How PromptOptBase rewrites and assembles prompts

PromptOptBase is a local structuring toolkit. We never send your drafts to OpenAI, Anthropic, Google, or any other model provider for generation, scoring, ranking, or storage. When you click Optimize or Generate, the output is produced by deterministic rules in JavaScript — not by a remote inference call.

That means we cannot "know" your product, audience, or compliance context unless you type it. We reorganize, label, trim, and scaffold what you provide. Outputs are labeled as structured rewrite or structured build so you can tell at a glance that nothing was inferred by a cloud model.

• No API keys required and none are called
• No server-side prompt storage or analytics on your text
• No quality score from a hidden model — only checklist-style changes you can read

The homepage optimizer takes an existing draft and walks a fixed, ordered checklist. Each step is visible in the before/after redline view so you can reject or edit anything that moved.

1. Name the task up front — one clear ask, not three competing goals
2. Separate audience, tone, and context when the draft mixes them in one paragraph
3. Pin output format: length, sections, list vs prose, language, or file shape
4. Cut filler phrases, hedging loops, and duplicate sentences that repeat the same constraint
5. Optionally wrap surviving content in a light scaffold: Task → Context → Constraints → Output
6. Show redline diffs so additions, moves, and deletions are explicit
7. Never fabricate product facts, metrics, customer names, or brand claims

If your draft is missing a labeled field — for example, no output format section — we may insert an empty labeled slot with placeholder guidance. That slot is a reminder for you to fill in, not AI-invented copy.

The prompt generator maps form fields into a consistent prompt skeleton. Required fields block generation with inline errors; empty optional fields are omitted entirely so the result stays tight.

The goal field is required. Without a stated objective, the assembler cannot know what to optimize for and will refuse to build output.

• Goal / task — what success looks like in one sentence
• Audience and tone — who reads this and how it should sound
• Context — background facts you already know and want preserved
• Constraints — must-follow rules, banned topics, length caps
• Output format — structure, bullets, JSON shape, or example layout

Generated text is assembly, not invention. If you leave context blank, we do not guess your market, SKU list, or campaign history.

The marketing prompt builder targets Ads, Email, Social, and Landing channels. Each channel loads static constraint blocks from templates maintained in code — not from trend scraping, competitor monitoring, or model calls.

• Ads — character discipline, CTA placement, variant count, compliance reminders
• Email — subject/preheader hooks, body scannability, single primary CTA
• Social — platform length habits, hook-first openings, hashtag/emoji policy toggles
• Landing — hero promise, proof points, objection handling, section order

You supply product facts, offers, and brand voice. The builder stitches channel-appropriate section labels and constraint reminders around your inputs. It does not research your category or write claims you did not provide.

The framework builder applies prompting techniques as explicit labeled blocks: chain-of-thought steps, few-shot example pairs, role framing, output parsers, and similar patterns.

• Chain-of-thought — numbered reasoning steps with a final answer slot
• Few-shot — example input/output pairs you provide; we do not invent demonstrations
• Step-back / decomposition — parent question plus sub-question list
• Structured output — schema or field list you define; empty fields stay empty
• Self-check — review checklist appended before the model is asked to respond

Technique blocks change layout and instructional phrasing only. Selecting CoT does not call a reasoning model; it inserts the scaffolding text you would have pasted manually.

The role and system prompt builder assembles persona, scope, safety boundaries, tool-use hints, and response habits into a system-style prompt. Persona text comes from your fields; we do not generate a fictional expert biography.

1. Persona — role name, domain, and voice cues you provide
2. Scope — what the assistant should and should not do
3. Rules — numbered behaviors, citation habits, refusal boundaries
4. Output habits — default format, language, brevity, or citation style
5. Optional user-facing preamble — how the assistant should greet or clarify

Treat system prompts as policy documents you own. Review scope and refusal rules for your org before deployment.

The length tuner reuses optimizer logic with length as the primary control. Two modes share the same honesty rules as the homepage tool.

• Expand — adds empty labeled fields (examples, output format, edge cases) when missing; never inserts brand claims or statistics
• Compress — removes filler, repeated constraints, and redundant examples while keeping the core ask
• Protect constraints — when enabled, constraint blocks are not deleted during compress passes

Expand makes a short prompt easier to complete; it does not make a vague prompt correct. Compress makes a long prompt paste-ready; it does not guarantee policy compliance.

ChatGPT, Claude, Gemini, and Generic presets adjust section labels and phrasing habits — for example, how system instructions are introduced or how constraint blocks are titled. They do not call those vendors' APIs, fetch live tokenizer tables, or simulate model responses.

• ChatGPT — common system/user split labels and concise directive tone
• Claude — XML-style section tags where helpful and explicit context framing
• Gemini — role/parts-friendly headings and safety reminder placement
• Generic — vendor-neutral headings when you paste into custom gateways or local runners

Presets are copy ergonomics, not routing. Your chat product still decides how instructions are interpreted.

The templates library ships copy-ready starters maintained as static content in the repository. Each template is a vetted scaffold you can paste as-is or run through the optimizer after swapping in real facts.

• Not a personalized feed — the same catalog for every visitor
• Not an AI-generated example bank — no nightly model refresh
• Not auto-filled with your company data — placeholders stay explicit

Templates accelerate structure; they do not replace research, legal review, or brand approval.

Some flows let you send output from the optimizer into the generator, or the reverse, without leaving the site. That handoff uses optional browser-only temporary storage for the current tab session — nothing is uploaded.

• Handoff is opt-in and can be cleared by closing the tab or starting fresh
• No account sync — another device or browser profile will not see your handoff buffer
• Sensitive drafts should be cleared after use on shared machines

For the full rule list in one place, see this methodology page. For tool-specific walkthroughs, use the guides linked from each workbench.

PromptOptBase is honest about what local rules cannot do. We optimize structure and clarity; we do not certify outcomes.

• No quality scoring — we do not rank prompts with a hidden model or claim a numeric "prompt score"
• No policy guarantee — ads, email, medical, financial, or HR compliance remains your responsibility
• No live web research — we cannot fetch competitor pages, pricing, or news
• No multilingual nuance model — locale UI translations do not auto-translate your prompt body
• No memory across visits unless you explicitly use same-tab handoff

If a tool label implies more than rearrangement and labeling, treat that as a bug and report it.

Structured output is still your content. Before pasting into ChatGPT, Claude, Gemini, or an internal gateway, you should verify facts, tone, and policy fit.

1. Read the redline or assembled sections — especially empty labeled fields you must complete
2. Replace bracketed placeholders with accurate product, legal, and customer details
3. Check channel rules: ad disclosures, email consent, social character limits, landing claims
4. Test with your actual model and temperature settings — our presets do not predict behavior

Report a misleading label, a rule that drops constraints, or a page that overclaims capabilities at contact.