constrained invalid states.

    Applies to
    OpenAI API tool-capable models
    Limitation
    Tool availability and strictness vary by model; examples may hurt reasoning models.
    How PromptingCraft.com uses it
    Do not invent tool names; carry confirmed capabilities/authority into a handoff.
    Apply when
    The user is building/configuring tools or agent capabilities. Not mandatory for every prompt.
    Known tension
    An API tool-schema recommendation is not a requirement to install MCP or add tools.

    OpenAI: Function calling· retrieved 2026-10-12· content hash ce460da5635f…ledger line 109 ↗

  7. OFFICIALGENERAL
    oai-untrusted-data

    Keep untrusted text out of privileged developer instructions and limit what data can flow to tools.

    Applies to
    OpenAI agent workflows; page is legacy Agent Builder guidance
    Limitation
    Mitigations reduce risk but cannot guarantee prevention; legacy workflow examples are not current product mandates.
    How PromptingCraft.com uses it
    Treat retrieved docs as reference data, never grants to upload secrets or deploy.
    Apply when
    Retrieved/tool content is included in a prompt. Not mandatory for every prompt.
    Known tension
    No requirement for a new guardrail SaaS or an approval step on every local action.

    OpenAI: Safety in building agents· retrieved 2026-10-12· content hash e3902b08bb7f…ledger line 128 ↗

  8. OFFICIALGENERAL
    oai-cache

    Keep useful shared instructions/reference prefixes stable and put changing task details later; measure total task cost and quality.

    Applies to
    OpenAI cache-eligible models; rules vary by model and settings
    Limitation
    Prefix matches alone may lack an eligible breakpoint. Thresholds, write costs and lifetimes vary; caching generates a new answer.
    How PromptingCraft.com uses it
    Keep entrypoint compact and disclose references on demand.
    Apply when
    Repeated API calls with measured eligible reuse, not ordinary manual chat. Not mandatory for every prompt.
    Known tension
    Do not pad every prompt to reach a cache threshold or treat cached tokens as task savings.

    OpenAI: Prompt caching· retrieved 2026-10-12· content hash 74dc45350d5d…ledger line 147 ↗

PromptingCraft.com methodology

Our own adaptations are recorded in the same ledger but are never labelled official.

  1. INFERREDEXPERIMENTAL
    contract-provenance

    Bidirectionally trace original classified requirements and final operative clauses; structural checks leave semantic review pending.

    Evidence
    Historical public v0.1.0 visible small-bug output introduced retention/network behavior. Candidate effectiveness is measured in quality-1, not established by this record.
    Limitation
    Annotations can misclassify entailment or omit an original requirement. This intervention is not a provider-validated guarantee.

    ledger line 337 ↗

Your requirements come first

Nothing on this page is a requirement for your task. A recommendation can shape how a prompt is written; it can never add a requirement, grant permission or override what you asked for.

Website research is not prompting evidence

This website’s design drew on screens browsed in Mobbin, and its page structure on keyword data from Ubersuggest. Those tools informed design and search decisions only. They are not, and cannot be, sources for how to prompt a model. Research disclosure.