prompt-atlas

A curated library of 38 reusable Prompt Cards organized by technical direction. Each card carries metadata, variables, examples, failure modes, and tuning notes — they are work assets, not snippets, and each explicitly states when to use a sibling card instead.

When to invoke this skill

Trigger when the user describes a prompt-engineering task in any of the covered directions, or asks for "a prompt for X" where X falls under:

• RAG: retrieval scoring, multi-hop eval synthesis, query rewriting, HyDE-style hypothetical answers, citation faithfulness, answer hallucination detection, chunk summarization for retrieval.
• Agent: ReAct planning, plan-and-execute, tool-call repair, trajectory reflection, long-context memory compression, multi-agent sub-task delegation.
• RLHF: pairwise preference labeling, pointwise reward scoring, constitutional critique-and-revise, best-of-N selection, defensive red-team prompt generation.
• SFT: instruction-set augmentation, self-instruct generation, (instruction, response) quality filtering, SFT response generation.
• Multimodal: structured image captioning, VLM caption verification, visual question answering with grounding, OCR structured extraction, chart and table extraction.
• Chain-of-Thought: structured reasoning with rationale summary, least-to-most decomposition, self-consistency aggregation, verify- then-finalize, tree-of-thoughts.
• Evaluation: LLM-as-judge rubrics, reference-based judging, per-claim factuality, pointwise quality scoring, safety output classification, position-bias-aware pairwise judging.
• Code: structured code review, test case generation, audience- calibrated code explanation, code evaluation judging, refactor suggestions.

Do not invoke for: jailbreaks, safety-bypass prompts, or attempts to extract proprietary internal reasoning traces. See docs/SAFETY.md.

Routing decision tree

Map the user's described task to the closest direction, then to the closest card. When in doubt between two cards, read both and pick the one whose Purpose section best matches; the Tuning Notes of each card also explicitly compare it to its siblings.

RAG

User describes...                                                                  → Card
─────────────────────────────────────────────────────────────────────────────────────────
"score whether a retrieved passage is relevant to a query"                         → rag/retrieval-relevance-evaluator
"build a multi-hop QA eval question from two passages"                             → rag/multihop-eval-synthesizer
"decompose / rewrite a user query into focused sub-queries before retrieval"       → rag/query-rewriting-decomposition
"generate a hypothetical answer to embed as a search query (HyDE)"                 → rag/hyde-hypothetical-answer-generator
"audit whether a cited span actually supports the claim it was attached to"        → rag/citation-faithfulness-scorer
"detect hallucinations in a RAG answer (per-claim grounding against context)"      → rag/answer-grounding-checker
"summarize a long document chunk for retrieval indexing"                           → rag/chunk-summarizer-for-retrieval
"compress retrieved passages into a smaller question-tailored context"             → rag/context-compression
"resolve a chat follow-up ('what about that?') into a standalone retrieval query"  → rag/conversational-query-resolver
"synthesize an answer from multiple sources, surfacing conflicts and citations"    → rag/multi-source-aggregator
"build a structured output (table / list / record) from RAG sources with citations" → rag/structured-rag-output-builder
"fuse multiple sub-query retrieval results into one ranked deduplicated set"        → rag/query-fusion
"resolve time-relative phrases ('last month', 'recent') into concrete time bounds" → rag/time-aware-retrieval-rewriter

Agent

User describes...                                                                  → Card
─────────────────────────────────────────────────────────────────────────────────────────
"agent loop, ReAct-style, every step emits a JSON tool call"                       → agent/react-planner-with-tool-schema
"produce a complete plan upfront for a goal whose structure is predictable"        → agent/plan-and-execute-planner
"fix a malformed tool call given the validation error message"                     → agent/tool-call-repair
"step back and reflect on whether the trajectory is on track"                      → agent/self-critique-reflection
"compress a long agent trajectory into structured memory before context overflow"  → agent/long-context-memory-summarizer
"split a complex task across specialized workers / agents (multi-agent)"           → agent/sub-task-delegator
"decide whether a goal is too ambiguous to act on; ask one good clarifying question" → agent/clarification-asker
"convert an OpenAPI / Swagger / JSON Schema spec into an agent tool catalog"       → agent/api-spec-to-tool-catalog
"decide retry / abort / escalate when an agent operation fails"                    → agent/error-recovery-strategy
"plan agent execution within a token / dollar budget"                              → agent/budget-aware-planner
"compress a tool's verbose output before adding to agent context"                  → agent/tool-output-summarizer
"reconcile conflicting outputs from multiple sub-agents (factual vs value-based)"  → agent/multi-agent-conflict-resolver
"translate raw API response into a user-readable answer"                           → agent/api-result-translator

RLHF

User describes...                                                                  → Card
─────────────────────────────────────────────────────────────────────────────────────────
"label A vs B preference (HHH dimensions) for reward model data"                   → rlhf/pairwise-preference-labeler
"produce a single-response scalar reward signal (no pair available)"               → rlhf/pointwise-reward-scorer
"critique a response against a constitution and produce a revised version (CAI)"   → rlhf/constitutional-critique-revise
"pick the best of N candidate responses (rank + select)"                           → rlhf/best-of-n-selector
"generate adversarial probe prompts for safety evaluation (defensive only)"        → rlhf/red-team-prompt-generator
"diagnose model refusal calibration (over-refusal vs under-refusal vs correct)"    → rlhf/refusal-calibration-probe
"generate iterative DPO (chosen, rejected) pairs targeting one principle"          → rlhf/iterative-dpo-pair-generator
"score whether a response matches a defined persona / brand voice"                 → rlhf/persona-consistency-judge
"score helpfulness vs harmlessness independently to detect over-cautious failures" → rlhf/helpfulness-vs-harmlessness-tradeoff
"label pairwise preference for long-form (long input + long output) responses"     → rlhf/long-context-preference-labeler
"detect reward hacking patterns (sycophancy, verbosity, format gaming)"            → rlhf/reward-hacking-detector
"audit whether a preference label's rationale actually justifies the pick"         → rlhf/preference-rationalization-judge

SFT

User describes...                                                                  → Card
─────────────────────────────────────────────────────────────────────────────────────────
"rewrite ONE instruction into N diverse variants (same task)"                      → sft/instruction-variant-expander
"generate NEW instructions in the same task family as seed examples"               → sft/self-instruct-from-seed
"filter (instruction, response) SFT pairs by quality before training"              → sft/data-quality-filter
"generate the response half of an SFT pair given an instruction"                   → sft/response-generator
"generate multi-turn conversation SFT data (chat training)"                        → sft/conversation-sft-pair-generator
"select best K few-shot demonstrations from a candidate pool for a query"          → sft/few-shot-example-selector
"generate a response in a defined persona / brand voice (with strictness)"         → sft/persona-controlled-response
"rewrite text in a target style (formal / casual / specific voice)"                → sft/style-transfer
"analyze SFT dataset coverage by topic / skill, find gaps and over-representation" → sft/data-coverage-analyzer
"classify an instruction's difficulty for a target model class"                    → sft/instruction-difficulty-classifier
"generate code-specific (instruction, response) SFT pairs at controlled difficulty" → sft/code-sft-pair-generator
"find semantic near-duplicates in an SFT instruction set (paraphrases, etc)"       → sft/instruction-deduplicator

Multimodal

User describes...                                                                  → Card
─────────────────────────────────────────────────────────────────────────────────────────
"verify whether a candidate caption matches the image (per-claim audit)"           → multimodal/vlm-image-description-verifier
"generate a structured caption (scene, objects, action, salient text)"             → multimodal/structured-caption-generator
"answer a question about an image with grounding region + confidence"              → multimodal/vqa-with-confidence
"extract typed fields from a document image (receipt / invoice / form / ID)"       → multimodal/ocr-structured-extraction
"extract data from a chart / plot / table image"                                   → multimodal/chart-table-extractor
"analyze a document page's layout (title, body, tables, figures, reading order)"   → multimodal/document-layout-analyzer
"extract graph structure from a diagram / flowchart / architecture image"          → multimodal/diagram-to-structured-data
"convert a UI screenshot into a component spec (component tree + layout)"          → multimodal/screenshot-to-spec
"classify an image into custom user-defined categories with confidence"            → multimodal/image-classification
"transcribe handwritten text with per-word confidence"                             → multimodal/handwriting-transcriber
"reverse-engineer an edit instruction from a before/after image pair"              → multimodal/image-edit-instruction-generator
"compare two images and explain similarities / differences"                        → multimodal/image-comparison-explainer

Chain-of-Thought

User describes...                                                                  → Card
─────────────────────────────────────────────────────────────────────────────────────────
"single-pass structured reasoning with sub-steps and a visible rationale"          → cot/structured-reasoning-with-rationale-summary
"decompose a complex compositional problem into easier sub-problems in order"      → cot/least-to-most-decomposition
"aggregate N independently-sampled reasoning paths into a consensus answer"        → cot/self-consistency-aggregator
"draft + verify before committing to a final answer (per-check verdicts)"          → cot/verify-then-finalize
"explore multiple approaches in parallel, evaluate, prune (tree-of-thoughts)"      → cot/tree-of-thoughts
"abstract the question into a principle first, then apply (step-back prompting)"   → cot/step-back-prompting
"critique and revise a candidate reasoning plan before execution"                  → cot/plan-critique-and-revise
"reason with explicit per-step uncertainty and a final confidence range"           → cot/uncertainty-quantification
"reasoning where every claim must cite a provided source"                          → cot/citation-grounded-reasoning
"contrast correct path against an articulated wrong path (anti-misconception)"     → cot/contrastive-self-consistency
"process external criticism of a candidate answer (accept / correct / reject)"     → cot/self-correction-protocol
"generate a reusable meta-prompt for a class of tasks from description + examples" → cot/meta-prompt-generator

Evaluation

User describes...                                                                  → Card
─────────────────────────────────────────────────────────────────────────────────────────
"LLM-as-judge rubric for open-ended outputs, fixed 4-dimension rubric"             → eval/llm-judge-rubric-open-ended
"score a model output against a gold reference (closed-form benchmark)"            → eval/reference-based-judge
"decompose an output into atomic claims and label each true/false/unverifiable"    → eval/per-claim-factuality-judge
"score one output on custom dimensions with self-reported confidence"              → eval/pointwise-quality-scorer
"classify a single output along a harm taxonomy (allow / review / block)"          → eval/safety-output-classifier
"pairwise judge with explicit position-bias detection (two-call protocol)"         → eval/pairwise-judge-with-position-bias-probe
"judge a multi-turn dialogue (per-turn + conversation-level scoring)"              → eval/multi-turn-dialogue-judge
"generate a domain-specific rubric with concrete level anchors (1-5)"              → eval/rubric-generator
"compare baseline vs candidate outputs and detect quality regressions"             → eval/regression-detector
"diagnose LLM judge biases (length / position / format / verbosity)"               → eval/judge-bias-probe
"check confidence calibration (predicted confidence vs actual accuracy)"           → eval/calibration-checker
"design a small human eval study with rubric, calibration plan, analysis"          → eval/human-eval-bootstrap
"build a multi-benchmark leaderboard with weighting and differentiator analysis"   → eval/leaderboard-builder

Code

User describes...                                                                  → Card
─────────────────────────────────────────────────────────────────────────────────────────
"structured code review with per-dimension findings (correctness/security/etc.)"   → code/code-review-checklist
"generate test cases for a function with happy/edge/error coverage"                → code/test-case-generator
"explain code at a specific audience level (junior dev / PM / domain expert)"      → code/code-explanation-generator
"judge whether candidate code fulfills a task (with optional gold + tests)"        → code/code-eval-judge
"suggest concrete refactors with rationale and impact (single goal per call)"      → code/refactor-suggestion
"translate code from one language to another (literal / idiomatic / balanced)"     → code/code-translation
"focused code security review with CWE-style findings and threat model"            → code/security-review
"summarize a git diff into a structured PR description (changes + risks + tests)"  → code/code-summary-for-pr
"plan a major version migration grounded in actual code (phased + ordered)"        → code/migration-plan-generator
"analyze the impact of changing a function / API signature on a codebase"          → code/dependency-impact-analyzer
"explain a stack trace / error message at a specific audience level"               → code/error-message-explainer
"generate a conventional / imperative / verbose commit message from a diff"        → code/commit-message-generator
"review an API design (REST / GraphQL / gRPC) for ergonomics and consistency"      → code/api-design-reviewer

For tasks not covered above:

1. Check INDEX.md (auto-generated) for the full card list grouped by direction and tag.
2. If still no match, the closest direction's existing cards may be adaptable — read their Tuning Notes sections for adjacent use cases.
3. If no card fits, tell the user and suggest opening an issue using .github/ISSUE_TEMPLATE/new-prompt-card.yml to request the card.

Tag dictionary (for fuzzier matches)

When the user's words don't directly map to a direction, search by tag intent:

• "score / rate / judge / evaluate" → scoring, llm-judge, rubric
• "label / annotate / classify" → preference-labeling, classification, pairwise
• "synthesize / generate examples / new instructions" → generation, synthesis, seed-expansion, data-augmentation
• "multi-step / decompose / break down / plan" → planning, decomposition, decomposition-cot, react
• "verify / check / ground / fact-check" → factuality, grounding, self-check, citation
• "multi-hop / cross-passage / chain" → multi-hop
• "tool / function call / repair / schema" → tool-use, structured-output
• "image / picture / visual / VLM" → vision, image-description, vlm-eval
• "reflection / step back / look at trajectory" → reflection, self-check
• "memory / compress / summarize trajectory / resume later" → memory, rationale-summary
• "constitution / critique-revise / harmless rewrite" → harmlessness, helpfulness, honesty
• "filter / quality / drop bad data" → scoring, classification, instruction-tuning
• "safety / harm / block / red-team / harmlessness" → safety, harmlessness
• "rationale / reasoning / sub-steps / least-to-most / self-consistency" → structured-reasoning, rationale-summary, decomposition-cot, self-check
• "reward model / RM training / preference data" → reward-modeling, preference-labeling, pairwise

The full controlled tag vocabulary is in docs/SCHEMA.md.

How to use a card

1. Read the entire card. Frontmatter + all six sections. Skipping Failure Modes and Tuning Notes is the most common cause of unsatisfying results — those sections are where the experience lives.
2. Match the user's variables to the card's variables block. If a required variable is missing, ask the user for it.
3. Substitute {{variable}} placeholders in the card's prompt body.
4. Match the model class to the card's models field. If the user is on a model class the card has not been validated for, mention this and adjust per Tuning Notes.
5. Surface the output schema. If the card's output_schema is structured-json, ensure the calling environment can parse JSON; if the user is in a freeform chat, propose using a JSON-mode call.
6. Check the sibling cards. Most cards' Tuning Notes name an adjacent card and explain when to switch (e.g. eval/reference-based-judge says use it for closed-form, eval/llm-judge-rubric-open-ended for open-ended). Verify the chosen card is the right sibling for the user's actual task.

Safety

Every card is reviewed against docs/SAFETY.md. If a user request would require violating that policy (jailbreaks, hidden-CoT extraction, harm-enabling content), do not adapt a card to fit — refuse and explain.

The eval/safety-output-classifier card is itself defensive (it detects harm to filter or label); do not invert it to generate harmful content.

Repository layout (reference)

prompt-atlas/
├── prompts/<direction>/<slug>.md     ← the cards (100 total)
├── templates/prompt-card.md          ← canonical template
├── docs/SCHEMA.md                    ← frontmatter + tag vocabulary
├── docs/SAFETY.md                    ← policy
├── INDEX.md                          ← auto-generated catalog
└── scripts/{validate,build_index}.py ← maintenance