Upwork cover letter examples that open with the job

You want the site rebuilt so your marketing team can change copy without a developer. The reason that usually fails is not the CMS choice — it is that the page templates end up too rigid to say anything new in. I rebuilt [site] on [CMS] last [year] with a block-based editor: about eight composable sections the team combines themselves. They have shipped [number] new pages since without opening a ticket. Which pages do your team change most often? That tells me which blocks are worth building first.

Your checkout breaks when the booking plugin is active, and only on mobile. That pattern is almost always two plugins enqueuing conflicting jQuery on the same hook. I fixed the same class of conflict on [site] — the booking plugin was loading its own jQuery build after the theme had bound the checkout handlers. The fix was dequeuing the duplicate and rebinding, not disabling either plugin, so the client kept both features. Can you tell me which booking plugin it is? The dequeue differs between them.

You have been rejected twice under guideline [number], which is almost always about how the app explains what it does with the data it asks for — not about the data itself. I took [app] through the same rejection last [year]. The fix was rewriting the permission prompts to name the specific feature each one enables, plus a privacy screen before the first request. It passed on the next submission. What does your current rejection notice say verbatim? The wording tells you which of the two problems it is.

Your dashboard feels slow on the table view specifically, which usually means the list is re-rendering every row on every keystroke rather than the one that changed. I fixed exactly that on [project]: virtualised the list, moved the filter state out of the row components, and the input went from visibly laggy to immediate. No backend changes. Roughly how many rows are you rendering at once, and is the filtering happening client-side? That decides whether virtualisation alone is enough.

Your API times out under load but the database looks idle. That combination usually points at connection-pool exhaustion or a synchronous call to a third party inside the request, not at query performance. I traced the same symptom on [project] to a payment provider call made inline on every checkout. Moving it to a queued job with a webhook callback took p95 from [seconds] to well under a second. Are there any outbound HTTP calls in the request path? That is where I would look first.

You need someone who can take this from schema to deployed without handing off three times. The risk in that is not capability, it is that a solo build has no second pair of eyes on the decisions that are expensive to reverse. I built [project] end to end — [stack] — and handled it by agreeing the data model and the deploy story with the client in week one, before writing feature code. Both survived to launch unchanged. Do you have a schema already, or is that part of the work?

You want the checkout rebuilt in React with Stripe Payment Element, and the existing coupon logic left intact. That last part is where most rebuilds go wrong — the discount rules are usually older and stranger than anyone remembers. I did the same migration for [client type] last [year]: single-page React checkout, Stripe Elements, Apple and Google Pay, with the legacy rules ported rather than reimplemented. The coupon edge cases took longer than the payment work. Could I see your current discount rules before quoting? That is the part that decides the timeline.

Your customisations keep disappearing when the theme updates, which means they were made directly in the theme files rather than through an app block or a theme extension. I moved [store]'s customisations into theme app extensions so updates stop overwriting them. Same visual result, but the changes survive every update and can be toggled off from the editor. How many customisations are we talking about, and is the theme from the Shopify store or custom-built? That changes which route is cheapest.

Your pipeline finishes without errors but the numbers are wrong some days. Silent partial failure almost always means an exception is being swallowed inside a per-row try block. On [project] I found the same thing: rows that failed validation were logged and skipped, but nothing surfaced the skip count, so a bad upstream file looked like a quiet day. Adding a failure threshold that halts the run fixed it in an afternoon. Does your pipeline record how many rows it processed versus how many it expected?

You want to stop taking the site down to deploy. The blocker is usually not the deploy tooling — it is that migrations and application code have to be able to run in either order during the rollout. I moved [project] to zero-downtime deploys by splitting every schema change into an additive step and a cleanup step released separately. The pipeline change was the small part; the migration discipline was the work. How often do your releases include schema changes? That decides how much of this you actually need.

Your suite is green but bugs still ship, which usually means the tests assert on implementation rather than on what the user can observe. On [project] I rewrote the highest-churn suites to drive the UI the way a user does and assert on visible state. The suite got slower and caught considerably more. I also quarantined the flaky tests rather than deleting them, so the team could see what they were losing. Roughly what share of your tests are unit versus end-to-end right now?

Your feature works in testing and produces nonsense on real inputs. That gap is nearly always the prompt being tuned on clean examples while production data is messy, truncated, or in another language. On [project] I fixed it by building an eval set out of real failing inputs first, then changing the prompt against that rather than against handpicked cases. The failure rate dropped and, more usefully, became measurable. Do you have examples of the inputs where it fails? Those are worth more than any prompt rewrite.

You need the brand rebuilt so it still reads at 24px in an app icon and on a trade-show banner. That range is where most identity work quietly falls apart. I solved the same problem for [client type] in [year]: one mark with a simplified glyph below 32px, a type scale a non-designer can set without breaking it, and a one-page rule sheet so their team stopped asking me to resize things. Is there an existing wordmark you want kept, or is this a clean slate?

People are signing up and not reaching the point where your product does anything for them. That is almost never a visual problem — it is usually that the first screen asks for work before it has shown any value. I reworked the same flow for [product type]: moved account setup after the first useful action, cut the initial form to two fields, and completion went from [x] to [y]. Do you have funnel data on where people stop, or is that part of the work?

You have [duration] of footage and you want shorts out of it. The part that eats the budget is not the cutting — it is that footage shot for long-form usually has no moment that works as a cold open. I handled the same brief for [client type]: pulled [number] clips, re-framed to vertical with the speaker held on the safe area, and captioned for silent playback, which is how most of it gets watched. Is the footage scripted or conversational? That decides how much is actually usable.

Sessions are up and the revenue has not moved, which usually means the pages that grew are the ones ranking for queries with no buying intent behind them. I found the same pattern at [client type] in [year]: about [share] of new traffic was landing on informational posts with no route onward. Adding a genuine next step to the ten biggest ones, and re-targeting four of them, moved assisted conversions without adding a page. Can you share Search Console access? I would want to see query intent before quoting.

The platform says the ads work and the accounting disagrees. That gap is usually double-counted conversions on a page people reach more than once. I untangled the same thing for [client type]: deduplicated the events, moved the primary goal server-side, and reported ROAS fell by about a third overnight. That was the real number, and the budget decisions after it were finally the right ones. Are you on a standard ecommerce setup or a custom one? That decides whether this is days or weeks.

A list this size that is not producing revenue is usually a deliverability problem wearing a content problem's clothes — once enough of the list stopped opening, the engaged part stopped reaching the inbox too. I rebuilt sending for [client type]: segmented by genuine engagement, suppressed the rest, warmed back over [duration]. Volume dropped and revenue per send went up, which is the trade worth making. What does your open rate look like on the last ninety days versus the last year?

You want enquiries triaged and routed with a summary attached. The part that breaks is not the routing — it is what happens to the ones the classifier is unsure about, because silently misfiling those is worse than not automating at all. I built the same flow for [client type]: classification with a confidence threshold, anything below it into a human queue rather than guessed, and a weekly report on what landed there so the prompt could be tightened against real cases. Roughly what volume a week, and is the CRM already decided?

It works in simulation and fails on the line, which usually means the model assumes a tolerance the process cannot actually hold. I worked the same problem for [client type]: measured what the process was really delivering rather than what the drawing asked for, moved the critical dimension onto the feature that could hold it, and the reject rate dropped without changing the supplier. Do you have measurement data from the failing batches, or would that be the first step?

You want pieces that rank without reading like they were written to rank. That is a harder brief than it sounds — most SEO writing solves the second problem by ignoring the first. I wrote [number] pieces in [subject area] for [client type], built from interviews rather than from the other top-ten results. They ranked, and they were also the pages the sales team actually sent to prospects, which is the test that matters. Do you have subject-matter experts I can interview, or is this desk research?

Marketing copy rarely survives a literal translation — the line that works in [source language] is usually working because of a rhythm or a pun that does not exist in [target language]. I handled the same set for [client type]: translated the intent rather than the sentence, flagged the three headlines that needed rewriting outright, and gave you both options so the decision stayed yours. Is there a glossary or existing localised material I should match?

You are spending your mornings on inbox and scheduling, and the real cost is the context-switching rather than the minutes. I ran the same setup for [client type] for [duration]: everything triaged before 9am against rules we agreed in week one, anything genuinely needing you batched into one afternoon slot, and a Friday summary so nothing needed chasing. After a month they stopped opening the shared inbox. What are you using for scheduling, and is there anyone else I would coordinate with?

Two years of unreconciled books before a filing deadline is a different job from ongoing bookkeeping, and the risk is that whoever takes it starts at month one and runs out of time. I took the same situation for [client type]: worked backwards from the closing balance to find where it diverged — a single mis-mapped clearing account repeating monthly — then reconciled forward from there. A fraction of the time a straight rebuild would have taken. Which software are the books in, and is the deadline fixed?