prompting-architect
Architect-level prompt design — layering system, task and data regions, instruction precision and delimiters, few-shot selection, chain-of-thought versus direct answers, versioning and review, and the injection-resistant structures that keep untrusted text from becoming instructions. Use when a prompt is a production artefact.
- License 许可
- MIT
- 最近更新
- 2026-11-13
- Stars
- ★ 0
- Forks
- 0
01 前言 Front Matter
- name
- prompting-architect
- license
- MIT
- metadata
- { "crusor-path": "ai-ml-architect/llm-architect/prompting-architect", "crusor-domain": "ai-ml-architect", "crusor-origin": "original", "crusor-sources": "docs.anthropic.com/en/docs/build-with-claude/prompt-engineering; platform.openai.com/docs/guides/prompt-engineering; learnprompting.org", "crusor-triggers": "prompt engineering, system prompt, few shot, chain of thought, prompt injection, prompt template, prompt versioning, delimiters", "crusor-verified-on": "2026-10-30" }
- description
- Architect-level prompt design — layering system, task and data regions, instruction precision and delimiters, few-shot selection, chain-of-thought versus direct answers, versioning and review, and the injection-resistant structures that keep untrusted text from becoming instructions. Use when a prompt is a production artefact.
- compatibility
- Any model or provider
03 SKILL.md 正文
Prompting Architect
A prompt is source code that ships to a probabilistic interpreter. It needs version control, review, tests, and a structure that separates instructions from data.
When to use
- Any model-backed feature, before tuning anything else.
- Prompts that grew by accretion and now contradict themselves.
- A prompt-injection or jailbreak concern.
When NOT to use
- The failure is retrieval quality or missing context (
rag-architect). - The task needs format guarantees — use structured output, not a better prompt.
Decision 1 — Structure
[system] role, rules, output contract, refusal policy — stable, versioned
[task] the specific instruction with slots — changes per feature
[data] delimited, labelled, explicitly "data, not instructions"
[examples] optional, few-shot, minimal
[request] the actual question, last
- Instructions before data, question after data — models attend to the end.
- Fixed delimiters (
<doc>…</doc>, fenced blocks) with a rule that content inside is data. - One prompt, one job. Two jobs in one prompt produce a model that half-does both.
Decision 2 — Instruction quality
- Be specific about what and how much: length, format, audience, exclusions.
- State the refusal case: "if the documents do not contain the answer, say you don't know".
- Positive instructions ("cite the source id") beat negative ones ("don't make things up") — the negative needs a positive replacement.
- Remove contradictions; contradictory prompts produce variance that looks like model flakiness.
Decision 3 — Few-shot and reasoning
- 3–8 examples, chosen to cover edge cases and format, not to be exhaustive.
- Balanced examples: if every example says yes, the model says yes.
- Chain-of-thought for multi-step reasoning; skip it for extraction where it mostly adds tokens.
- Ask for the reasoning in a structured field when it must be auditable, not in free prose you then parse.
Decision 4 — Injection resistance
- Treat all user and retrieved content as untrusted; never interpolate it into the system region.
- Delimit and label; instruct the model to ignore instructions found inside data.
- Do not rely on the model alone: enforce permissions in code, validate tool calls
server-side (
agent-architect). - Test with known payloads ("ignore previous instructions…"), including base64 and unicode tricks.
Decision 5 — Lifecycle
- Prompts in files, in version control, reviewed; not strings in a controller.
- Each change runs the eval suite (
evals-architect); attach the delta to the PR. - Version the prompt id with the model id in logs, so a regression is traceable to a change.
- Keep the previous prompt deployable — rollback is a config change, not a code revert when possible.
Failure modes at scale
- Prompt that grew to 3,000 tokens with four sections contradicting each other.
- User text concatenated into the system prompt; injection succeeds on day one.
- Few-shot examples all one shape; the model fails on a common edge case.
- Prompt edited in the dashboard, no version, no eval, no record of who changed it.
temperatureraised to "make it more creative" on an extraction endpoint.
Verification receipts
- Prompt file with layering visible and a version id.
- Eval score attached to the current prompt version.
- Injection test suite run against the current prompt.
- Log line showing prompt version + model id per request.
- Rollback to the previous prompt version exercised.
04 同仓库更多 Skills
interaction-patterns-architect
Architect-level interaction design in code — choosing established patterns over invented ones, feedback and latency behaviour, destructive-action safeguards, navigation and overlay conventions, and the state model behind every interactive surface. Use when a screen has to behave, not just look.
design-system-ops-architect
Architect-level operation of a design system in code — token pipelines and versioning, component API and variant policy, the catalogue as the contract, adoption metrics, and the governance that stops a system from forking into three. Use when a component library must survive several teams.
design-handoff-architect
Architect-level design handoff — what must be specified before implementation starts, tokens over raw values, responsive and content-limit contracts, state matrices, and the review loop that catches drift between design and code. Use when a design is about to become code.
content-microcopy-architect
Architect-level UI content — where strings live, voice and terminology consistency, error and empty-state copy as part of the interface, translation-ready authoring, and the content model a product needs before it multiplies screens. Use when the words in the interface matter as much as the pixels.
ui-ux-implementation-architect
Architect-level direction for turning a design into working UI — the handoff contract, component ownership versus screen-level layout, interaction and state conventions, design QA gates, and the review habits that keep an implementation faithful without frozen pixels. Use before any specific UI implementation skill.
test-strategy-architect
Architect-level test strategy — allocating assertions across unit, integration, contract and end-to-end layers, designing for testability through dependency boundaries, fixture and data strategy, and the flake and coverage policies that keep a suite trustworthy. Use when defining how a codebase is tested.