simplify-code Skill Review: 14,696 Bytes Sending Four Reviewers at Your Diff

The skill's whole design argument fits in seven words: "Four narrow reviewers beat one broad reviewer." That line sits in a 14,696-byte SKILL.md inside NousResearch/hermes-agent, a repository with 256,820 stars, 56,176 forks and 211 SKILL.md files in my clone. The stars belong to the agent, not to this skill. Public discussion of the skill is thin: one GitHub issue, withdrawn by its author, and no Hacker News story.

I read the file from a shallow clone (commit eb1b4c5, 2026-10-24) and ran none of it. The skill directory is one markdown file, no scripts. Nothing went through any API; every token figure below is file bytes divided by 4.

What the skill actually says

Three phases. Phase 1 captures git diff (working tree first, then staged, with variants for "the last commit" or "this branch"). Phase 2 hands the complete diff to four reviewers in one delegate_task batch call. Phase 3 dedupes, drops false positives, and applies fixes by risk tier. Two lines carry the attitude:

Findings come back as file:line → problem → cost → fix | confidence | risk. The mandatory cost field is the best idea in the file: a finding that can't say what it costs is probably a nit. It also says to run only when asked ("Do NOT auto-run this after every edit") and supports dry-run, focus and scope modifiers. Four commits touch the path since 2026-06-08; the 2026-07-24 one added the fourth (altitude) reviewer and an inline fallback, and its message says it tracks the current Claude Code design.

What it costs you in context

Arithmetic from file size, divisor 4 bytes per token, rounded per row. Real tokenizer counts will differ; nothing here was measured in a live session.

The file is the small part. Each of the four children gets the full diff, runs its own repository searches, and bills separately; the repo's delegation config warns that "each child consumes API tokens independently." The skill itself says to warn the user above 2,000 changed lines. I can't give a per-run figure and didn't try.

What the community reports

The last two rows concern Claude Code's /simplify, the tool this skill was modelled on, not the Hermes port. They split: one commenter finds it a net win in a multi-tool pipeline, the other finds it marginal. The 3 comments mentioning skills in the 52-point Hermes launch thread (item 48419000) don't discuss this one. Verdict on the port itself: little public discussion.

The withdrawn issue still describes the file correctly: "review my code" is a trigger, Phase 3 edits files, and the opt-out is a phrase you must add. The reporter decided that's intended autonomy. Fair call, but know it before typing "review my code".

How to install it

Verified against the README at the commit I read. The skill ships in the bundled skills/ folder, and the repo's prompt-test fixtures list it in the default skill index, so there is no separate skill install step.

# Linux, macOS, WSL2
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash

# Windows (PowerShell)
iex (irm https://hermes-agent.nousresearch.com/install.ps1)

# then
source ~/.zshrc    # or ~/.bashrc
hermes
# inside Hermes: /skills, or /simplify-code

I haven't run the installer or read install.sh; piping a remote script to bash is your call. On Claude Code, skip this file: Anthropic ships its own /simplify, and this one names Hermes tools (delegate_task, search_files).

Where it earns its keep, and where it doesn't

It earns its keep on a few-hundred-line diff in an interactive Hermes session, just before a PR. The SAFE/CAREFUL/RISKY split is concrete, the git blame check guards against deleting code that only looks dead, and the inline fallback tells the agent to disclose when it ran as a single pass.

Now the criticism.

• The headline claim is unmeasured. No file under evals/ mentions simplify-code. Support for "four beat one" is an assertion in issue #379 plus the fact that Anthropic landed on the same design. Agreement on a design isn't evidence it works.
• "Four parallel" isn't universal. The skill says four fits the concurrency budget "on any default install". The default is 10 (tools/delegate_tool_config.py), but for finite one-shot -q sessions oneshot_max_children defaults to 2, and _oneshot_spawn_budget in delegate_tool.py refuses a 4-task batch. I read that code and didn't run it; the skill's inline fallback should catch it, which means headless runs get one sequential pass.
• Trigger and edit policy collide (see above).
• Cost scales with your diff. Four copies of the diff plus four search sessions; the only guard is advisory prose at 2,000 lines.

Skip it on Claude Code, for a read-only review (say "just report"), for diffs in the thousands of lines, or in headless Hermes runs.

FAQ

What does the simplify-code skill do?

It captures your git diff, sends the complete diff to four parallel reviewers (reuse, quality, efficiency, altitude), merges their findings, and applies fixes by risk tier: SAFE edits automatically, CAREFUL edits with test verification, RISKY ones only flagged. SKILL.md is 14,696 bytes, MIT licensed.

How much context does simplify-code use?

SKILL.md is 14,696 bytes, so bytes ÷ 4 gives an estimated 3,674 tokens. That is arithmetic from file size, not a measured session figure. The real bill is larger, because each of the four sub-agents receives the full diff and searches the repository on its own.

Is simplify-code the same as Claude Code's /simplify?

Not the same code. The frontmatter says it was inspired by Claude Code's /simplify, and Anthropic's docs describe the same four angles: reuse, simplification, efficiency and abstraction level. The Hermes version adds SAFE/CAREFUL/RISKY tiers and dry-run, focus and scope modifiers, and it names Hermes tools such as delegate_task.

How do I install simplify-code?

It ships in the bundled skills folder of Hermes Agent. Install Hermes with curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash on Linux, macOS or WSL2, start it with hermes, and invoke the skill with /simplify-code. I did not run the installer.

Skill by Nous Research and Hermes Agent contributors, MIT licensed per the repository LICENSE (Copyright 2025 Nous Research) and the SKILL.md frontmatter. Source: github.com/NousResearch/hermes-agent. Quoted lines are short excerpts for review purposes. Reviewed October 25, 2026 from source, GitHub data, Anthropic's public docs and public Hacker News comments; nothing was run through any API. Community reception shifts over time, so check the issue tracker for current status.