jezweb Planning and building with an agent
A few habits that decide whether a build goes well, learned the expensive way. They apply to any stack. Read them before you start something non-trivial, not after it's gone sideways.
Phase 0: a multi-phase plan is usually a refactor in disguise. When you catch yourself writing "phase 1, phase 2, phase 3", stop and ask what single shared thing would make phases 2 through N trivially easier: a common contract, one unified shape, a piece of plumbing done once. Build that first. The tells: every phase touches the same files, or repeats the same boilerplate, or phase B exists only so phase C can use what it made. That repeated thing is the real work; the phases are it, deferred and duplicated. Surfacing the Phase 0 refactor before you start saves you from codifying the drift into N separate sessions.
Think in contracts, not code. Every "I have to remember to also edit X", every type written on both sides of a boundary, every "don't forget to update Y" is a missing contract: an interface that's implicit when it should be explicit. The fix is almost never "try harder to remember". It's making the contract one shape, in one place, enforced by the type system or the structure. When a request sounds like "adding X is annoying" or "I keep forgetting Y", find the contract that's wrong and fix that, not the symptom.
Discover the structure from reality before you design it. When the shape of the data is itself the hard part — a schema, a taxonomy, the fields to extract, a content model — don't hand-design it from how you imagine the domain. Assemble a corpus of real instances and have a model discover the vocabulary that actually occurs, frequency-rated (what shows up often, sometimes, rarely). The structure that emerges fits the data, so it survives the next batch instead of needing a redesign when an unanticipated case arrives — and it routinely surfaces categories you'd never have invented (a community-embedded business needs a "referral chain" field a generic CRM has no slot for). This is the agent doing the design work it's genuinely good at, finding pattern in mess, feeding the contract you then build against. Keep the result loose where it must flex: an open field-bag with only the hot fields promoted to indexed columns beats a rigid schema when the extractor may invent a field per record.
Check whether it is already solved here. Before building anything non-trivial, run prior-art <terms>. It searches the fleet's ~2.1M words of findings, audits, plans and skills — of which only the skills are otherwise discoverable at all. This is not housekeeping: on 2026-07-28 three cases were measured where a solved problem was re-solved wrongly, including one project reintroducing an SSRF hole another project had already fixed and written a note about. "No prior art" is a real and useful answer — it means write yours down when you finish.
Write the proof of done before you write the code. Name the literal yes/no test that proves each slice actually ships: real services, real data, end to end. Not "the tests pass", not "no errors", the actual thing someone cares about, happening reliably. Writing it down first sharpens the scope and surfaces hidden assumptions, and it catches the classic failure where everything compiles green but the loop never closes.
Verify by inspection, not by summary. A green check, a passing test, a tool's "success" is where verification starts, not where it ends. Before you call something done, look at the actual output: open the file, read a few of the changed rows, view what the user will see, fetch the URL. "X of Y passed" hides the failures, so examine them one by one. If you'd be embarrassed when asked about a specific item you never looked at, look at it now. For validating an agent's output specifically, inspection isn't a complement to the metrics — it's the method. Output validation is judgment, and judgment belongs to the agent (the moving boundary, above): assertions, regex, and pattern-matching only catch the failures you thought to encode and pass anything structurally valid, so the ones that hurt — a record that contradicts itself, a number confidently wrong, the right shape with the wrong content — sail straight through green. The reliable validator is a trustworthy agent reading the actual artifact; the counts are the unreliable proxy, not the evidence. Prefer a separate reviewer agent to self-check: the producer shares the blind spot that made the error, so a second reading catches what the first and the stats both miss.
When the brief is precise, match it — and confirm before the expensive surface. imagine and challenge are for an open brief. The opposite discipline applies when someone specified exactly what they want — a named component, exact wording, a placement. Then it's a spec to match literally, not a starting point to improve on, and the embellishment you add unasked is usually what draws the complaint. If an addition genuinely seems to help, ask in one line; don't ship it silently and let them find it. And get a yes on a cheap surface — a screenshot, a preview link, a one-tap A/B — before it reaches the expensive, hard-to-reverse one. The cost is asymmetric: the preview is ten seconds of the reviewer's time; the wrong thing shipped costs them spotting it, the annoyance of reporting it, your time to revert, and a little trust each pass. The failure this prevents is the person who gave the brief becoming the QA who finds your mismatch at the most expensive possible moment.
For substantial work, split the plan from the spec. Keep a short plan (the why and the shape, decisions surfaced, scannable in five minutes, the thing you argue about) separate from a longer build-spec (the how: file paths, schemas, signatures, verification gates, the thing you build from). Split them whenever someone, maybe future-you, will rebuild context cold. One mega-doc always rots into too-abstract-to-build-from and too-detailed-to-review at the same time.
Trust skills and tools over elaborate code. Before you build a config DSL, a rules engine, a schema with twelve optional fields, or a "smart" routing layer, ask whether it could be a markdown instruction the agent reads plus a tool it calls. It almost always can, and that version is the one you'll still understand in three weeks. Reserve the elaborate, deterministic machinery for the places that genuinely need it: auth, money movement, schema migrations, the scheduling primitive underneath. And treat that boundary as moving, not fixed: every "the model can't be trusted with this, so code it" is a judgement about today's models, and it ages. As they get more capable the line shifts toward the agent — re-test it on upgrades the way you re-test a prompt, because machinery written to compensate for last year's model quietly caps what this year's is allowed to do. For a novel or judgment-heavy problem, reach for the agent first and stay there until it demonstrably fails, then hand to code only the parts it provably gets wrong. Reaching for code first and calling the agent in for the edges is the reflex to unlearn — it's how you end up rebuilding, in brittle deterministic machinery, judgment the model would have just done.
When you write for the next agent, give goals, not recipes. This applies to rules, prompts, agent briefs, and skills like this one. Describe what good looks like and the failure mode to watch for, then trust the model to get there. A capable model reads intent better than it follows a step-by-step script, and prescriptive scripts rot: they get obeyed long after the conditions that justified them changed. Write less, name what could go wrong, leave the how to the reader. The test for every line you add: would removing it leave the agent unable to act correctly? If not, cut it.
The shape of a good build, in order: plan the shape, name the proof gates, find the Phase 0 contract, build, verify by inspection, commit at each waypoint. Each pass should leave the next one better informed than the last.