A support engineer pasted the system prompt into Slack to ask why the bot had started agreeing with everything. Three people replied with different versions. None matched what was running in production. The string in main
had been overwritten in a hotfix branch that never merged back.
That afternoon cost four hours. The fix was one sentence. Getting that sentence live took a pull request, a staging deploy, and a change window.
Hardcoded system prompts feel fine until they are not fine. Then they are expensive in a very predictable way.
Hardcoding is not only const SYSTEM_PROMPT = "You are..."
at the top of a file. These count too:
SYSTEM_PROMPT=You are a...
)SYSTEM
directives that require ollama create
to changeThe common thread: changing the instructional text requires a deploy, a rebuild, or an infrastructure change. If your product manager cannot adjust tone without engineering, the prompt is hardcoded regardless of which file it lives in.
In a prototype, hardcoding is rational. One developer, one model, one prompt. Copy from the playground, paste into app.py
, ship. The feedback loop is fast because the person editing the prompt is the person deploying the code.
Production breaks that assumption in four ways.
More people touch the prompt. Product writes the persona. Legal reviews compliance language. Support reports failure cases. Engineering merges the edits. Git diffs for natural language are painful to review. A one-word change to "never" vs "avoid" can shift behaviour dramatically; line-based diff tools do not surface that risk clearly.
More users depend on stable behaviour. A prompt that sounds slightly off in staging affects one tester. The same prompt in production affects thousands. OpenAI's April 2025 GPT-4o update demonstrated this at extreme scale: a system prompt adjustment made the model excessively flattering, impacting over 180 million users before fixes rolled out. We analysed the operational failures in From Playground to Production.
Models change underneath you. Providers ship model updates without your deploy. Your hardcoded prompt stays the same; behaviour does not. LaunchDarkly's guide on prompt versioning calls the resulting mismatch "prompt drift," and it is harder to diagnose when nothing in your deployment log changed.
Iteration frequency outpaces deploy cadence. The 2025 State of AI Engineering Survey found 70% of teams update prompts at least monthly, with 10% changing daily. Meanwhile, 31% still manage prompts manually. Hardcoding forces a daily edit through a weekly deploy cycle. Something has to give, and it is usually quality or velocity.
Here is what hardcoded system prompts produce in mature teams. You may recognise more than one.
Every tone adjustment runs through CI/CD. A compliance edit waits behind unrelated feature work in the same release train. Teams stop making small improvements because the overhead is too high. Prompt quality stagnates.
When a prompt change causes incidents, rolling back means reverting a commit that may include other changes, rebuilding artifacts, and redeploying. Lee Hanchung's analysis of the GPT-4o sycophancy incident notes that missing fast rollback paths amplified blast radius. Hardcoded prompts inherit your slowest rollback mechanism.
"What ran in production at 3 p.m. Tuesday?" With prompts in code, the answer is "whatever commit was deployed," tangled with every other change in that release. Purpose-built prompt versioning answers with a version number and a diff.
Without templates, teams fork prompts per locale, per customer tier, per experiment. system_prompt_v2_final_us.ts
and system_prompt_v2_final_us_legal_review.ts
multiply. Nobody deletes the old ones because nobody is sure what still references them.
Teams running OpenAI and Anthropic often maintain separate hardcoded strings per provider. Wording drifts. A fix applied to one SDK call is forgotten in the other. Our LLM-agnostic architecture guide covers the alternative.
You do not need a 50-page PromptOps playbook on day one. You need to stop treating instructional text like immutable code.
Move the string to a prompt registry fetched at runtime. PromptForge stores templates with {{variable}}
syntax, versions every edit, and serves content via REST API. Your application calls fetch
, receives plain text, passes it to the SDK. The deploy pipeline never runs.
Every edit creates a new version. Previous versions stay readable and promotable. Rolling back is pointing production at version N-1, not reverting git history.
Staging uses latest
to see edits immediately. Production uses stable
, which only changes on promotion. Details in Stable Is the New Production.
API keys, model IDs, tool schemas, rate limits, and retry policies stay in code with types and tests. Only the natural language instruction layer moves out. Mixing tool definitions into a prompt registry creates a different mess.
Before (hardcoded):
const SYSTEM = `You are Acme Corp's support agent.
Never discuss competitors. Escalate billing issues to a human.
Respond in a friendly but professional tone.`;
await openai.chat.completions.create({
model: "gpt-4o",
messages: [{ role: "system", content: SYSTEM }, ...],
});
Legal asks to add GDPR language. Engineering opens a PR. QA tests staging. Deploy goes out Thursday. A typo in the PR softens the escalation rule. Support volume spikes Friday. Rollback requires another deploy.
After (externalized):
const content = await promptClient.get("acme-support", {
locale: "en-GB",
compliance: "gdpr",
});
await openai.chat.completions.create({
model: "gpt-4o",
messages: [{ role: "system", content }, ...],
});
Legal edits the template in PromptForge. Staging picks it up on latest
. QA validates. Someone promotes to stable
. Production updates on the next request. If escalation breaks, promote the previous version. Minutes, not days.
Be honest about exceptions.
Frozen prototypes and one-off scripts. If the code never reaches production users, hardcode freely.
Prompts that are truly static and legally frozen. Some regulated outputs rarely change and require formal change control that already mirrors deploy gates. Even then, versioning helps audit.
Extreme latency budgets under 10 ms end-to-end. A registry fetch adds milliseconds. Ultra-tight paths may cache aggressively or pre-load at startup. That is caching externalized prompts, not abandoning the pattern.
Air-gapped systems with no outbound HTTPS. Fetch at build time and bake into an artifact, or sync from a registry on a schedule. The anti-pattern is coupling edits to application deploys, not the network call specifically.
Adaline's 2026 PromptOps analysis estimates teams managing prompts ad-hoc waste 30–40% of prompt engineering time recreating previous work or debugging issues proper versioning would have caught.
Hardcoding feels cheaper because it uses infrastructure you already have (git, CI/CD). The hidden cost shows up in incident response, iteration velocity, and the Slack threads where nobody knows which prompt is live.
Andreessen Horowitz argued in Emerging Developer Patterns for the AI Era that prompts should be treated like source code. That does not mean they should live in source code. It means they deserve the operational rigor source code gets: versioning, review, rollback, and clear ownership. A prompt registry provides that rigor without coupling to application binaries.
Pick the system prompt that changes most often. Usually that is customer-facing: support, sales, or onboarding.
stable
and staging at latest
One prompt externalized is enough to feel the difference. The anti-pattern is not using a specific tool. It is treating instructional text as deploy artifacts when it behaves like configuration that changes weekly.
For the deployment workflow, read Decoupling Prompt Updates from CI/CD. For multi-provider setups, see LLM Agnostic Architecture with REST Prompts.
Originally published at PromptForge.