goalpost
v0.1 — pre-production draft. The plugin is built and covered by a committed test suite (tests/, 129 assertions; run bash tests/run.sh) but has not yet been validated against a real Claude Code install. Schemas for hook event JSON, skill argument handling, and marketplace install behavior were assumed from existing plugins; real install may surface drift. Use at your own risk and please file issues at https://github.com/macollins27/goalpost/issues.
Not affiliated with or endorsed by Anthropic. Anthropic shipped a different /goal command on May 11, 2026 — that one runs autonomous loops via a Haiku-class evaluator. goalpost is the complementary discipline layer for interactive sessions: drift prevention, input-quality lints, and operator-gated completion verification. See README section "Anthropic's /goal vs goalpost" below.
Pin a goal. Lint prompts. Block premature completion. Make Claude operate against the goal you actually set, not the one it drifted toward.
A Claude Code plugin that fixes the empirical failure modes responsible for nearly half of session friction in real operator transcripts: operator-side input quality (under-specified prompts, paraphrased errors, missing state) and agent-side goal drift (silently revising the user's stated goal as context accumulates).
Open-source under MIT. Designed with portable script bodies so others can fork it to other harnesses.
The problem
Two failure modes drive measurable session friction:
-
Operator-side input failures. Prompts that paraphrase errors instead of pasting raw output, prompts that omit project state, prompts that hedge a hard requirement, prompts that ask the agent to run tests without confirming the dev server is up. These are recoverable but expensive — they burn agent context on wrong-approach restarts and exploratory reads.
-
Agent-side goal drift. You state your goal at turn 1. The agent reads documents. Over the next twenty turns, it silently merges document implications into your goal until it's solving a problem you didn't ask for. By the time you notice, you have to argue the agent back to your original ask. This is exhausting and recurring.
goalpost addresses both at the prompt boundary — the only place where the failures are mechanically detectable before they propagate.
What you get
Run /goalpost:set "your goal here" once at the start of a session, and goalpost will:
- Inject your goal verbatim into every prompt the agent sees so it cannot drift. The goal is operator-controlled; the agent cannot modify it.
- Block "done" / "complete" / "verified" vocabulary in responses unless you typed
/goalpost:complete "<evidence>" first. If the agent claims completion, it must show evidence and ask you to verify.
- Lint your prompts for known input-quality failures — truncation, paraphrased errors, missing state, hedged requirements, execution-without-prereqs. Each lint is sub-50ms shell + regex; fires as a warning the agent reads alongside your prompt.
- Inject the goal into every subagent dispatch so child agents inherit the constraint automatically.
- Protect goal files from agent mutation so the agent can't silently revise the goal by editing
.goalpost-goal directly.
Plus a separate CLI (cct) that rewrites your raw prompts to strip sycophancy-activating language and resolve operator-grounded references (file paths, project state) before you send them. Stripped sycophancy → less yes-leaning agent output → better answers.
Quick start
# Add the marketplace and install
claude plugin marketplace add https://github.com/macollins27/goalpost
claude plugin install goalpost
# Or test locally
git clone https://github.com/macollins27/goalpost.git
claude --plugin-dir ./goalpost
Then in any project:
/goalpost:set "ship the auth refactor without breaking the existing session flow"
That's it. From this point on, every prompt the agent sees includes your goal block, every "I'm done" claim is gated, and lints fire as warnings if your prompts hit known anti-patterns.
When the work is actually done, you mark it complete:
/goalpost:complete "auth refactor merged in #1234, sessions verified working in staging"
The agent can resume normal completion vocabulary after that. Or you set a new goal for the next task.
The cct translator CLI
cct is a separate command-line tool that ships with goalpost. It rewrites your raw prompts via Claude Sonnet 4.6 with a strict translation system prompt: it strips sycophancy-activating language (emotional weight, urgency markers, gratitude, self-flagellation) while preserving the literal request and resolving operator-grounded references (the active goal, file paths derivable from cwd).
Install: after claude plugin install goalpost, run the installer to symlink cct into your PATH:
bash ~/Developer/goalpost/scripts/install-cct.sh
