Prompt Failover Kit — Model-Agnostic Prompts With a Fallback Playbook for Claude, GPT, Gemini & Local LLMs

Prompt Failover Kit is a library of eight original, model-agnostic prompts — each with per-model tuning notes and a shared fallback playbook — built for people who rely on a single chat LLM for real work and have no plan for when it refuses, degrades, times out, or hits a usage cap. No gateway to deploy, no API key, no self-hosting required.

Why this exists

In September 2026, one of the fastest-rising repositories on GitHub is OmniRoute, a free, MIT-licensed AI gateway with roughly 64,300 stars. It puts a single endpoint in front of 352 AI providers — Claude, GPT, Gemini, GLM, DeepSeek, MiniMax and others — and automatically cascades across four tiers (Subscription → API Key → Cheap → Free) when a provider fails or runs out of quota, using 19 routing strategies and "last-known-good-provider" stickiness. It stacks compression engines that cut token usage 15–95% (about 89% on average) to stretch free-tier budgets further, and it runs three self-healing layers — circuit breakers, exponential backoff, and per-model lockouts — so a single bad provider never takes the whole system down. It's a genuinely well-built piece of infrastructure, and it's free to self-host.

But it's infrastructure. Using it means running a gateway service, wiring up provider credentials, and pointing your own code at it — not something you do from inside a ChatGPT, Claude, or Gemini tab. Most people doing real work with an LLM aren't calling an API at all; they're typing into one chat window with no plan for what happens when that one model refuses the request, quietly degrades, times out, or they simply run out of messages for the day. Prompt Failover Kit takes the same underlying discipline — not any of OmniRoute's code, config, or routing logic — and rebuilds it at the prompt layer: every entry works as a portable master prompt across models, comes with the tuning it needs on each major one, and sits inside a fallback playbook that tells you exactly when to switch and what to switch to, by hand, in under a minute.

What's inside

• 8 Universal Master Prompts for high-value work tasks, each portable across Claude, GPT, Gemini, and local/open-weight models
• Per-model Tuning Notes for every prompt — the specific phrasing fix for each model's known failure pattern
• A Portability Compiler — turn any single-model prompt you already rely on into a version that works everywhere
• A Compression Pass — shrink any prompt-and-output pair into a budget version for when you're close to a usage cap
• A Fallback Playbook — the decision tree for what to do, and which model to switch to, when your primary one refuses, degrades, times out, or runs dry

Best for

• Anyone doing real work through a single chat subscription with no backup plan for outages, rate limits, or a model update that broke a prompt that used to work
• People who've had a model suddenly refuse, ramble, or go generic mid-project and didn't know whether to fight it or switch
• Teams whose "AI process" is really just "whoever's logged into which tool," with no shared way to move a task between models without starting over

What you'll need

• Access to at least one major chat LLM — Claude, ChatGPT, or Gemini — no plugins, agents, or API access required
• Ideally a free-tier account on a second model, so the fallback playbook has somewhere to send you
• The raw facts for whatever you're producing — each master prompt asks for these upfront

How tuning notes work (read this first)

Every prompt below is written in plain, portable language on purpose — no model-specific syntax. Paste the Universal Master Prompt as-is into any chat model. Then check the Tuning Notes for the model you're using and add the one line it suggests before you send, or append it as a follow-up if the first response already shows the failure pattern it describes.

The 8 prompts (copy/paste)

1. Weekly Status Report Generator

Universal Master Prompt: "Write a weekly status report. Inputs: project [NAME], period [DATES], completed this week [LIST], in progress [LIST], blockers [LIST IF ANY], next week's focus [LIST]. Structure: Summary (2-3 sentences), Completed, In Progress, Blockers (with what's needed to unblock each), Next Week. Under 400 words, no filler transitions between sections."

Tuning notes: Claude — often adds a closing offer to help further; add "End after Next Week, no closing offer." GPT — tends to merge Blockers into In Progress; add "Blockers must be its own section even if empty." Gemini — sometimes drops the bracketed section headers; add "Use these five headers exactly, in this order." Local/open-weight — drop the word-count constraint; smaller models trade off accuracy against arbitrary limits.

2. Meeting Transcript → Action Items

Universal Master Prompt: "Here is a meeting transcript: [PASTE TRANSCRIPT]. Extract every action item discussed, whether or not it was explicitly assigned. Output a table: Action, Owner (write 'Unassigned' if unclear from the transcript), Due Date (write 'Not set' if unclear), Context (one line quoting or paraphrasing where it came from)."

Tuning notes: Claude — reliable at full extraction on long transcripts; no adjustment usually needed. GPT — on transcripts over ~3,000 words it can stop early; add "Continue until the transcript ends — confirm you reached the final line." Gemini — sometimes infers an owner that wasn't stated; add "Only assign an owner if a name is explicitly tied to the action in the text." Local/open-weight — split long transcripts into two or three chunks and run the prompt on each rather than pasting the whole thing at once.

3. Cold Outreach Email

Universal Master Prompt: "Write a cold outreach email. Inputs: recipient's role and company [DETAILS], why I'm reaching out [REASON], the specific value I can offer them [VALUE], one piece of genuine research about their situation [RESEARCH NOTE], the ask [SPECIFIC ASK]. Under 120 words, one clear call to action, no generic flattery."

Tuning notes: Claude — well-calibrated on tone by default. GPT — can default to a generic "I hope this finds you well" opener; add "Do not use a generic greeting line — open directly with the research note." Gemini — sometimes produces a shorter, blunter draft than the word count allows; that's usually fine here, but ask it to "use the full word budget if it improves specificity" if the draft feels thin. Local/open-weight — provide the research note as a complete sentence rather than a fragment; weaker models paraphrase fragments poorly.

4. Competitive Landscape Brief

Universal Master Prompt: "Write a competitive landscape brief. Inputs: our product [DESCRIPTION], competitors to cover [LIST 3-5], what I know about each one's positioning and pricing [NOTES PER COMPETITOR]. Structure: one paragraph per competitor covering their apparent strength, apparent weakness, and how we differ, then a closing 'Where we should focus' paragraph. Base every claim only on the notes provided — flag anything you're inferring rather than stating as fact."

Tuning notes: Claude — good at flagging inference versus fact when asked; keep that instruction in, don't drop it to save length. GPT — may fill gaps with plausible-sounding but unverified competitor details; add "If my notes don't cover something, say 'not provided' rather than guessing." Gemini — tends toward shorter per-competitor paragraphs; ask explicitly for "3-4 sentences per competitor, not one." Local/open-weight — feed one competitor at a time across separate turns rather than all 3-5 in one prompt.

5. Contract Clause Risk Scan

Universal Master Prompt: "Here is a contract clause: [PASTE CLAUSE]. This is not a request for legal advice — flag language patterns only. List: (1) any term that's undefined but load-bearing to the clause's meaning, (2) any obligation that's one-sided or asymmetric between the parties, (3) any deadline, cap, or threshold stated ambiguously (e.g. 'reasonable time' with no number). For each, state the specific phrase and why it's ambiguous or asymmetric. Do not rewrite the clause."

Tuning notes: Claude — sometimes adds a disclaimer paragraph before answering; add "Skip the disclaimer, I've already stated this isn't legal advice — go straight to the list." GPT — generally direct and thorough on this task. Gemini — may occasionally rewrite the clause anyway despite the instruction; if it does, resend with "List only — if you rewrite anything, redo the response." Local/open-weight — this task needs careful reading; if the clause is dense legal prose, prefer a larger hosted model over a small local one for this specific prompt.

6. Data Table Insight Summary

Universal Master Prompt: "Here is a data table: [PASTE TABLE OR CSV]. Identify the 3 most notable patterns — largest changes, clearest trends, or biggest outliers. For each, state the specific numbers involved, not just a description. Do not recommend actions, just report what the data shows. If the table has fewer than 3 clear patterns, report only the ones that are actually there."

Tuning notes: Claude — accurate on cited numbers; still spot-check one row yourself, since any model can transpose digits. GPT — good at pattern-finding but occasionally rounds numbers without saying so; add "State exact values from the table, not rounded ones." Gemini — may pad to exactly 3 patterns even when only 1-2 are genuinely notable; the "only the ones that are actually there" line matters most on this model. Local/open-weight — cap the table at roughly 30-40 rows per prompt; larger tables risk silent truncation.

7. Long Document Executive Summary

Universal Master Prompt: "Here is a full document: [PASTE FULL TEXT]. Write a one-page executive summary: the single most important finding first, then 3-5 supporting points each with the number or evidence behind it, then the recommended action if one is stated in the document. No claims beyond what's in the text."

Tuning notes: Claude — handles long documents in one pass well; good default choice for anything over 5,000 words. GPT — on very long documents, verify the summary covers material from the end of the document, not just the opening — it can weight early content more heavily. Gemini — strong at long-context recall; if the summary feels thin, ask it to "cite the section each point came from." Local/open-weight — most local models have a much smaller context window; paste in sections and summarize each before producing a final combined summary.

8. Customer Support Reply Drafter

Universal Master Prompt: "Draft a customer support reply. Inputs: customer's message [PASTE], the actual resolution or answer [FACTS], our tone [e.g. warm and direct / formal / casual]. The reply must directly answer what the customer asked first, before any context or apology. Under 150 words."

Tuning notes: Claude — can default to opening with an apology even when nothing went wrong; add "Only apologize if something on our end actually went wrong." GPT — reliable at matching the specified tone. Gemini — sometimes answers slightly indirectly before the direct answer; the "must directly answer first" line is the fix, keep it in. Local/open-weight — give the tone as a short example sentence rather than an adjective; smaller models match examples better than abstract tone labels.

Supporting framework: Portability Compiler

Use this to turn any prompt you already rely on — tuned for one specific model — into a version that works across all of them.

"I currently use this prompt on [MODEL NAME]: [PASTE YOUR EXISTING PROMPT]. Rewrite it to be model-agnostic: remove any phrasing, formatting syntax, or assumptions specific to that one model, state every output requirement explicitly rather than relying on that model's defaults, and keep every bracketed input placeholder intact. Output the portable version only."

Supporting framework: Compression Pass

Use this when you're close to a message cap or daily limit and need the same result in fewer tokens.

"Here is a prompt and a typical response length it produces: Prompt: [PASTE PROMPT]. Typical response: [PASTE OR DESCRIBE A SAMPLE OUTPUT]. Rewrite the prompt to produce the same essential output in roughly half the length — cut supporting explanation and formatting overhead, keep every fact, number, and decision the original output contained. Output the compressed prompt only."

Fallback Playbook

Run through these in order whenever a model isn't giving you a usable result. This is the manual version of what an automated gateway does for API traffic — you're the router.

1. Refusal or heavy hedging: Don't argue with the refusal. Reframe the request as analysis rather than action (e.g. "identify the risk patterns in this clause" rather than "is this clause enforceable"), or move to a different model — some models are measurably more willing on borderline-but-legitimate business tasks than others.
2. Quiet degradation (output goes generic, vague, or stops following your structure): Don't keep re-rolling the same prompt. Switch to that prompt's Tuning Notes above, or apply the Compression Pass — degradation often means the model lost track of a long instruction, and a shorter, tighter prompt recovers it.
3. Timeout or outage: Have your second model's tab already open before this happens, not after. Paste the same Universal Master Prompt — it was written to be portable for exactly this moment — apply that model's Tuning Notes, and continue.
4. Rate limit or usage cap reached: Apply the Compression Pass to your remaining tasks first. If you're still capped, that's the moment to fail over to model two — don't wait until you're fully blocked to open the backup tab.
5. Three failed attempts on the same model: Stop. Three re-rolls of a prompt that isn't working is a signal to switch models, not to try a fourth phrasing. This is the prompt-layer version of a circuit breaker — it exists to stop you from burning a budget on a path that's already shown it's not working right now.

How to use

1. Pick the prompt matching the task you need to do.
2. Paste the Universal Master Prompt into whichever model you have open, filling in the bracketed inputs.
3. Check that model's Tuning Notes and add the suggested line if the failure pattern it describes shows up.
4. If the result still isn't usable, walk the Fallback Playbook — don't just keep re-rolling.
5. For a prompt not covered here that you already rely on, run it through the Portability Compiler once so it's ready for failover before you need it.
6. When you're close to a cap, run your next task through the Compression Pass before you hit the wall, not after.

FAQ

Is this OmniRoute? No, and it doesn't contain any of that project's code, configuration, or routing logic. This library takes the discipline behind an automated multi-provider gateway — cascading fallback, compression, and circuit breakers — and rebuilds it as something you run by hand from inside an ordinary chat window, for people who aren't deploying infrastructure.

Do I need to self-host anything, or use an API key? No. Every prompt is plain language pasted into any chat LLM's normal chat window.

Is this affiliated with Anthropic, OpenAI, Google, or the OmniRoute project? No. This is an independent, model-agnostic prompt library and isn't affiliated with or endorsed by any AI lab or the maintainer of the project that inspired it.

Do the Tuning Notes assume a specific model version? They describe general tendencies observed across current-generation chat models as of this library's publication date, not a specific version number — treat them as a starting adjustment to test, not a permanent rule, since providers update their models regularly.

What if my second model also fails? Walk the playbook again against a third option, or apply the Compression Pass and try your original model once more — degradation is sometimes transient and a shorter prompt succeeds where a long one didn't.

Notes: Digital product, delivered as the copy/paste content on this page — no separate file download required. All prompts, tuning notes, and frameworks are original writing.