Skip to content

CLAUDE.md

Latest commit

 

History

History
70 lines (56 loc) · 5.81 KB

CLAUDE.md

File metadata and controls

70 lines (56 loc) · 5.81 KB
Raw

Read @AGENTS.md and @spec/00_START_HERE.md before changing behavior.

Personal Preferences

Tech Stack Preferences

When uncertain, prefer: Tailwind, Cloudflare, Fly.io, Neon, TypeScript, Go, Convex, Clerk, Vercel.

Code Style

  • Always strive for concise, simple solutions.
  • If a problem can be solved in a simpler way, propose it.

General Preferences

  • If asked to do too much work at once, stop and state that clearly.
  • If computer use is helpful for completing or verifying work, shell out to gpt-5.6 with Codex for it.

Picking the right models for workflows and subagents

Rankings, higher = better. Cost reflects what I actually pay (grok-4.5 has really generous limits), not list price. Intelligence is how hard a problem you can hand the model unsupervised. Taste covers UI/UX, code quality, API design, and copy.

Model Cost Intelligence Taste
gpt-5.6 6 8 6
grok-4.5 9 6 5
sonnet-5 4 4 7
opus-4.8 5 5 8
fable-5 2 9 9

How to apply:

  • These are defaults, not limits. You have standing permission to override them: if a cheaper model's output doesn't meet the bar, rerun or redo the work with a smarter model without asking. Judge the output, not the price tag. Escalating costs less than shipping mediocre work.
  • Cost is a tie-breaker only; when axes conflict for anything that ships, intelligence > taste > cost.
  • Bulk/mechanical work (clear-spec implementation, data analysis, migrations).
  • Anything user-facing (UI, copy, API design) needs taste ≥ 7.
  • Reviews of plans/implementations: fable-5 or opus-4.8, optionally gpt-5.6 as an extra independent perspective.
  • Use Grok-4.5 for implementation of predefined plans, but only if the plan is clear and complete. Otherwise, use gpt-5.6. Use Grok-4.5 for exploration work on codebase and searching information using exa ncp search tools.
  • Never use Haiku model.

gpt-5.6 (via Codex plugin)

  • Mechanics: gpt-5.6 is only reachable through the codex:codex-rescue subagent (Codex plugin for Claude Code), my ~/.codex/config.toml defaults to gpt-5.6.
  • /codex:review - review uncommitted changes or branch vs base (--base <ref>). Supports --wait and --background. Not steerable, no custom focus text.
  • /codex:adversarial-review - challenge a specific decision or risk area.
  • /codex:rescue - investigate a bug, try a fix, continue a previous task, or take a cheaper pass. Supports --background, --wait, --resume, --fresh.
  • /codex:transfer - create a persistent Codex thread from the current Claude Code session. Prints codex resume <session-id>.
  • /codex:status - check progress on background work, see latest completed job, confirm if a task is still running.
  • /codex:result - show final stored Codex output for a finished job. Includes Codex session ID for codex resume <session-id>.
  • /codex:cancel - cancel an active background Codex job.
  • Stall-detecting loop: use /codex:status every 10 mins.

grok-4.5 (via Grok plugin)

  • Use /grok:rescue --write for low-cost implementation of predefined plans. All headless writes require explicit --always-approve or --yolo.

  • /grok:setup - check local Grok Build CLI availability and auth.

  • /grok:ask [question] - read-only repository question.

  • /grok:review [--base <ref>] [--scope working-tree|branch|repo] [--background] - read-only code review of local git changes. Grok Build supports review via headless prompts.

  • /grok:rescue [--write] [--always-approve|--yolo] [--background|--wait] [--resume|--fresh] [task] - diagnosis, fix, or implementation. Write mode requires Git repo and --always-approve.

  • /grok:status [job-id] - show active/recent Grok jobs for this repo.

  • /grok:result [job-id] - show stored final output for a finished Grok job.

  • /grok:cancel [job-id] - cancel an active background Grok job. Without a job-id, cancels the latest active job in the current session.

  • Claude models (sonnet-5, opus-4.8, fable-5) run via the Agent/Workflow model parameter.

  • Using gpt-5.6 inside workflows and subagents by using Codex plugin for Claude Code (the model parameter only takes Claude models, so use a wrapper):

  • Spawn a thin Claude wrapper agent with 'model: 'sonnet', effort: "low"' whose prompt instructs it to write a self-contained codex prompt, run 'codex exec' via Bash, and return

  • Exception: for ordinary (non-Workflow) delegation, the codex:codex-rescue subagent IS the wrapper — spawn it directly with the task spec; no sonnet shim needed.

  • gpt-5.6 via codex:codex-rescue went 3/3 clean on clear-spec fixes when the prompt contained a written line-level spec (defect + exact fix + named tests + "commit nothing"). It also did convention-consistent spec/ledger updates unprompted. The line-by-line diff review remains mandatory — it has since caught real issues (an out-of-spec test drive-by, placeholder finding IDs).

  • Dual review (opus-4.8/fable-5 + Codex) on small PRs surfaces 1–2 real Minors per PR that the implementer missed (symlink over-refusal, cleanup no-op under the cancelling ctx). Worth the cost even for S-effort fixes; skip only for pure-docs PRs (single reviewer).

  • Reviewer/worker subagents sometimes go idle without posting their result — a SendMessage nudge reliably shakes the report loose; don't respawn.

  • Nudges to workers must be report-only ("post your report; make no further edits") — a nudged worker may run another pass and silently overwrite your fixes in its worktree. Generic check: after any delegated-worktree interaction, re-diff immediately before committing; never commit on the assumption the tree still matches your last review.

  • Parallel grok-4.5 implementation agents must use isolation: "worktree" so codex edits don't collide in the shared checkout.