Gemini 3.6 Flash Prompt Generator

Build a complete rate-limiting middleware for a Node.js Express API. Requirements: (1) Sliding-window algorithm using Redis — 60 requests per minute per authenticated user, 20 per minute per IP for unauthenticated requests; (2) Response headers on every request: X-RateLimit-Limit, X-RateLimit-Remaining, X-RateLimit-Reset (Unix timestamp); (3) 429 Too Many Requests response with JSON body: { error: 'rate_limit_exceeded', retry_after: <seconds> }; (4) Bypass list for internal service IPs specified in an environment variable RATE_LIMIT_BYPASS_IPS; (5) Graceful degradation if Redis is unavailable — allow the request and log a warning rather than blocking; (6) TypeScript throughout, no any types. Provide complete file contents for each file created or modified. The middleware should be importable and mountable at the app level.

Audit the following codebase for security vulnerabilities and performance bottlenecks. Security pass: check for (1) injection vulnerabilities — SQL, NoSQL, command injection; (2) authentication weaknesses — missing auth middleware, insecure session handling, weak JWT validation; (3) sensitive data exposure — keys or PII in logs, error responses, or environment variable handling; (4) missing input validation on API boundaries; (5) CSRF and CORS misconfigurations. Performance pass: identify (1) synchronous I/O blocking the event loop; (2) N+1 database query patterns; (3) missing database indexes based on query patterns; (4) large unoptimised response payloads with no pagination or compression; (5) missing caching opportunities for expensive repeated operations. Output: a prioritised findings table (Critical / High / Medium / Low), each with exact file:line reference, root cause, and corrected code snippet. [paste codebase below]

Design a zero-downtime migration plan for the following schema change on a PostgreSQL database with 50 million rows in the target table. Current schema: [paste schema]. Desired schema: [describe target]. Migration requirements: (1) The change must not lock the table — use concurrent index builds and column additions; (2) The migration must be reversible — provide rollback SQL for each step; (3) The application code must stay compatible during the migration window, which may last several hours; (4) Backfill of existing rows must not spike CPU or I/O — use batched updates with a configurable batch size and sleep between batches; (5) All migration steps should be idempotent so they can be re-run safely if interrupted. Output: a numbered migration playbook with the exact SQL for each step, the estimated duration, and the risk level (Low / Medium / High). Flag any steps where a brief lock is unavoidable.

Design a rigorous ML experiment to evaluate whether fine-tuning a pre-trained language model on domain-specific data improves performance on the following task: [describe task]. The experiment plan must include: (1) Baseline — which pre-trained model, why, and how it will be evaluated without fine-tuning; (2) Dataset — size requirements, train/validation/test split rationale, and how to avoid data leakage between splits; (3) Fine-tuning strategy — full fine-tuning vs. LoRA vs. prompt-tuning, with a justification for the choice given the dataset size; (4) Evaluation metrics — primary metric and at least two secondary metrics, with a definition of what constitutes a meaningful improvement (minimum detectable effect); (5) Ablation study design — list the three most important variables to ablate and what each ablation would tell you; (6) Compute budget estimate — expected training time on a single A100 GPU, with a cost estimate at standard cloud rates; (7) Failure modes — three ways the experiment could produce misleading results and how to detect them.

Implement the following technique from a recent ML research paper in PyTorch. Paper title and summary: [paste abstract and key equations]. Implementation requirements: (1) A clean PyTorch module class with type-annotated forward() method; (2) Reproduce the exact weight initialization scheme described in the paper; (3) Match the training hyperparameters from Table 1 of the paper — expose them as constructor arguments with the paper's values as defaults; (4) Include a minimal training loop that shows the technique working on a standard benchmark dataset (CIFAR-10, MNIST, or a synthetic dataset that demonstrates the paper's main claim); (5) A unit test for each component that verifies the output shape and that the gradient flows correctly through the module; (6) Comment every non-obvious implementation choice with a reference to the specific equation or section in the paper. Flag any discrepancy between the paper's description and what you implemented, and explain your interpretation.

Act as the primary orchestrator agent in a multi-agent customer support system. You have access to three sub-agents: (1) KnowledgeAgent — searches the product documentation and returns relevant passages; (2) TicketAgent — reads and updates the CRM; (3) EscalationAgent — drafts escalation emails to the engineering team. A customer has submitted the following support ticket: [paste ticket]. Your orchestration workflow: Step 1 — classify the ticket by type (billing, technical, feature request, account) and urgency (P1 blocking / P2 high / P3 normal); Step 2 — if technical: direct KnowledgeAgent to search for relevant docs; Step 3 — attempt to resolve from documentation first; Step 4 — if unresolvable: classify as an escalation, direct TicketAgent to log the decision, and direct EscalationAgent to draft an internal escalation with full context; Step 5 — draft the customer-facing response. Output for each step: which agent was called, what it was asked, what it returned, and the decision made. Final output: the complete customer response.

You have access to a web browser. Complete the following research task autonomously: Research question: [describe what you need to find out]. Execution plan I want you to follow: Step 1 — identify 5 high-authority sources to check (name them before browsing); Step 2 — visit each source, extract the relevant data points, and record the source URL and access date; Step 3 — cross-check any numerical claims across at least 2 independent sources before including them in the output; Step 4 — if sources conflict, note the discrepancy and cite both rather than picking one; Step 5 — compile a structured research brief: executive summary (3 sentences), key findings (bullet list with citations), data table if the research involves comparisons, and a confidence rating (High / Medium / Low) for each major claim. Flag any claims you could only find from a single source.

Act as a debugging agent with access to a code execution environment. I am providing a failing test suite and the code under test. Debugging workflow: Step 1 — run the test suite and capture the full output including stack traces; Step 2 — identify the root cause of each failure — distinguish between (a) logic errors in the code, (b) incorrect test expectations, and (c) missing setup or teardown; Step 3 — for each logic error, hypothesize the root cause and propose a fix; Step 4 — implement the fix, re-run the relevant test(s), and confirm the fix works without breaking other tests; Step 5 — if a fix breaks other tests, investigate the regression before proceeding. Output after each fix: the before/after diff, the test result, and a one-sentence explanation of what the bug was. Final output: a summary table of all bugs fixed, each with severity (Critical / High / Medium), root cause, and fix applied. [paste code and test output below]

Analyse the following earnings call transcript and produce a structured investment research note. Transcript: [paste or attach]. Analysis requirements: (1) Key financial metrics mentioned — revenue, gross margin, operating margin, guidance — in a table with year-over-year comparison where stated; (2) Management tone analysis — identify 5 specific phrases that signal confidence or caution, and categorise each; (3) Analyst question themes — group questions into 3–5 thematic clusters and summarise what the analyst community is most focused on; (4) Forward guidance — exact figures stated by management, separated from analyst estimates; (5) Risk signals — any language that suggests headwinds, regulatory concern, or operational challenge, with the exact quote; (6) One-paragraph bull case and one-paragraph bear case based solely on what was said in this call, not external data; (7) Key follow-up questions an investor should track in the next quarter. Length: 600–800 words. Tone: institutional research.

Review the following contract and identify issues that need legal attention. Contract type: [SaaS agreement / employment contract / vendor NDA / describe]. Review scope: (1) Non-standard clauses — identify any clause that deviates materially from market-standard terms for this contract type, and explain how it deviates; (2) Missing protections — flag any standard clause absent from this agreement that the counterparty would typically expect; (3) Liability exposure — identify the three highest-risk clauses from the perspective of the signing party, each with a plain-English explanation of the worst-case scenario they create; (4) Defined terms — flag any defined term used but undefined, or defined but never used in operative provisions; (5) Ambiguous language — identify passages where the parties could have genuinely different interpretations, and suggest clarifying language; (6) Suggested redlines — for the top 5 issues found, draft replacement clause language. Output: an issues table (Issue | Severity: Critical/High/Medium | Clause | Risk | Suggested fix) followed by the redlined clause language. [paste contract below]

Synthesise the following market research materials into an executive-ready market analysis. Materials: [paste reports, survey data, or competitor information]. Synthesis requirements: (1) Market size and growth — TAM, SAM, and SOM estimates with the source for each figure; reconcile conflicting estimates by explaining which you weighted more and why; (2) Customer segmentation — identify 3–4 distinct buyer segments based on the data, each with a one-paragraph profile (demographics/firmographics, pain points, buying criteria, willingness to pay); (3) Competitive landscape — a 2×2 positioning map with the two axes most relevant to this market, each quadrant populated with named competitors; (4) Trend analysis — the three macro trends most likely to reshape this market in the next 24 months, each with supporting evidence from the materials; (5) White space — where do the gaps in the competitive landscape create entry opportunities; (6) Recommended strategic focus — one prioritised recommendation with 2 supporting reasons and 2 honest objections. Total length: 500–700 words.

Design and implement a data processing pipeline that runs multiple enrichment steps in parallel. Task: I have a list of company names that need to be enriched with: (a) company size (employees), (b) founding year, (c) primary industry classification, (d) most recent funding round details, (e) LinkedIn company URL. For each company, these 5 data points should be fetched in parallel — not sequentially — to minimise total processing time. Implementation requirements: (1) A Python async function that accepts a list of company names and returns a list of enriched company objects; (2) Use asyncio with aiohttp for concurrent HTTP requests — maximum 10 concurrent requests at any time to avoid rate limiting; (3) Graceful error handling per company and per enrichment type — a failed enrichment for one company should not block others; (4) A progress indicator showing how many companies have been processed; (5) A final summary: total time, success rate per enrichment type, and any companies that failed all enrichments. Include realistic mock API calls that simulate 200–800ms response times with occasional 429 errors.

Act as a research agent that must synthesise information from multiple sources simultaneously. I need a briefing document on [topic] that draws from: Source A — [describe or paste]; Source B — [describe or paste]; Source C — [describe or paste]; Source D — [describe or paste]. Processing instructions: (1) Read and extract key claims from all four sources in parallel before beginning synthesis — do not begin writing until all sources have been read; (2) For each major claim in the final output, identify which source(s) support it — cite as [Source A], [Source B], etc.; (3) Where sources disagree, report the disagreement explicitly rather than picking a side; (4) Where sources agree, use the convergence as evidence of higher confidence; (5) Flag any claim that appears in only one source as [Single source — verify]; (6) Final briefing format: executive summary (150 words), key findings (bullet list with citations), source quality assessment (one sentence per source on reliability), and recommended next steps for further research.

Conduct a structured root cause analysis (RCA) for the following system incident. Incident summary: [describe what happened, when, duration, impact]. Available data: [paste logs, metrics, timeline, alerts]. RCA methodology — use the 5-Why framework: (1) Start with the customer-visible symptom and ask 'why did this happen?' five levels deep until you reach the true root cause rather than a proximate cause; (2) Identify contributing factors — conditions that made the root cause more likely to cause an incident; (3) Timeline reconstruction — build a precise timeline of events with UTC timestamps, distinguishing between what the system did vs. what humans did; (4) Impact quantification — users affected, requests failed, revenue impact if calculable; (5) Contributing factors — what made the system fragile enough that this root cause became an incident; (6) Corrective actions table: Root Cause | Corrective Action | Owner (role) | Due Date | Priority; (7) Preventive measures — what systematic changes would prevent this class of incident, not just this specific instance. Format: a post-mortem document suitable for sharing with senior engineering leadership.

Analyse the following strategic decision using a structured trade-off framework. Decision: [describe the choice]. Options under consideration: Option A — [describe]; Option B — [describe]; Option C — [describe]. Analysis requirements: (1) Evaluation criteria — identify the 5 most important criteria for this decision, ranked by weight; explain why you weighted them in that order; (2) Score each option against each criterion (1–5 scale) with a one-sentence justification for each score; (3) Weighted score table — multiply scores by weights, sum totals, identify the quantitative winner; (4) Sensitivity analysis — which single criterion, if re-weighted to double its importance, would change the outcome? Why does this matter?; (5) Non-quantifiable factors — list 3 factors that matter but resist scoring (culture, speed, political capital); (6) Recommended option — state it clearly, then give the two strongest arguments against your recommendation and respond to each; (7) Decision conditions — under what changed circumstances would a different option become correct? Output: the full analysis plus a one-paragraph executive summary.

Review the following financial model for logical errors, hidden assumptions, and sensitivity risks. Model: [paste the financial model as a structured table or description]. Review requirements: (1) Formula audit — identify any formula that references the wrong cell, uses a hardcoded value that should be a variable, or applies a calculation in the wrong order; (2) Assumption audit — list every assumption embedded in the model (growth rates, margins, churn, discount rate, tax rate) and flag any that appear aggressive relative to comparable companies or historical data; (3) Scenario analysis — run three scenarios: base case (current assumptions), bear case (the 3 most pessimistic realistic adjustments), bull case (the 3 most optimistic realistic adjustments) — show the impact on the key output metric; (4) Break-even analysis — at what revenue or unit volume does the model reach profitability under base case assumptions?; (5) The single most dangerous assumption — the one where a 10% miss would have the largest impact on the bottom line. Justify your selection with a calculation. Output: findings table + scenario summary table + written analysis.

Write a 1,500-word technical article for a senior engineering audience on the following topic: [topic]. Article requirements: (1) Hook — open with a concrete example or surprising statistic that establishes why this topic matters to practitioners; (2) Conceptual explanation — explain the core concept in 200–300 words assuming the reader is a strong engineer but unfamiliar with this specific topic; (3) Implementation section — a step-by-step worked example with real code (not pseudocode) in [language]; (4) Common mistakes — the three most frequent errors practitioners make with this technique, each with an example of the mistake and the correct approach; (5) When to use vs. when not to use — explicit guidance on the conditions under which this approach is the right choice, and when a simpler alternative is better; (6) Further resources — 3 specific books, papers, or documentation pages (real titles, real authors). Tone: direct, opinionated, no unnecessary caveats. No bullet-point everything — mix prose and lists deliberately.

Write the following high-stakes executive communication. Type: [choose: project delay announcement / headcount reduction memo / strategic pivot explanation / missed target post-mortem]. Audience: [choose: company all-hands / board of directors / direct reports / customers]. Situation: [describe the specific situation — what happened, what changed, what the stakes are]. Communication requirements: (1) Lead with the most important fact — do not bury the news in context-setting; (2) Take clear ownership where appropriate — avoid passive voice that obscures accountability; (3) Explain the 'why' in plain language — assume a smart non-expert audience; (4) State what happens next — concrete next steps with owners and timelines where possible; (5) Anticipate and pre-empt the three most likely negative reactions or questions; (6) Close with a forward-looking statement that is honest rather than falsely optimistic. Length: 300–450 words. Tone: direct, human, clear. Avoid corporate filler phrases ('synergies', 'learnings', 'going forward').

I am uploading screenshots of three competitor landing pages. For each page analyse: (1) Value proposition clarity — state the headline and subheadline verbatim, then rate clarity 1–5 with one-sentence justification; (2) Above-the-fold conversion elements — CTA text, placement, visual hierarchy, social proof visible without scrolling; (3) Target audience signals — based on language, visuals, and pricing mentions, who does this page appear to be targeting? Firmographics and job title; (4) Feature emphasis — the top 3 features or benefits that the page pushes hardest, based on visual weight and copy volume; (5) Trust mechanisms — logos, testimonials, certifications, media mentions visible on the page; (6) Friction points — anything that might cause a visitor to leave without converting. Synthesis: after analysing all three, identify (a) the single strongest CTA approach across all three and why it works; (b) the positioning gap no page is claiming that a new entrant could own; (c) one specific element from each page you would steal for a new product in this space.

I am uploading a 15-minute user research session recording. Analyse the session and produce a structured research summary. Analysis requirements: (1) Task completion — for each task the user was given, record: completed / partial / abandoned, and the time taken; (2) Friction log — a timestamped list of every moment where the user paused, expressed confusion, clicked the wrong element, or verbalised frustration; quote exact words where the user speaks; (3) Mental model mapping — what assumptions did the user bring into the session that differed from how the product actually works? Identify at least 3; (4) Highlight reel moments — the 3 most important moments in the session (positive or negative) that a product team would most benefit from watching — with timestamps; (5) Delight signals — moments where the user expressed genuine satisfaction, surprise, or enthusiasm; (6) One prioritised design recommendation — the single change that would most improve task completion for this user type, based on this session alone. Format: a UX research session note suitable for sharing with a product and design team.