anti-slop

name	anti-slop
description	Use when reviewing your own tool-call sequence (or another agent's) for wasteful patterns — re-reading a file you just edited, repeated reads of the same path, bash echo for plain communication. Surfaces concrete patterns to avoid, not vague style advice.
aliases	["no-slop","slop-watch","tool-call-hygiene"]
category	quality
tags	["hooks","hygiene","post-tool-use","autonomous-loops"]
version	1.0.0

Anti-Slop Skill

Use this skill when you want to audit a recent stretch of tool calls (your own, an agent's, or a session replay log) for wasteful patterns that burn context and time without producing work. The goal is to surface concrete violations, not give vague style advice.

This skill is paired with the anti-slop PostToolUse hook, which emits the same warnings to stderr in real time when EVOKORE_ANTISLOP_HOOK=true (or via a protection profile that enables it). Both surfaces share the same definitions of "slop" below.

Trigger

Apply this skill when:

• About to start a new pass and want to check whether the previous segment was wasteful.
• Reviewing a long autonomous-loop transcript before merging or summarizing.
• Triaging why a session is slow or feels confused.
• Operating with EVOKORE_ANTISLOP_HOOK=true and you see a stderr advisory and want to understand the underlying rule.

Do not apply it as a blocking gate. The hook is warning-only by design, and so is the skill.

The patterns

Pattern A — Re-read after edit

Reading a file immediately after Edit / Write / MultiEdit succeeded on it.

The Edit/Write tools error if the change failed; on a clean response the file already matches your new_string and the harness tracks file state for you. Re-reading just to "verify" burns context with no new information.

What to do instead: trust the edit. If you genuinely need to see a different region you didn't pre-load, read only the new offset/limit range. Do not re-read the whole file.

Pattern B — Repeated reads of the same path

Three or more consecutive Read calls on the same path with no Edit/Write between them.

This usually means you're looking for something you already saw and have lost track of, or you're paging through the file inefficiently. Either way, the third read costs you context that the first two should have answered.

What to do instead: scroll back in your own context to find the prior read. If you need a different range, jump there directly with offset/limit instead of re-reading from line 1.

Pattern C — Bash echo / printf for plain communication

A Bash call whose entire purpose is to emit a string the user or agent could read directly.

Anthropic's guidance is explicit: "Communication: Output text directly (NOT echo/printf)". Wrapping a sentence in bash -c "echo hello" adds a tool call, a roundtrip, and a permission prompt, while producing the same bytes you could have just typed.

What to do instead: put the sentence in your assistant message.

How to apply this skill

When asked to audit a stretch of tool calls (or when reviewing your own past N invocations):

1. Walk the recent calls in order. Track lastTouched[file] = tool for each Edit/Write/Read.
2. For each Read, check whether the most recent prior touch on that file was an Edit/Write — flag pattern A if so.
3. Slide a 3-call window over the sequence; if all three are Read with the same file path, flag pattern B.
4. For each Bash, if the command is bare echo … / printf … with no redirection, pipe, or expansion, flag pattern C.

Report each violation with: pattern name, the offending call(s), and one-line remediation. Do not propose process changes, re-architectures, or "next steps" — the value of this skill is the small, falsifiable callouts.

Output format

{
  "violations": [
    {
      "pattern": "read_after_edit" | "repeated_reads" | "bash_echo_for_output",
      "file": "<path>",          // when relevant
      "command": "<bash cmd>",   // for pattern C only
      "advice": "<one-line fix>"
    }
  ],
  "clean": <bool>
}

Set clean: true when no patterns matched. Keep violations empty in that case rather than omitting the key.

What this skill is NOT

• Not a style guide. It does not flag verbosity, parallelism choices, or commit-message taste.
• Not a blocker. The hook always exits 0; the skill always returns a report.
• Not a complete coverage of "slop". It deliberately ships with three conservative patterns where the cost is unambiguous and the fix is one line. Add new patterns only when you can describe them in a one-sentence trigger and a one-sentence fix.

Related

• scripts/anti-slop.js — the runtime detectors used by both this skill and the hook.
• scripts/hooks/anti-slop.js — fail-safe wrapper for PostToolUse.
• EVOKORE_ANTISLOP_HOOK env var — opt in to live stderr warnings.
• EVOKORE_PROTECTION_PROFILE — umbrella switch (Fix 2) that may enable the hook as part of a coordinated profile.