Most AI-generated plans look good on paper but fall apart during execution. They contain vague steps, hidden assumptions, weak verification, and risks that were never properly addressed.
The result is rework, missed scope, and eroded trust in the planning process.
grokplan produces plans whose tracker rows are built to resist poor execution.
It does this with two mechanisms:
- A Row Quality Standard (including explicit anti-gaming rules) that the planner must satisfy.
- A mechanical execution contract in
launch.mdthat forces re-reading of the current row, literal quoting of text being reinterpreted, anti-backfill discipline, and periodic contract re-engagement.
The output is a plan and tracker with built-in defenses against scope drift, fake verification, and gradual rationalization. It is stronger than default planning, but it is not magic. The final result still depends on how rigorously the planner applied the standard and how strictly the implementer follows the contract.
Every serious run goes through four phases with hard gates:
RECON → COMPOSE → STRESS-TEST → GO / NO-GO
- RECON: Deep exploration and active falsification.
- COMPOSE: First honest synthesis of the plan and tracker.
- STRESS-TEST: Adversarial review from multiple perspectives.
- GO / NO-GO: Mechanical rubric and zero-tolerance ratchet.
Nothing moves forward until it meets the exit criteria for that phase.
One public example is included: adding an export command to an existing CLI data processing tool.
The difference is explicit risks, machine-checkable verifications, and anti-drift mechanisms.
The public payload contains only what a new user needs:
grokplan/: The complete skill (the main deliverable).README.md,LICENSE,CHANGELOG.md, and supporting files.examples/: Real before/after comparisons.
Everything else (research, internal notes, pre-release artifacts, experimental Companion TUI source) is excluded via .gitignore and lives under experimental/ or other gitignored directories.
Install through the Grok plugin system when available.
# Global
mkdir -p ~/.grok/skills/grokplan
cp -a grokplan/* ~/.grok/skills/grokplan/
# Project-local
ln -s /absolute/path/to/grokplan ./skills/grokplanThe Companion TUI is currently in production. It will not be required to get the improved planning experience or the launch.md handoff. The core value works with plain Grok and file tools. See companion/README.md before attempting to use it.
Companion TUI is to show the updated grokplan, because it does not automatically pop up like the regular plan.md from grok build, and to show a live tracker that gets checked off, as there is currently no access to manipulate the TUI.
Once installed, invoke it naturally:
- "Plan this with grokplan at Medium depth"
- "Use grokplan on the auth refactor before we hand it to /implement"
/grokplan "your task" at Deep depth
It will ask for depth and run the full adversarial protocol.
grokplan produces a complete execution package:
plan.mdtracker.mdrubric-report.md- Perspective critiques
- Research trace and decision log
All artifacts are designed to survive long implementation runs.
Bug reports with reproduction steps are welcome. Pull requests that improve the protocol or documentation will be reviewed.
MIT License. See LICENSE.