-
Notifications
You must be signed in to change notification settings - Fork 0
FAQ
Quick answers about Wild_Root_Prompt — what it is, what it costs, whether your data leaves your machine, what hardware it needs, and how it compares to writing prompts by hand.
A local prompt engineering tool that rewrites a rough idea into an expert-level prompt or a full 12-section instruction manifest, by applying 173 prompt engineering techniques before your model ever sees the input.
No. Generation runs against a local backend over the loopback interface. The only outbound traffic is the optional web enrichment — a search on your task text — and --offline disables even that. There is no telemetry, no account, and no API key in the codebase.
No. It runs on Ollama or any OpenAI-compatible local server. There is nothing to pay for and nothing to sign up for.
Yes — MIT licensed, for personal and commercial use.
A 3B model needs roughly 4 GB of RAM and works fine, especially in Quick mode. 7–8B models want ~8 GB and are the sweet spot for Full manifests. It runs on Android via Termux too. See Installation.
No. Double-click the launcher and use the Web UI in your browser. The CLI is there for people who want technique bundles, scripting, and parallel dual-model runs.
Quick returns one enhanced prompt to paste into any LLM. Full returns a 12-section instruction manifest meant to be executed by an agent. Iterate in Quick, deliver in Full. --draft gives you sections 1–2 of a Full manifest as a cheap direction check.
Prompt engineering methods — Chain-of-Thought, Tree-of-Thought, ReAct, MECE, Constitutional AI, red teaming, and 160-plus more — across 15 categories, stored as editable JSON. The complete catalogue is at Prompt Engineering Techniques.
Start with the defaults (15, balanced). For a specific job use a bundle: --techniques "bundle:Audit / securite". If you're unsure, --recommend-techniques picks based on your task's own wording.
No. Past roughly 30 active techniques, small models start producing a mushy compromise between competing instructions. A coherent bundle of 10–20 beats a 100-technique pile.
Inline modifiers typed in your task text — /expert /tableau /sources — that control persona, format, depth, reasoning style and epistemic standards. There are 60. See Slash Metacommands.
Because their disagreements are informative. Pair a systematic model with a creative one and the synthesis pass has genuinely different material to merge, rather than two paraphrases of the same answer.
Yes — --backend openai_compatible with the server's URL. The backend must run on the same machine (loopback only, by design). See Configuration.
Completely, with --offline. That skips web enrichment and connectivity checks entirely. Everything else — pre-processing, technique injection, generation, synthesis — is local anyway.
Yes. Both are plain JSON in the repo root: prompt_expert_methodology.json and prompt_templates.json. Edit them and the changes apply on the next run — no code change, no rebuild. Contributing if you want to share them.
Handing a complete, unambiguous brief to an LLM agent — including ambiguity zones, error handling, control loops and a reproducibility checklist. It's written to be executed, not just read. Structure: Architecture.
Yes. The manifests are model-agnostic — that's the point. Generate locally, execute anywhere.
Yes, in memory/, injected as context so a sequence of related tasks stays coherent. Disable per-run with --no-memory, per-prompt with /neuf, or wipe it with memory clear. It can be encrypted at rest.
Yes — a small REST API with SSE streaming, or the CLI with --quiet --output. See REST API.
No. It binds to loopback and has no authentication because it has no remote attack surface by design. If you need remote access, tunnel over SSH rather than exposing the port.
It isn't notarized — that requires a paid Apple developer account. The source is public and you can build it yourself. Workaround in Troubleshooting.
The methodology file was authored bilingually and the identifiers were kept stable rather than renamed, since they're referenced by ID and by name in saved configurations. Bundle names are matched case-insensitively, and the numeric form (bundle:6) works everywhere the name does.
The issue tracker. Include the command you ran and the full error.
Next: Examples — real before/after runs · Comparison — versus other approaches · Glossary — every term defined · Use Cases · Quick Start.
Start
Reference
Understand
Help