Make it sound like us
In one line
One request, two different problems — the facts change weekly and belong in retrieval, the voice is stable and can be taught — and the ladder tells you how far to climb for each.
The situation
The support lead reads the assistant's draft replies and says: "These are accurate, but they don't sound like us. Make it sound like us." The engineering lead hears "fine-tune it", books three weeks, and starts assembling a dataset.
Before that, it is worth asking what "sound like us" contains — because it contains two things, and only one of them is a fine-tuning problem.
Two requests hiding in one sentence#
Read a hundred of the team's actual replies and the pattern is clear. There is a voice: first name greeting, two short paragraphs, a concrete next step, no hedging, sign-off with the agent's name. And there are facts: the refund window, the plan limits, the escalation path — all of which changed in the last quarter and will change again.
Voice is stable. Facts are not. Any technique that bakes facts into the model is wrong by construction, because the next policy change means another training run, and in the meantime the model confidently states last month's terms. Facts belong in : the current policy page, fetched per request, cited in the reply.
Voice is what techniques are good at, and the ladder orders them by cost.
Climbing only as far as needed#
Prompt and examples first. Put four real replies in the system prompt as examples of the voice. This takes an afternoon, is reversible in a deploy, and — measured on an where the support lead rates twenty drafts — gets most of the way. In this case, from "sounds like a robot" to "sounds like a new hire who has read the style guide".
Then, if the gap is still real, a small . Two thousand past replies, adapters, a day of training. Now it sounds like the team rather than like someone imitating the team. The adapter is megabytes, it swaps over the same base model, and it carries only form — the facts still come from retrieval on every call.
The three weeks that were booked would have gone into a full fine-tune that also memorised last quarter's refund window. The ladder prevents that not by forbidding the top rung, but by insisting each rung is measured before the next one is climbed.
Questions to ask
- Which parts of "like us" are stable, and which changed in the last quarter?
- Have I tried examples in the prompt before anything that trains?
- Who rates the drafts, and against what?
- If the policy changes tomorrow, which mechanism updates and which does not?
Mindset
Split every "make it do X" request into what changes and what does not. Changing things go in the prompt via retrieval; stable things are taught, starting with the cheapest teaching available and climbing only while the evals say to.
Where it connects
Phase 3 · 7 min
Retrieval & RAG
Fetch the relevant text at request time and put it in the prompt, because almost everything worth answering about is either private, recent, or needs a citation.
Phase 3 · 6 min
The adaptation ladder
Five ways to make a general model fit your task, ordered by cost and by how hard each is to undo — and the discipline is to stop climbing the moment the evals stop improving.
Got the shape of it?
Examples do not count toward phase progress — that stays on the topics. This is just so the list remembers what you have seen.