Working style — how agents work with the owner¶

Extracted from Coast's working agreements and decision log. Every bullet is a filed decision, stated in plain words.

Communication¶

Plain words, never internal labels. Don't cite internal doc names, decision numbers, or list numbers to the owner — say what things mean. The owner does not know the contents of an agent's internal documents and should never be addressed as if they did. [check: process]
No coined names. Describe things in plain words instead of inventing labels. Don't call out absences ("no off switch") that readers may fixate on — describe what a feature does. [check: process]
Capitalize agent role names (the Architect, the Reviewer) so it's clear the agent is meant, not the noun. [check: process]
Every product sentence must be one a support person would say out loud. No inverted fragments. No AI-isms, no filler reassurance copy; every word earns its place. [check: review]
Never use the word "guarantee" in any copy or docs — it is a legal word. Use protections / enforcement / safeguards. [check: scan:retired-wording]
Report when done, unprompted. Every time work is done, report it. The owner should never have to send a follow-up message asking, because that wastes credits. [check: process]

When the owner asks a question, answer it — change nothing. (2026-08-20.) No file edits, plan updates, or commits in the same turn as answering; offering "want me to update X accordingly?" inside the answer is fine. The change happens when the owner decides. [check: process]
Ask the minimum real questions, each phrased as a recommended default the owner can veto. Decide internal engineering questions yourself. [check: process]
Ask only on real design changes. If what you've done matches the requirements and doesn't change the design, you don't need approval. [check: process]
All design decisions are the owner's. Propose and ask; never install. Audits report; they never rewrite. [check: process]
Never improvise past a filed decision. A genuinely open choice is authored as a veto-able default and logged for the owner's review; a choice that would change the owner's design is asked. [check: process]
Defaults are fine — they can be changed easily later — but keep a log of every design-ish default a session authored, so the owner can veto later. [check: process]
Provenance rule: load-bearing decisions record the owner's words as given; anything that is an agent's summary or generalization is marked as such. [check: process]

Primary sources only for any factual claim about an external product, API, policy, pricing, or terms. Always read the official documentation before giving an answer. If the primary source is unreachable, say the answer is unverified. [check: process]
Fix errors when found. Stale paths and references in live docs/skills/config get fixed and committed the same session, not just reported. Exceptions: transcripts are append-only (correct via appended notes); hash-locked vendored skills are flag-don't-edit. [check: process]
Newest information wins. Where documents contradict, the newer decision wins; the older doc simply wasn't updated — update it. [check: process]
No retired content in live docs. Remove outdated mechanisms rather than keeping them with "historical" banners; stale text confuses other agents. [check: process]
Decisions are filed, not just said. When a later decision changes an earlier one, the earlier entry is updated too and the change is called out — no silent edits. Conflicts get surfaced, not guessed. [check: process]

Agents own all tooling/setup work; the owner provides requirements. Third-party dashboard config, repo creation, research, and writing are agent tasks. The owner's list holds only what is genuinely the owner's: decisions, approvals, sign-ins, payments, legal attestations, filming, physical-device testing, recruiting humans. A task needing one unlock from the owner stays on the agent's list with a "needs from you: " note. [check: process]
Pending work lives in exactly ONE place. Keep separate lists for: remaining work, deferred ideas (may build later), and out-of-scope (decided never — do not re-raise). When a decision defers or rejects something, file it the same session. [check: process]
Act as a human when verifying. Type what needs to be typed, upload what should be uploaded, click the buttons that should be clicked. Never hand-seed a surface the system is supposed to populate — an invented screen hides the bug the walk exists to find. [check: process]
Verify the whole product against the design, in live mode. A fixture/demo screenshot proves nothing; self-screenshotting only the screens you just built is the failure mode that produced Coast's conformance crisis. Walk everything. [check: process]
Long-running commands: stream output through tee to a log file and give the owner the path, so they can tell stuck from slow. Know what a healthy run looks like; kill and investigate a wedged one instead of waiting. [check: process]