Roadmap

A living document. Items here are intentions, not commitments — open issues and PRs are the authoritative state of work in flight.

v0.1 (released 2026-05-10)

Shipped. See CHANGELOG.md. 32 cards across 7 directions, schema + validator + CI, bilingual README, safety policy, dual license.

v0.2 — broaden coverage and tighten boundaries

Goal: get every direction past the "minimum representative" line and into "covers the common production tasks". Roughly +10–15 cards across the seven directions.

Likely additions per direction:

• RAG: chunk summarization for retrieval, relevance feedback / iterative retrieval, context compression.
• Agent: multi-agent supervisor, sub-task delegation, tool catalog generator from API specs.
• RLHF: red-team prompt generator (defensive), iterative DPO data generator, refusal calibration probe.
• SFT: persona controller, style-controlled response variants, conversation-format SFT pair generator.
• Multimodal: chart / table extraction, image + text joint classifier, video frame description aggregator (audio direction reserved for a later release).
• CoT: tree-of-thoughts variant, plan critique + rewrite, step-by-step rationale with citations.
• Eval: pairwise judge with position-bias probe, calibration test harness, multi-turn dialogue judge.

Vocabulary additions in v0.2 will follow the standard process: each new tag/audience/model entry must be justified in the PR that introduces the first card using it.

v0.3 — discovery and integration

Goal: make the library easier to find the right card in.

• README catalog table generated from INDEX.md so the front page doubles as the category index.
• scripts/build_index.py extension: secondary index slices by audience and by model class.
• Skill-side improvements: tag-search hints in SKILL.md so Claude Code can route on intent phrases that don't directly map to a direction.
• A short "tour" doc that walks through 3–4 representative end-to-end workflows (a RAG eval pipeline, an SFT data construction pipeline, an agent loop) using the cards in sequence.

Beyond v0.3 — open questions

These are not commitments; they are open design questions where the right answer depends on contributor traffic and use signals.

• Versioned card retrieval. As cards evolve (version: 0.2.0 etc.), is there value in keeping previous revisions accessible rather than overwriting? Likely yes for cards that change semantically; not worth it for cosmetic edits.
• Sibling-card cross-references as machine-readable links. Today the "see also" pointers live in prose inside Tuning Notes. A structured related_cards frontmatter field would let tooling surface cluster diagrams, but adds maintenance cost.
• Per-card eval harness. Some cards have a known "if calibrated, agreement with gold should be >X%". Stamping cards with a reproducible eval would make quality auditable, but it's a real engineering project.
• Localization beyond bilingual README. Card prompt bodies are English-first by policy (most production prompts use English to maximize model performance). If demand emerges for native-language prompt variants, that would need a per-card language: zh variant alongside the English original — schema-supported but currently unused.
• Plugin / package distribution. Today, installation is git clone. A pip install prompt-atlas or similar is doable but only worth it if the cards are consumed programmatically, not just read.

Non-goals

Intentionally out of scope for the foreseeable future:

• A SaaS layer or hosted UI. The repo + skill model is the product.
• Generation prompts whose primary purpose is harm enabling, jailbreak research, or proprietary-prompt redistribution. See docs/SAFETY.md.
• Becoming a generic "prompt store" by including every prompt anyone contributes. The PromptOps quality bar (Failure Modes, Tuning Notes, sibling-card relationships) is the entry criterion.

How to influence the roadmap

• Open an issue with the new-prompt-card template proposing a card you'd use.
• Open an issue describing an integration / discovery / quality gap with concrete examples.
• A PR with the card or fix is the strongest form of influence.