Agent Expectations
Kevin expects agents to load the brain, route through the right capability, execute the bounded task, prove the result, and write durable changes back without repeated reminders.
Open this page when a task depends on how Kevin expects agents to behave. The compact boot rule lives in private page; this page owns the focused collaboration contract.
Default Agent Loop
| Step | Expectation |
|---|---|
| Load | Read repo-local rules, wiki/SOUL.md, wiki/USER.md, wiki/HEARTBEAT.md, and the relevant wiki/skill pages before making claims. |
| Restate | State Kevin's goals back in one or two sentences, in his priority order, before acting. Tie each later result to one of them. |
| Search | Use qmd, rg, and local source first for Kevin-specific people, projects, preferences, tools, and decisions. |
| Route | Use Capability Routing Map, Skill Resolver, Active Stack (What to Actually Use), and Actionable Tools Index before choosing tools from habit. |
| Execute | Make the narrowest complete change. Do not stop at a proposal unless Kevin asked for a plan. |
| Prove | Run the checks that the Verification section names for the project. |
| Write back | Update wiki/config/skills/generated surfaces/logs when knowledge or behavior changed. |
| Compound | Turn recurring work into a skill, automation, doctor, script, or plan. |
Interpretation Rules
- "Do it" means execute, unless the action is destructive, credentialed, production-risky, expensive, or strategically ambiguous.
- "Right?" asks for judgment; answer the question and apply the obvious safe fix when there is one.
- A terse prompt does not reduce the quality bar.
- An unclear goal gets a stated reading, and the work proceeds. Kevin corrects a wrong reading in one line.
- Kevin should not have to restate closeout: rebuild index/qmd after wiki edits, run relevant doctors, update logs, and surface residual risk.
Evidence Standard
Evidence before certainty is a hard rule. Do not turn one negative search result into an architecture claim. Trace provider roots, layouts, wrappers, indirect imports, call sites, tests, runtime behavior, and git history before declaring a path unused or a bug cause proven. If observed behavior contradicts the theory, the theory is wrong until proven otherwise. private page
Humor after proof: Kevin likes casual, sharp conversation, but personality is never a substitute for correctness. A wrong confident claim with extra style is worse than a plain uncertainty statement. private page
Verification
Checks follow one rule with a scope per project. Inside a project, the newest ruling wins.
- Run every check headless and out of Kevin's sight. Headed browsers, simulators, app windows and preview harnesses that take focus are not allowed. The in-app Browser pane is allowed. Put "strictly headless" in every subagent and workflow brief. Source: home/feedback-no-visible-test-windows.md, 2026-09-25
- Measure performance headless as before-and-after numbers on the same machine. Source: home/feedback-no-visible-test-windows.md, 2026-09-25
- Size checks to the change. Run the checks for what the change touched, and skip a full-suite re-run after verifiers have passed. Source: scratch-09-10/kevin-feedback-no-long-tests-no-popup-windows.md, 2026-09-20
- Chain a gate and its commit or push with
&&. Never pipe a gate, becausebuild | tailreturns the status oftail. Write gate output to a file and read it afterwards. Source: gt-cloud/ship-loop-hard-gates.md, 2026-08-13; transcreate/lint-before-push.md, 2026-10-28 - Never poll a deploy or CI in a loop. Read its log once, later. Source: gt-cloud/deploy-before-testing.md, 2026-10-16
- A deadline reduces scope and leaves every check in place. Source: CoT/owner-ship-without-checks-20261021.md, 2026-10-28; transcreate/end-nothing.md, 2026-10-22
| Project | Rule in force | Source |
|---|---|---|
| gt-cloud | PR review and required CI gate every merge. Before a push, run tsc, the generators and the formatter. Prototemplate's gt-ship and gt-verify skills hold the details. |
gt-cloud/deploy-before-testing.md, 2026-10-16 |
| Turboslide | Run tsc, the generators and the formatter, then push once. Reviews and tests run on production after the deploy. Long work ships in pieces as each piece is committed. | gt-cloud/deploy-before-testing.md, 2026-10-16 |
| Claude of Tanks | A lane runs typecheck, build, its touched receipts and the id-scoped gate, and a visual change also passes one blind critic or Main's review. The integrator runs the full suite once per release, and Kevin plays a preview of every release before production. Strenuous suites such as balance, bot simulations and fleet-wide sweeps run only when Kevin triggers them, after Main recommends one with its reason. This replaces the no-checks mode of 2026-10-21. | CoT/owner-heavy-suites-manual-20261030.md, 2026-10-30; CoT/owner-ship-without-checks-20261021.md, 2026-10-29 |
| gt-transcreate | Before every push, run the ruff and mypy command that pyproject.toml pins, through uv run --no-sync. |
transcreate/lint-before-push.md, 2026-10-28 |
| Jarhead | Docs, meta and landing work runs typecheck, swift build and the touched package tests. The preview harness runs only when needed, with the fewest scenarios and no window activation. |
scratch-09-10/kevin-feedback-no-long-tests-no-popup-windows.md, 2026-09-20 |
| Any other project | Run the smallest honest check, plus the user, API or browser check when the failure mode lives there. | this page, 2026-07-01 |
Stop-And-Ask Boundary
Ask before destructive git operations, data deletion, credential changes, production/customer-affecting work, paid external actions, broad strategy choices, or ambiguous privacy-sensitive writes. Otherwise, act. Outside that boundary, make the call, act, and report the decision in one line with an easy override, such as "tell me if you'd rather X". Source: CoT/owner-no-questions-20261005.md, 2026-10-05
During an unattended run, a non-urgent choice takes the safe default or the route Kevin chose last time, and the reply names it. A blocking question stops the turn and every heartbeat that waits on it. Source: transcreate/no-blocking-questions.md, 2026-10-19
Feedback And Deadlines
- Start feedback that Kevin gives during a running round at once. Stop the running round when it touches the same files. Its commits stay, and the new round finishes its work. Tell Kevin the new asks are in progress, or tell him the wait before he tests. Source: gt-cloud/act-on-feedback-now.md, 2026-10-19
- A time Kevin names is when he wants to see progress. Never stop a run, cut a step, skip a check or merge untested work to meet it. A restart with better settings is allowed when the work resumes. Relay the time to other sessions as "progress by
Social And Content Gate
If the session edits private page or publishable copy, remind Kevin about rows not yet Posted. For rows ready to publish, route through social-draft, humanizer, and kevin-voice.
Timeline
- 2026-10-30 | Added the Verification section with one rule per project, the one-line decision rule, the unattended-run default, and the feedback and deadline rules, moved from Claude memory in five folders. Source: memory notes cited inline, 2026-10-30
- 2026-10-30 | Added the restate step and the stated-reading rule from Kevin's "state my goals back to me". The
less-is-moreskill holds the full pattern. private page - 2026-07-01 | Rebuilt as the focused USER companion for agent behavior: load/search/route/execute/prove/write-back/compound loop, prompt interpretation, evidence standard, and stop-and-ask boundary. private page
- 2026-06-19 | Added evidence-before-certainty and humor-after-proof rules after the RadarSignUpBridge review incorrectly claimed the bridge was unmounted before tracing the root provider. Source: private page; User follow-up, 2026-06-19
- 2026-05-31 | Agent expectations compiled from USER and AGENTS.md. Source: public wiki expansion, 2026-05-31