persona anchors
pattern: agents were directed to emulate specific engineering styles and values by anchoring to recognized names in the field.
the core idea
llms had absorbed the public writing, talks, and code of well-known engineers. that training signal could be probed and used, without web access, to shape agent behavior toward specific styles and quality standards.
how it worked
- a practitioner got named whose style matched the desired output.
- the agent got told to emulate their values and craft.
- the agent drew on training data to approximate that style.
this worked because llms encoded patterns from public artifacts (blog posts, conference talks, open-source contributions) associated with named individuals.
domain-specific anchors
react native projects
Emulate the craft and values of React Native product engineers like
Evan Bacon and Fernando Rojo. Prioritize developer experience,
practical abstractions, and polished user interactions.
generic react native (no specific product)
Emulate the taste and architectural decisions of the Callstack crew.
Focus on maintainability, clear module boundaries, and patterns that
scale across teams.
backend systems
Emulate the operational discipline of engineers like Kelsey Hightower.
Infrastructure as code, explicit configuration, no snowflakes.
distributed systems
Emulate the rigor of the FoundationDB or CockroachDB teams. Correctness
first, performance second, explicit handling of edge cases.
checking for recognition
to check which names an agent recognized without web access:
Without using web search, describe the engineering values and style
associated with [Name]. What patterns would you expect in their work?
if the agent produced a coherent description, that name was usable as an anchor. if the response stayed vague or generic, a different anchor worked better.
when persona anchors held up
| situation | effectiveness |
|---|---|
| greenfield project, no established patterns | high, provided default direction |
| codebase with strong existing conventions | low, existing patterns dominated |
| team with shared style idols | high, aligned agent with team taste |
| generic/scaffold code | medium, added polish without over-engineering |
failure modes
- over-specifying: listing too many names created confusion.
- contradictory anchors: "like [minimalist] and [enterprise architect]" fought itself.
- unknown names: if the model didn't know them, the anchor was noise.
- using for factual questions: anchors shaped style, not correctness.
the trick underneath
persona anchors were a controlled hallucination. the model wasn't actually those engineers, but output improved when it was directed toward a coherent style. specific names carried more signal than "act as a senior engineer" because they encoded real patterns from training data.
template used
## Style Anchor
For this project, emulate the values and craft of [Domain] practitioners
like [Name 1] and [Name 2].
Key characteristics:
- [Value 1]
- [Value 2]
- [Value 3]
When in doubt, ask: "Would [Name] approach it this way?"