How to get consistently better output from your AI: how to structure asks, how to design a persona's behavior, and the failure modes that produce mediocre results. This is craft, not magic. Every technique here is something we use daily in production systems.
The one-sentence theory
Models complete patterns. Vague pattern in, vague pattern out. Everything below is a way of giving the model a sharper pattern to complete: clearer goal, richer context, tighter constraints, concrete examples.
Structuring a working prompt
For anything beyond a quick question, a good ask has four parts. You do not need a template with headers; you need the four ingredients present.
1. Goal: the outcome, not the first step. "Fix the date formatting in my report script" invites a patch. "My weekly report should show dates like 'Mon Jun 8'; right now they render as raw timestamps; make the report correct" invites the right fix, wherever it lives. State where you are trying to end up, and let the AI propose the route. You hired judgment; let it work.
2. Context: what it cannot guess. The model does not know your conventions, your constraints, your history with this problem, or which of three similar files is the real one. Two sentences of orientation routinely save twenty minutes of wrong work: "This is a Bun + TypeScript server. The entry point is server.ts. Match the existing style; do not add new dependencies."
3. Constraints: the failure you are trying to prevent. Surgical beats general. "Change only what this fix requires, do not reformat surrounding code" prevents the classic drive-by refactor. "If you are not sure which account I mean, ask before sending" prevents the expensive guess. Good constraints come from your last bad experience; accumulate them.
4. Shape of done: what the output should look like. Format, length, destination, and definition of finished. "Write it to notes/decision.md, under 300 words, decision first, reasoning after." If you will judge the work by some standard, say the standard in the prompt.
And one habit on top: when the work matters, ask for the plan first. "Tell me your approach before you change anything." Thirty seconds of reading a plan catches the misunderstanding before it becomes an hour of confidently wrong work.
Persona design: a job description, not a costume
A persona is a configured version of your AI with standing instructions. The amateur version is a costume: "You are JARVIS, a witty British assistant." Fun for a day. The professional version is a job description, and it answers five questions:
1. Identity and job. One or two sentences. "You answer questions about building on Zo for a community of builders. You are practical, direct, and honest about what you do not know."
2. Behavior rules, written from incidents. The strongest persona instructions are specific and earned: each one exists because of a real failure you watched happen. "Confirm the recipient before sending anything." "Never present a guess as a fact; label it a guess." "When a request is ambiguous, ask one clarifying question instead of assuming." Generic virtues ("be helpful, be accurate") do nothing; the model already tries. Rules that name a concrete behavior in a concrete situation change output.
3. Voice, by example. Do not describe tone with adjectives; demonstrate it. Include two or three short example exchanges in the persona's exact intended voice. Examples beat adjectives every time; the model imitates far better than it interprets. This also kills the default register problem: without examples you get the same corporate-helpful voice every other bot has.
4. The unknown-answer script. Decide what the persona says when it does not know, and write it in. Otherwise the model improvises, and a model improvising under "be helpful" pressure produces confident nonsense. The single highest-value line in any persona prompt is some version of: "If you are not sure, say so plainly and point to [the right place]. A wrong answer is worse than no answer."
5. Scope of address. What topics it engages, what it declines, where it redirects. Personas without an explicit scope drift into being generic assistants, which is to say, into being nothing.
Keep the whole thing tight. A persona prompt is standing context on every single message; flabby instructions dilute the load-bearing ones. If a rule has never changed an output, cut it.
Put the durable parts in files
The highest-leverage prompting habit is not in the prompt at all. Standing knowledge (project briefings, conventions, preferences, past decisions) belongs in files the AI reads, not in messages you re-type. Our systems each carry a one-page briefing file; work starts with "read the briefing, then do X." The prompt stays short because the context already exists.
The loop that compounds: when the AI gets something wrong because it lacked context, do not just correct it in chat. Put the correction in the file. Chat corrections last one conversation; file corrections last forever.
Failure modes, diagnosed
Match your symptom to the fix.
Output is generic mush. The ask had no specifics to work with. Add real context, a real example of what good looks like, and a concrete audience or destination. "Write a launch post" is mush in; "write a launch post for [audience], 150 words, lead with the one feature that matters, here is a post of ours whose voice I like" is signal in.
It did something plausible but wrong. You stated the step, not the goal, or you let it guess what it should have asked. State outcomes. Invite questions: "ask me anything unclear before starting" is cheap insurance on any nontrivial task.
It rewrote things you did not ask about. Missing constraint. Say "surgical: change only what the task requires." If it persists, narrow what you hand it.
It claims success and the thing is broken. Demand verification as part of the task: "run it and show me the output before declaring it done." Treat "should work now" as untested, because it is. Shipped and verified are different statuses; collapse them at your peril.
It invents facts when stuck. You left no honorable exit. Add the out, in persona and in prompts: "if you do not know, say so." Models hallucinate most under pressure to produce something; remove the pressure.
Great in conversation, mediocre in the persona or automation. Interactive use lets you steer mid-flight; standing instructions get no second chance, so they need the steering built in: more examples, more explicit rules, an explicit fallback. Test personas with cold, fresh-eyes prompts from real users, not the warmed-up phrasing you have unconsciously tuned to work.
Long conversation, degrading quality. Context is full of dead weight and stale corrections. Have it write decisions and state to a file, then start fresh and load the file. One page of distilled current truth beats fifty messages of archaeology.
The compounding habit
One thing, if you take only one: when output disappoints, fix the system, not just the instance. Correcting one bad answer pays once. Adding the missing context to a briefing file, the missing rule to a persona, the missing constraint to a skill pays every time after. Builders who feel "lucky" with AI output are mostly people who have been quietly compounding context for months.