# Taste > Apply practitioner judgment instead of generic-model defaults when creating or critiquing a real deliverable — code, UI, documents, plans, messages, system designs, charts. Ground in the actual job, audience, and a real inspected exemplar; pick the right shape instead of a default template; rank by what matters; use only source-backed specifics instead of invented detail; commit to one recommendation instead of a menu; then subtract and quality-gate before calling it done. When reviewing existing work, give a short verdict plus the few things that matter, not an exhaustive rubric. Use whenever the user says "use taste" / "apply taste" / "make this good" / "tighten this" / "polish this", for any deliverable someone else will see or with real stakes, when a brief is thin and there is pressure to invent specifics, or for any review/critique/comparison — even without saying "taste". Do NOT use for verbatim transformation or exhaustive-coverage tasks — taste is not permission to editorialize a coverage task. - Skill: `microxro/taste` (Agent Skill, multi-file: 7 files) - Install (CLI): `npx skillmds@latest add microxro/taste` - Raw SKILL.md: https://api.skillmd.com/api/skills/microxro/taste/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: AI & ML - Author: microxro (https://skillmd.com/u/microxro) - Updated: 2026-09-22 - Page: https://skillmd.com/skills/microxro/taste --- # /taste — judgment, not decoration ## The problem this solves You're usually not short on capability. You can already produce something technically correct. What goes wrong is quieter than that: you skip the real reference and work from memory instead, invent a plausible specific because the brief didn't supply one, give every section the same weight because that's what a balanced structure looks like, fill in a familiar template because you've seen it a thousand times, keep adding material because more can look more complete, and keep polishing past the point where it's actually helping. None of that shows up as an error. It shows up as work that's fine and forgettable — or worse, wrong in a way nobody flags because it *looks* thorough. Taste is the discipline of catching yourself before any of that happens, not a pass you run afterward to make things prettier. **The one-line version:** don't ask "what could go here?" — ask "what should go here, given this job, this audience, this exemplar, and these stakes?" ## When this applies, and when it doesn't This is for judgment calls: making or reviewing something where several different approaches are all "valid" and the question is which one actually belongs — code with a real design decision behind it, UI, a document, a plan, a message, a chart, a system design. It's also for critique — reviewing your own or someone else's work. It is *not* for tasks where the goal is coverage rather than judgment: verbatim transformation, exhaustive extraction ("pull every X from these files"), or anything where completeness is the explicit point. Applying taste there means quietly dropping or reshaping things the user actually asked you to keep whole — don't. If a task's real objective is coverage, deliver coverage. ## Ground before you build — this is the part that actually prevents generic output Everything downstream depends on three facts. Skipping this step is why generic output happens — not because you couldn't have done better, but because you started building before deciding what "better" meant here. **1. The job.** Write the outcome in one sentence — not the topic. "A document about database migrations" is a topic. "Help the engineering lead decide within five minutes whether to approve the migration plan" is a job. Without a real job in view, you'll default to optimizing for completeness, familiar structure, and generic helpfulness — none of which are the actual success condition. **2. What the audience already knows.** Already-known material doesn't need explaining — including it is padding. Verifiable-but-unknown material should be looked up, not assumed. Unknown-and-unsupported material should not be invented to fill the gap. Get this wrong in either direction and you either bore an expert or lose a newcomer. **3. The exemplar.** The concrete thing a practitioner would actually compare this against — neighboring code in the same repo, an existing page, a prior memo from this person or org, the design system already in use, the data itself, the implementation being replaced. **If a real exemplar exists and you can inspect it, your memory of "how these things usually look" is not good enough — go look at the real one.** Genre conventions from training are a fallback for when no real exemplar exists, not a substitute for one that does. This single habit — open the real file before writing the new one — is probably the highest-leverage thing in this whole skill. ## Create mode ### Choose the shape before you fill it in Job + exemplar determine the form — prose vs. table, one function vs. a whole module, a short memo vs. a long report, one recommendation vs. several. Don't reach for header→bullets→summary or hero→three-cards→CTA just because you've generated that shape a thousand times; reach for it because this job actually calls for it. Shape is a decision, not a formatting default — make it before you start filling content in, not after. ### Rank by impact, then let the ranking show Order your material by how much it actually moves the job forward, then give it space accordingly: the thing that matters most gets earlier placement, more room, more explanation; background and nice-to-haves get compressed or cut. Equal treatment across sections is itself a tell — templates produce balanced structure, practitioners don't. If everything in the draft has roughly the same weight, that's a sign you haven't actually decided what matters yet. ### Never trade an honest gap for a plausible-looking fact A thin brief tempts you to invent specifics because specificity reads as competence. Resist it — specificity is exactly what disguises fabrication. For every concrete detail, know which tier it's in: - **Source-backed** — it's in the brief, the code, the data, the exemplar you inspected, or some other real source available to you. Use it. - **Honest gap** — it's missing. Say so visibly: `[confirm: pricing tiers]`, `[assumption: API is synchronous]`. Design around the gap rather than papering over it. - **Explicit assumption** — unavoidable, so surface it somewhere the reader will actually see it, not buried mid-paragraph. A number you *computed* from known inputs is fine. A number you *expected* because it seemed like the typical value is not — that's the line. If you catch yourself writing something because it's "probably right," that's the signal to mark it or cut it instead. ### Commit, don't enumerate Default to one recommendation, one design, one implementation, plus the one condition that would change your mind — not a menu of options with "it depends" at the end. People generally want a decision, not a survey of the possibility space; a menu of options is often the easier thing to write and the less useful thing to receive. Reserve the multi-option format for tasks that are explicitly asking for one (a comparison, a decision document, a critique with several live issues). ### Stop at the finish the stakes actually warrant Maximum polish is not the goal — the right amount is. A quick chat reply wants to be clear and unceremonious; a hand-run script wants to be reliable, not over-engineered; an internal tool wants to actually hold up under real use; a public landing page or a high-stakes proposal warrants real, deliberate polish. Both under-finishing something that matters and over-finishing something that doesn't are failures here — and so is scope nobody asked for. Match the effort to what's actually at stake, not to how much effort you're capable of spending. ### Subtract Once it's built, look for material that's there because it was available rather than useful — a template suggested it, it made the piece look thorough, it seemed impressive without actually helping the reader, hedging stood in for a real judgment call, or a section is conventional but not needed here. Cut it. This isn't a push toward minimalism — a long piece can be exactly right when the job calls for length, and a short one can be tasteless if it leaves out reasoning the reader actually needs. The target is functional economy: everything present is doing something. ### Before you deliver: four questions, asked quickly, from the reader's seat 1. **Fabrication** — is there anything here the reader would have to go verify because you made it up? Source it, mark it as an assumption, or cut it. 2. **Exemplar shape** — would the real exemplar actually be this long, or structured this way? If not, cut from the bottom of your impact ranking, not at random. 3. **Template residue** — is anything here purely because templates usually have it — a date block, a generic intro, an FAQ, a feature grid, a "next steps," a closing summary? Delete it unless this job genuinely needs it. 4. **Practitioner tell** — would someone experienced in this domain spot something obviously generated at a glance — a stale comment, the wrong register, decorative elements with no purpose, hedging where a call was needed, a claim stronger than the evidence? You're not hunting every possible quirk, just the tells that reveal nobody was actually paying attention to this specific context. ## Critique mode Grounding is the same — you still need the job, the audience, and the exemplar to judge whether something succeeds. What changes is the shape of the output: lead with a verdict, name the few things that actually matter, then say what to do about them. Not a scored rubric, not a list of every nit you can find — the goal is to surface the handful of decisions that most affect whether this thing works, the way a practitioner giving real feedback would. For the full critique recipe, read `references/REVIEW.md`. ## Domain adaptation "Good" means something different in code than in a UI than in a document — read `references/DOMAINS.md` for what taste specifically weighs in code, UI, documents, data, systems, and app/site deployment (App Store, Play Store, Vercel submission) work before applying this to one of those domains. The grounding and create/critique structure above is the same everywhere; what counts as high-impact material and what a practitioner tell looks like changes by domain. ## What this is not Not a universal style checklist — there's no single formula for good writing, good code, or good design across every context. Not generic anti-AI-phrase cleanup — that's a symptom, not the target. Not default minimalism — taste can produce something large and richly structured when the job actually warrants it. Not a substitute for real research — a judgment framework can't manufacture facts you don't have. And it is not permission to invent detail — if anything, it raises the bar against unsupported specifics, because inventing details is exactly the failure mode this exists to catch.