Migrating Prompts From Claude/GPT To Hermes: Gotchas

A prompt is not pure logic — it is logic plus model-specific phrasing that the original model is good at following. Migrating it is part translation, part re-tuning, part discovering which parts of the original were actually doing work. Plan for each prompt to need a few rounds of iteration before it matches the source-model output.

Common gotchas, by source model

Things that usually port for free

• Plain task descriptions ('summarize this article in 100 words').
• Schema-locked output requests with one example.
• Anti-rules ('do not include code fences', 'do not preamble').
• Domain-specific glossaries and definitions.
• Few-shot examples — these often help Hermes more than they helped the source model.

Things that need real work

• Tool-using prompts — different grammar, different harness expectations.
• Long persona prompts — Hermes responds in kind, sometimes more than you wanted.
• Anything depending on memory or persistent state — you build the state layer yourself.
• Anything depending on a specific provider feature (web search, image gen, code interpreter) — those become separate components.

Migration workflow

1. 1Pick the highest-value prompt in your library.
2. 2Snapshot its current output on the source model for 10 representative inputs.
3. 3Translate to Hermes using the gotchas table.
4. 4Run on the same 10 inputs. Compare line by line.
5. 5Iterate on the system prompt until Hermes outputs are equivalent or you accept the gap.

Applied exercise

1. 1Open one prompt that you know works well on Claude or GPT.
2. 2Identify every line that depends on a model-specific feature using the gotchas list.
3. 3Rewrite those lines for Hermes.
4. 4Run on three sample inputs and compare. Note which adjustments mattered most for your use case.

The big idea: porting prompts is a translation, not a copy. Mind the gotchas, evaluate honestly, and trim what is no longer carrying weight.