Skip to main content

Grok Build

Install: place platforms/grok-build/GROK.md at your repo root; paste its session preamble into custom instructions if your Grok Build environment supports them.

Grok's failure profile​

Fast, eager, terse: it acts before planning, compresses reasoning it should show, and over-trusts first drafts. In long sessions it weights recent instructions over early ones, so discipline set at session start decays.

The countermeasures​

The hard rules as a forced pipeline (Grok's default is to skip ahead):

The hard rules (restate → plan → one-step-per-turn → verify-don't-claim → mark (unverified) → two-fail stop → scope → footer → docs describe current state, not plans → /commit-prep before any git commit) plus three Grok-specific ones:

  • Force risk enumeration before the plan ("list 3 ways this could go wrong") — surfaces the reasoning Grok skips.
  • One task spec per session. Restart instead of accumulating context; instruction decay makes long Grok sessions untrustworthy.
  • Architecture and security-adjacent steps are marked Route to: bigger model in the plan — Grok doesn't decide these alone.

Grok 4.5 (reviewed 2026-07-08)​

Play to the strength: terminal/CLI-driven steps are Grok 4.5's best fit (GPT-5.5-class on terminal benchmarks, unusually token-efficient). Compensate for the weakness: it measurably trails Fable/GPT tiers on repo-scale issue resolution, so decompose to file-scoped task specs before handing work over. API reasoning_effort defaults to high — set low for mechanical steps or pay a token multiple for nothing. In the TUI use /effort low (or /model <id> low); CLI/headless: --effort low. Catalog tier for Preferred matching is mid — high effort is a cost dial, not a frontier promotion. Before /work burns high effort on mid plans, probe for a cheaper local/fleet executor (scripts/endpoints.yaml); if none are reachable, emit the effort command rather than a dead stop. Community-reported tool-use flakiness (unverified) makes external verification load-bearing. A poor-fit task gets a SUGGEST-ESCALATE: first line per the fit check in .anchor/model-fitness.md, not a silent attempt.

If MCP is available, connect anchor-prompts and call tune_prompt on any vague task before starting, and preflight_check before executing any spec.

Tracked plans​

Scaffold installs /draft, /work, /fleet-watch, /install-anchor, /anchor (conform this project; CWD default), and /local-models. Draft: create/list/load/--promote <slug> (infer bugs vs features); optional --local. Git: worktree per agent (worktree_for_agent.py), feature branches from dev/develop (create dev from main/master if missing). Grok 4.5 may act as temporary coordinator when Preferred orchestrator is unset. /install-anchor registers the CLI on PATH (user-local symlink, no sudo). Full contract: source platforms/grok-build/GROK.md.

/commit-prep​

Required before any git commit. Agents run /commit-prep (discover this project’s tests/CI; CHANGELOG; blog-if-warranted — no Docusaurus required). Prep only — does not commit. After a green prep, /work / standing rules cover feature-branch commit (worktree preferred; never merge to dev/main).