"Claude Code Q&A: What Works, What Doesn't, and What Will Save You Hours" - Katie Parrott

Why this is in the vault

This is the written-up Q&A from Every's first Claude Code Camp (July 2025). Kieran Klaassen answered subscriber questions, with Dan Shipper facilitating. It is an early, compact statement of the habits that became compound engineering. Most answers still hold; a few are now dated and worth knowing as dated. Full text from the paid re-read (piece updated 2026-09-09).

The core argument

Three patterns Parrott pulls from the session:

Planning is paramount. Plans are "maps for a journey," not instructions for how to walk.
Structure helps more than it constrains. PRs, branches and tests become checkpoints for AI work, even for solo developers.
Front-load context so the agent acts as an informed teammate rather than "a stranger dropped into the project."

Klaassen's answers, by theme:

Planning/prompting: Use a mega-prompt for the plan (every step, the what), but leave the how flexible so the agent can work agentically.
Delegation: He rarely corrects mid-flight; planning and review are the important steps. On agents going rogue: "If you have a precise, detailed plan, it never goes rogue." The PR is the checkpoint, letting the AI finish a full pass before a human steps in.
TDD: Define the tests, say "use TDD," and remind Claude not to write the full solution before running the tests, because it gets "eager."
Architecture: Tests, good code, good reviews. He has a command that pulls in all PR comments and resolves them in one pass.
Sub-agents vs. extended thinking: Sub-agents for many unrelated small tasks; "think ultra hard" for one genuinely hard question.
Bribes and threats ("I'll give you $1000," "I'll fine you $100"): he doesn't think these work anymore. He once tried a logging system that "fined" Claude $100 per failure.
CLAUDE.md: Use both a root file and directory-scoped files so the right context loads per area.
MCP: "Add one for anything you do manually."
Safety: --dangerously-skip-permissions is "very safe" inside containers. Use normal Git feature branches merged after review; Claude can revert its last change well.
Working style: Pomodoro, and batch work by type (an hour of review, an hour of research) to avoid fatigue across agents.
Learning: He lets several AI reviewers critique the same PR without picking a winner, to learn alternatives. Good ideas get codified into style rules so they apply next time. This is the seed of the "compound" step.
Errors: Don't paste logs; "let the agent decide" how much context it needs.

Other demos mentioned: parallel-research slash commands, running three features at once without conflicts, voice-to-text prompting, and Claude Code on a phone over SSH.

Mapping against Ray Data Co

"Precise plan → it never goes rogue" is the stated rationale for Ray's plan → tests → implementation order and the verify-dispatch gate. The Q&A gives the same causal claim, from the practitioner who popularized it, for why Ray's dispatch prompts carry acceptance contracts. Klaassen's "plan is a mega-prompt, how stays flexible" is a clean test for dispatch quality: specify outcome and constraints exhaustively, and leave method to the subagent.

Multi-reviewer "nice tension" → codified rules is the same mechanism as Ray's /improve loop and the self-review rubric. Critic disagreements should become rule updates, not one-off fixes.
Dated or contradicted advice: "Never goes rogue" is overstated. Later Every writing and Ray's own experience with false "verified" stamps and pointer returns show that a detailed plan reduces drift but does not remove it; independent verification is still required. "Dangerously-skip-permissions is very safe in containers" ignores credentials inside the container; see [[2026-08-14-every-ai-employee-security-framework]] on credential proxies. "Add an MCP for anything manual" predates the security-review SOP for installs and the tool-pool bloat it causes.
"Let the agent decide how much error context" fits the AGENTS.md subagent-routing rule: give the agent pointers to logs, not pasted dumps.

Why it matters for RDCO / The Denominator

Do: add one line to the verify-dispatch rubric: "Is the WHAT exhaustive and the HOW left open?" That is Klaassen's plan/how split as a checkable item.
Do (factory): adopt the "resolve all review comments in one pass" command pattern for the human review gate. Reviewers leave line comments and the agent addresses them together, which saves review round-trips (the Level 5 constraint).
Write (The Denominator): a short piece on what aged well and what didn't from mid-2025 Claude Code advice. Planning and checkpoints survived; "it never goes rogue" and "MCP for everything" didn't. That is a credible, non-derivative reframe for an audience of enterprise builders.

⚠️ Sponsorship

House promo: the piece markets Every's subscriber-only Claude Code Camp (next session August 22) with repeated subscribe CTAs, and the footer plugs Spiral, Sparkle, Cora, Monologue, the consulting practice and Compound Engineering. No third-party sponsor. Bias implication: answers come from Every's own product GM and are framed to sell the Camp. Confident claims ("never goes rogue," "very safe") read as event-pitch certainty, not tested results.

[[2026-01-26-every-claude-code-shipping]] - Klaassen's full Claude Code workflow
[[2026-01-28-every-stop-coding-start-planning]] - the planning-first argument expanded
[[2026-02-09-every-compound-engineering-guide]] - where the codify-the-learnings habit ended up
[[2026-03-13-every-compound-engineering-camp]] - a later Camp session write-up
[[2025-10-27-every-six-engineers-ai-workflows]] - the same engineers' stacks two months later
[[2026-04-15-thariq-claude-code-session-management-1m-context]] - context management, updated guidance
[[2026-08-14-every-ai-employee-security-framework]] - the security counterweight to "skip permissions in containers"