Prompt Strategy

With Eicra Soft, a first engagement costs a fixed $1,799: $300 for the prompt audit (up to 10 existing prompts, one use case), $600 for the pilot (one workflow, up to 5 production prompts) and $899 for the prompt library setup, built alongside the pilot. After that the Managed Prompt Desk is $999 a month for up to 10 prompt updates. For comparison, Upwork publishes a median AI engineer rate of $50 an hour, and Fiverr prompt gigs start around $15 with no testing or documentation.

A first engagement is live in about four weeks: a four-day audit, then a two-to-three-week pilot with the library setup built alongside it. Once the desk is running, prompt updates are delivered inside the monthly allowance. Delivery from Dhaka (UTC+6) gives overnight turnaround for US teams and same-day for the UK and Singapore.

A prompt strategy is the system around your AI instructions rather than the instructions themselves. It covers which business tasks are worth prompting, how each prompt is structured, what context it receives, what output format it must return, how it is tested, how versions are tracked and who maintains it. The deliverable is a prompt library with test sets and documentation, not a list of clever phrases.

One-off prompt writing gives you a phrase that worked once for one person. A prompt strategy gives you templates anyone on the team can run, a test set that says whether a change made things better or worse, version history, and documentation so the person who wrote it is not the only one who can maintain it. The difference shows up when the author leaves or the model updates.

Prompts that were tuned to one model version can behave differently on the next. That is why the test set exists. Under the Managed Prompt Desk, when a model version you use changes we re-run your test set on the new version and report which prompts behave differently, then fix the ones that matter within the monthly update allowance. Without the desk, you still hold the test set and can run it yourself.

No, and be careful with anyone who does. We do not publish accuracy percentages, hallucination-reduction figures or ROI guarantees, because the result depends on your data, your model and your use case. What we do commit to is testable: prompts are accepted against a test set and written criteria you agreed before the build, and rework is free when a prompt fails them.

1. We collect 20 real inputs from your own work, including the awkward ones. 2. You define what an acceptable output looks like for each, in writing. 3. The prompt is run across all 20 and the results are scored against those criteria. 4. Failures are diagnosed as a prompt problem, a context problem or a model limit, and the prompt is revised. 5. The test set ships with the prompt so the same check can be repeated later.

Every package has a published cap: 10 prompts and one use case for the audit, one workflow and 5 production prompts for the pilot, 15 templates across 3 use cases for the library, and 10 updates a month for the desk. More use cases, more prompts or an embedded engineer are quoted as a custom engagement after the audit, priced in writing before any work starts. The caps are deliberate, because a prompt that is written but never tested is not a deliverable.

Your company does, from day one. The prompt templates, system instructions, test sets, evaluation criteria, version history and written documentation live in your own drive or workspace, and we build inside your accounts rather than ours. At handover you keep every file, our access is removed, and nothing is retained on our side. There is no platform to stay subscribed to in order to keep using what you paid for.

Less than most people expect. We need sample inputs and outputs for the use case, your written quality criteria, and access to the AI tool account only if you want prompts tested in your own environment rather than ours. You create any accounts in your own name, grant access for the engagement, and we remove it at handover. Work starts after a mutual NDA, personal data is handled under a data processing agreement, and no sub-processor is added without your written approval.

Yes, and the audit is built for it. We read the prompts your team or a previous provider already wrote, record where each one fails and why, and rebuild the ones worth keeping as documented templates with a test set. You keep the audit findings and the rewritten templates whether or not you continue with us, so nothing you have already paid for is thrown away.

Prompt work is not model training, and it is not application development. We do not fine-tune or host models, build RAG pipelines or agent infrastructure, or write the software that calls your prompts, which is AI agent development and AI integrations work rather than prompt work, and we do not sign off on legal, medical or financial output that a licensed professional has to approve. Where a use case needs engineering rather than prompting, the audit says so in writing instead of selling you a prompt library that cannot fix it.