--- name: isatvon-prompting description: Converts any raw prompt into an ISATVON-structured prompt for any AI platform (ChatGPT, Claude, Gemini, Perplexity, Copilot, Grok, ...), with the response also returned in ISATVON structure. Use when the user asks to improve, structure, convert, or rewrite a prompt, asks for an "ISATVON prompt", or wants a COSTAR-style prompting framework. license: Apache-2.0 --- # ISATVON Prompting You rewrite a raw, unstructured prompt into an **ISATVON-structured prompt**: seven markdown sections (I-S-A-T-V-O-N) that work on any AI platform and instruct the model to return its **answer in ISATVON structure too** — so both the ask and the response are auditable. Think COSTAR, plus the two mechanisms it lacks: the **A section's self-verification** (the model checks word counts and constraints before answering) and a hard **anti-fabrication rule** (never invent figures or citations). In blind benchmarking these were exactly the failure classes that separated the two — CO-STAR outputs invented statistics; ISATVON outputs did not. Element guide: [references/prompting-guide.md](references/prompting-guide.md) · COSTAR mapping: [references/costar-comparison.md](references/costar-comparison.md) ## When to use - The user asks to "improve / structure / rewrite / convert my prompt" or for an "ISATVON prompt". - The user wants a reusable, high-quality prompt for ChatGPT, Claude, Gemini, Perplexity, Copilot, Grok, or any other AI platform. - The user mentions COSTAR or asks for a prompting framework. Where it measurably helps most: **writing/content and constraint-heavy tasks** (blind benchmark: 4.70/5 vs 3.40 baseline). For pure reasoning or coding one-liners, baseline models already score ~4.8–5.0 — use **Lite mode** (below) or skip conversion. ## How NOT to use Send the converted prompt as the **user message**. Do not condense ISATVON's rules into a system prompt or "custom instructions": in benchmarking, that mechanism was ignored in 8/8 runs (replicated in a clean session and on a stronger model) and scored *below* baseline. The framework only works as a full prompt rewrite. Do NOT use for generating ISATVON **skill manifests** (JSON contracts for agent loops) — that is the [isatvon](https://github.com/isatvon/isatvon) skill. This skill produces **text prompts**, not manifests. ## Procedure ### 1. Extract intent Read the raw prompt. Identify the task, the implicit audience, and the desired deliverable. Ask at most 1–2 questions, and only if the goal is genuinely ambiguous — otherwise state your assumptions in the converted prompt's N section and proceed. Then triage the mode: - **Lite (I + O + N)** — the task is a self-contained reasoning, coding, or one-shot question with no constraints to enforce and no external sources to ground. Use [templates/prompt-template-lite.md](templates/prompt-template-lite.md). - **Full (all 7 sections)** — everything else: writing/content, specs, multi-constraint, or source-grounded tasks. This is the default. ### 2. Fill the seven sections Start from [templates/prompt-template.md](templates/prompt-template.md): | Section | What goes in it | |---|---| | **I — Instructions** | Role for the model, the task in one imperative sentence, and the rules that bound it. | | **S — Source** | Context, data, and references the model must use — and what it must not assume. | | **A — Automation** | The step-by-step method, ending with a **self-verification step** the model runs before answering (e.g. "check every claim traces to S"). | | **T — Tech stack** | Capabilities/tools to use or avoid: web search, code execution, no fabricated citations. | | **V — Variables** | Constraints: length, tone, audience, format limits — plus a fallback ("if a constraint can't be met, say which and why; ask at most 1 clarifying question"). | | **O — Outcome** | The exact response format. Always embed the ISATVON response structure from [references/response-format.md](references/response-format.md). | | **N — Notification** | Meta-reporting the model must include: assumptions made, confidence, what was skipped. | Every section must be filled — if the raw prompt gives nothing for a section, supply a sensible default rather than dropping it. **Exception (precedence rule):** if the *deliverable itself* is unknowable from the raw prompt (e.g. "Help me improve this." with nothing attached), do not invent one. The converted prompt's I section instructs the model to ask 1 clarifying question first; V's fallback governs. "Fill every section" applies to *supporting* sections (S, T, V defaults), never to fabricating the task. ### 3. Require an ISATVON response The O section always instructs the model to structure its reply as I (task as understood) → S (sources used) → A (how it verified) → V (constraints honored or broken) → O (the deliverable itself) → N (assumptions & confidence). Copy the skeleton from [references/response-format.md](references/response-format.md) verbatim into O. Lite prompts embed the compact 3-line skeleton (I / O / N) from the same file instead. ### 4. Tailor to the platform (if named) - **Perplexity / search-first tools** — S names the search scope and recency window; T requires cited sources. - **Copilot / coding assistants** — S names the workspace files; T pins languages/frameworks; A ends with "run or type-check the result". - **ChatGPT / Claude / Gemini / Grok** — general purpose; tune T to available features (browsing, code interpreter) and V to tone. - No platform named → keep it platform-agnostic (no tool assumptions in T beyond "don't fabricate"). ### 5. Emit Output the converted prompt as **one copy-pasteable markdown block**, then 1–2 sentences on what was added or tightened versus the raw prompt. Do not explain the framework itself unless asked. ## Worked examples Before/after conversions showing the expected concreteness: - [examples/generic-summarize.md](examples/generic-summarize.md) — platform-agnostic baseline - [examples/chatgpt-marketing-email.md](examples/chatgpt-marketing-email.md) - [examples/claude-code-review.md](examples/claude-code-review.md) - [examples/gemini-research-brief.md](examples/gemini-research-brief.md) - [examples/perplexity-market-research.md](examples/perplexity-market-research.md) - [examples/copilot-refactor.md](examples/copilot-refactor.md) - [examples/grok-social-post.md](examples/grok-social-post.md)