Changelog

All notable changes to prompt-atlas are documented here. The format follows Keep a Changelog and this project adheres to Semantic Versioning at the repository level. Individual Prompt Cards carry their own version field and changelog inside the card frontmatter / body.

[Unreleased]

Post-v0.1.0 polish + expansion. Will become v0.1.1 (or v0.2.0 if breaking schema changes accumulate).

Added

Beginner-friendly experience

• docs/QUICKSTART.md — 5-minute walkthrough for non-technical users (covers what {{variable}} means, what JSON output means, where to find the right card). Bilingual.
• README "I want to..." section — 38 user-facing goals mapped to exact card paths, grouped by intent.
• ## Quick Use section is now a required body section on every card. Three lines (Use when / Fill in / You'll get) in plain English. Required-section list grew from 6 to 7; validate.py enforces it.
• INDEX.md "Use when" column auto-extracted from each card's Quick Use line.

New Prompt Cards (6)

• rag/chunk-summarizer-for-retrieval — produce search-friendly summaries of long document chunks for retrieval indexing.
• agent/sub-task-delegator — multi-agent foundation: split a complex task across specialized workers with explicit input/output contracts and dependency edges.
• rlhf/red-team-prompt-generator — defensive safety probe generator (refuses sexual_minors category outright; embeds expected refusal posture in each probe).
• multimodal/chart-table-extractor — extract structured data from chart, plot, and table images with per-data-point confidence.
• cot/tree-of-thoughts — explore multiple distinct reasoning branches, evaluate, prune.
• eval/pairwise-judge-with-position-bias-probe — pairwise judge with explicit two-call position-bias detection protocol.
• sft/conversation-sft-pair-generator — multi-turn conversation SFT data generator with explicit multi-turn-behavior coverage.
• sft/few-shot-example-selector — pick best K demonstrations from a candidate pool for a target query, ordered for primacy effect.
• eval/multi-turn-dialogue-judge — judge multi-turn dialogues with per-turn and conversation-level scoring.
• rag/context-compression — compress retrieved passages into a smaller question-tailored context with verbatim spans and source citations.

New code direction (5 cards)

• code/code-review-checklist — structured per-dimension code review (correctness / readability / performance / security / testability / idiomaticity) with severity-tagged findings and approve / request- changes / comment-only verdict.
• code/test-case-generator — generate test specs for a function with explicit happy-path / edge-case / error-handling / boundary / regression coverage and per-test priority.
• code/code-explanation-generator — audience-calibrated code explanation (junior / senior / non-technical / domain-expert) with key concepts and likely confusion points.
• code/code-eval-judge — judge candidate code against a task description, optional reference, and optional test cases. Strict security gate: security<=2 cannot pass.
• code/refactor-suggestion — concrete refactor suggestions oriented toward a single goal (readability / performance / testability / modularity / type_safety) with behavior-change-risk and test-breakage-risk estimates.

Vocabulary additions: code direction; tags code-review, test-generation, documentation.

Round-balance expansion (6 cards across 6 directions)

• code/code-translation — translate code between languages with literal / idiomatic / balanced strategies; flags behavioral differences and untranslatable constructs.
• agent/clarification-asker — front-door of agent loops: decide proceed / clarify / out_of_scope on ambiguous goals; ask exactly one high-value clarifying question.
• cot/step-back-prompting — abstract a question into a principle question, answer the principle, then apply back to the original.
• eval/rubric-generator — generate a domain-specific scoring rubric with concrete level-1/3/5 anchors and weighting recommendation.
• multimodal/document-layout-analyzer — identify regions on a document page (title / body / table / figure / caption / footer etc.) with reading order and hierarchy.
• rlhf/refusal-calibration-probe — diagnose whether a response is appropriately calibrated to prompt safety: catches over_refusal on benign prompts AND under_refusal on unsafe prompts.

Per-direction now: rag 8, eval 8, agent 7, sft 6, code 6, cot 6, multimodal 6, rlhf 6. Total 53. All directions ≥6.

Batch expansion to 68 (15 new cards)

• rag/conversational-query-resolver — multi-turn RAG follow-up resolver: turn "what about that?" into a standalone retrieval query.
• rag/multi-source-aggregator — synthesize an answer from multiple retrieved sources, surface conflicts, attribute citations per-claim.
• agent/api-spec-to-tool-catalog — convert OpenAPI / Swagger / JSON Schema into the tool catalog format used by agent loops.
• agent/error-recovery-strategy — decide retry / abort / escalate on operation failure with backoff suggestions.
• rlhf/iterative-dpo-pair-generator — generate (chosen, rejected) DPO pairs targeting one specific behavioral principle, with meaningful_delta and incidental_changes audit fields.
• rlhf/persona-consistency-judge — score whether a response matches a defined persona on multiple dimensions.
• sft/persona-controlled-response — generate responses in a defined persona with strict / balanced / loose strictness.
• sft/style-transfer — rewrite text in a target style with meaning-fidelity controls.
• multimodal/diagram-to-structured-data — extract graph structure (nodes / edges / relations) from diagram images.
• multimodal/screenshot-to-spec — convert UI screenshot into a framework-agnostic component spec.
• cot/plan-critique-and-revise — critique a candidate plan before execution and produce a revised version.
• cot/uncertainty-quantification — reasoning with per-step confidence + evidence-type labels and a final confidence range.
• eval/regression-detector — detect quality regressions of candidate outputs against baseline on specific dimensions.
• eval/judge-bias-probe — diagnose systematic biases (length / position / format / verbosity / self-preference) in an LLM judge.
• code/security-review — focused code security review calibrated to threat model (web_server / data_processor / client_app / library / internal_tool); CWE-style findings.

Per-direction now: rag 10, eval 10, agent 9, sft 8, multimodal 8, cot 8, rlhf 8, code 7. Total 68.

Batch 2 to 83 (15 more cards)

• code/code-summary-for-pr — git diff → structured PR description with categorized changes, risks, and test suggestions.
• code/migration-plan-generator — phased plan for major version upgrades grounded in actual code patterns.
• code/dependency-impact-analyzer — assess impact of changing a function / API signature; classifies call sites by required work.
• agent/budget-aware-planner — plan agent execution under token / dollar budget; surfaces tradeoffs honestly.
• agent/tool-output-summarizer — compress verbose tool output before context inclusion (between tool execution and next reasoning step).
• rlhf/helpfulness-vs-harmlessness-tradeoff — score the two HHH axes independently and identify over_cautious / unsafe_helpful failures.
• rlhf/long-context-preference-labeler — pairwise preference for long-form responses with per-section judgments and long-form failure mode detection.
• sft/data-coverage-analyzer — diagnose SFT dataset coverage by topic / skill; identify gaps and over-representation.
• sft/instruction-difficulty-classifier — classify per-instruction difficulty calibrated to a target model class.
• multimodal/image-classification — custom-category image classification with single_label / multi_label modes.
• multimodal/handwriting-transcriber — handwriting transcription with per-word confidence.
• cot/citation-grounded-reasoning — reasoning where every factual claim cites a provided source (or is marked inference / common knowledge).
• cot/contrastive-self-consistency — articulate a plausible wrong path and contrast with correct path; anti-misconception.
• eval/calibration-checker — bucket predictions by confidence and measure ECE; identify over-confident / under-confident regions.
• rag/structured-rag-output-builder — produce structured outputs (table / list / record) from RAG sources with per-cell citations.

Per-direction: rag 11, eval 11, agent 11, sft 10, multimodal 10, cot 10, rlhf 10, code 10. Total 83. All directions ≥10.

Batch 3 to 100 (17 final cards)

• rag/query-fusion — fuse N sub-query retrieval results into one ranked deduplicated set.
• rag/time-aware-retrieval-rewriter — resolve "last month" / "recent" into concrete time bounds for the retriever.
• agent/multi-agent-conflict-resolver — reconcile conflicting outputs from multiple sub-agents (factual / value-based / methodological / scope).
• agent/api-result-translator — turn raw API JSON into a user-readable answer in a chosen style.
• rlhf/reward-hacking-detector — detect sycophancy / verbosity / format gaming / qualifier flooding in post-RLHF responses.
• rlhf/preference-rationalization-judge — audit whether a preference label's rationale actually justifies the pick.
• sft/code-sft-pair-generator — generate code SFT pairs at controlled difficulty in a target language with test sketch.
• sft/instruction-deduplicator — find semantic near-duplicates (paraphrases) in an instruction set.
• multimodal/image-edit-instruction-generator — reverse-engineer the edit instruction from a before/after image pair.
• multimodal/image-comparison-explainer — structured similarity / difference explanation between two images.
• cot/self-correction-protocol — process external criticism with accept / correct / reject decision.
• cot/meta-prompt-generator — generate a reusable meta-prompt for a class of tasks from description + examples.
• eval/human-eval-bootstrap — design a small human eval study (rubric + annotator instructions + calibration + analysis).
• eval/leaderboard-builder — multi-benchmark leaderboard with weighted aggregation, differentiator detection, and saturation flagging.
• code/error-message-explainer — explain stack traces / errors calibrated to junior_dev / senior_dev / non_technical audiences.
• code/commit-message-generator — generate commit messages from git diff in conventional / imperative / verbose styles.
• code/api-design-reviewer — review API design (REST / GraphQL / gRPC) on consistency / ergonomics / evolvability / security / performance.

Per-direction: rag 13, eval 13, agent 13, code 13, sft 12, multimodal 12, cot 12, rlhf 12. Total 100. All directions ≥12.

[0.2.0] — 2026-05-10

100-card milestone release. The post-v0.1.0 polish + 3 batches of expansion (15 + 15 + 17) brought the library from 32 cards in v0.1.0 to 100 cards across 8 directions, with all "Unreleased" content above moved to this release line.

Bilingual at-a-glance (场景 blockquote)

• Every card now opens with a single-line Chinese > 🎯 **场景**：... blockquote summarizing the use case. Aimed at non-English readers who need to identify "is this card for me?" in one glance. Retrofitted to all 47 cards, added to the template, documented in SCHEMA / CLAUDE / CONTRIBUTING.

Repository organization

• Root directory tidied: CONTRIBUTING.md and CODE_OF_CONDUCT.md moved to .github/ (GitHub auto-detects from there); CHANGELOG.md and ROADMAP.md moved to docs/. Root went from 11 markdown files to 5.

Changed

• SKILL.md routing tree expanded from 28 to 38 entries.
• templates/prompt-card.md updated with ## Quick Use template.
• CLAUDE.md and .github/CONTRIBUTING.md updated to describe the Quick Use quality bar.
• scripts/build_index.py "Use when" extraction now reads the **Use when:** line from ## Quick Use (cleaner than parsing Purpose).

[0.1.0] — 2026-05-10

First public release. The initial schema, tooling, and a curated set of 32 Prompt Cards covering all seven directions.

Added

Repository structure and tooling

• Hybrid form: GitHub repository and Claude Code skill (one SKILL.md at the repo root makes the same content installable to ~/.claude/skills/prompt-atlas).
• Dual-license policy: LICENSE documents MIT for code (scripts/, CI) and CC-BY-4.0 for prompt content (prompts/, templates/, docs/).
• Schema definition in docs/SCHEMA.md: required frontmatter fields, required body sections, controlled vocabulary, and the placeholder-↔-variables consistency rule.
• Canonical controlled vocabulary in scripts/vocab.yml. validate.py reads it directly so vocabulary changes only require touching one source plus its prose mirror in docs/SCHEMA.md.
• scripts/validate.py: enforces frontmatter shape, ID/path/direction consistency, controlled-vocabulary membership, body section ordering, and {{placeholder}} ↔ variable declaration consistency.
• scripts/build_index.py: regenerates INDEX.md (catalog by direction + tag matrix) from card frontmatter.
• CI workflow at .github/workflows/validate.yml: runs validation on every push and PR; fails the build if INDEX.md is stale.
• Standard OSS hygiene files: README.md (English/Chinese bilingual), CONTRIBUTING.md, CODE_OF_CONDUCT.md, .gitignore, .editorconfig, .github/PULL_REQUEST_TEMPLATE.md, structured issue templates for new prompt cards and bug reports.
• Safety policy in docs/SAFETY.md: explicit prohibition on jailbreaks, hidden-CoT extraction, harm-enabling content, and proprietary leaks; explicit endorsement of defensive and evaluation prompts.

Prompt Cards (32)

RAG (6):

• retrieval-relevance-evaluator — score passage relevance to a query.
• multihop-eval-synthesizer — synthesize multi-hop QA eval questions.
• query-rewriting-decomposition — split a query into focused sub-queries.
• hyde-hypothetical-answer-generator — HyDE technique for retrieval.
• citation-faithfulness-scorer — audit per-claim citations.
• answer-grounding-checker — per-answer hallucination detection.

Agent (5):

• react-planner-with-tool-schema — ReAct loop with strict JSON tool calls.
• plan-and-execute-planner — upfront linear plan for predictable goals.
• tool-call-repair — schema-driven repair of malformed tool calls.
• self-critique-reflection — meta-level continue/switch/escalate.
• long-context-memory-summarizer — compress trajectories into structured memory.

RLHF (4):

• pairwise-preference-labeler — HHH dimensions, pairwise.
• pointwise-reward-scorer — single-response scalar reward.
• constitutional-critique-revise — Constitutional AI critique-then-revise.
• best-of-n-selector — rank N candidates, supports both inference and data construction.

SFT (4):

• instruction-variant-expander — rewrite ONE seed into N variants.
• self-instruct-from-seed — generate NEW instructions in the same task family.
• data-quality-filter — keep / review / drop SFT pairs by quality.
• response-generator — produce the response half of an SFT pair.

Multimodal (4):

• vlm-image-description-verifier — audit caption against image (per-claim).
• structured-caption-generator — captions as discrete schema fields.
• vqa-with-confidence — visual QA with grounding region and confidence.
• ocr-structured-extraction — schema-driven OCR + extraction for documents.

CoT (4):

• structured-reasoning-with-rationale-summary — single-path reasoning with rationale_summary.
• least-to-most-decomposition — decompose into strictly easier sub-problems.
• self-consistency-aggregator — majority vote across N pre-sampled paths.
• verify-then-finalize — draft + verify + finalize in one prompt.

Eval (5):

• llm-judge-rubric-open-ended — fixed 4-dimension rubric for open-ended outputs.
• reference-based-judge — score against a gold reference.
• per-claim-factuality-judge — atomic claim decomposition + factuality labels.
• pointwise-quality-scorer — custom dimensions + self-reported confidence.
• safety-output-classifier — defensive harm-taxonomy classifier.

Design decisions documented

• Cards are work assets, not snippets. Every card includes ## Failure Modes (how it breaks, how to detect) and ## Tuning Notes (model differences, temperature, sibling cards).
• Tuning Notes explicitly compare each card to its siblings so readers know when to switch (e.g. open-ended vs reference-based judges, pairwise vs pointwise reward, ReAct vs plan-and-execute).
• "Hidden chain-of-thought extraction" is a banned use case; cards use reasoning_summary / rationale_summary / decision_basis field names by convention. Normal step-by-step reasoning in visible output is fine.
• INDEX.md is auto-generated, never hand-edited. CI rejects PRs that ship a stale INDEX.md.

Removed

The following tags were proposed in early drafts but never used by any card and were removed from the controlled vocabulary in this release: audio, persona, style-control, video. They can be re-added in the future via the standard vocabulary change process when an actual card needs them.