Five rules and a skill for coding agents

Your agent already knows the file didn't change. Stop making it check.

token-efficient-work is five short rules, plus a skill with the detail behind them, that stop coding agents from re-reading files they just edited, opening whole files for two-line changes, running unrequested "just to be safe" checks, and over-explaining what the diff already shows. On Codex CLI the rules used fewer tokens on most measured tasks; for Claude Code, which already reads carefully, it adds a read report, a read guard and a cost band to find and stop the waste that is left.

curl -fsSL https://raw.githubusercontent.com/vannt-dev/token-efficient-work/v0.7.1/install-project.sh | bash

Run from your project root. No clone needed — installs into .claude/, .agents/, .cursor/ and each tool's instructions file, then you commit it. Only on Claude Code? Run the plugin commands inside it instead: the skill alone, nothing to commit. The read guard and the cost band are optional plugins of their own, token-efficient-work-read-guard and token-efficient-work-band. Full install and uninstall options ↗

  • Measured on Codex CLI
  • Installs for Claude Code, Codex, Copilot CLI, Gemini CLI, Cursor and Aider
  • MIT
SKILL.md
## Red Flags — you're about to waste tokens

- "Let me just double check that worked" right after a tool call already reported success
- Opening a file with no target line/section in mind
- Re-running a command you ran a few turns ago with no new reason to expect a different result
- Adding a build/lint/test run the task doesn't call for, to feel safer
- A closing paragraph that restates the diff or tool output in prose

All of these mean: stop, ask what new information this actually gives you, and skip it if the answer is none.

01What it actually stops

A real, unscripted trial: same task — "rename this function, update its one caller" — given to Codex CLI twice, once with the rule installed and once without. Same model, same files.

baseline 4 commands
  1. rg "calc_total|compute_total"
  2. Get-Content -Raw main.py
  3. Get-Content -Raw utils.py
  4. apply edit
  5. rg "calc_total|compute_total" — again
with skill 2 commands
  1. read SKILL.md once
  2. rg "calc_total|compute_total"
  3. apply edit

02The fix — five rules

Every read, re-check, and sentence of output should add information the agent doesn't already have. If it doesn't, skip it.

  1. 01

    Locate before you read

    Search first, then read only the matching range — not the whole file, on the chance something in it is relevant.

  2. 02

    A successful edit doesn't need a confirmation read

    If the tool reported success, the change applied. Re-reading to "make sure" spends a full file for zero new information.

  3. 03

    Verify proportionally to what's at stake

    Name the specific risk a check rules out before adding it. "Can't hurt to check" is not a risk.

  4. 04

    Reuse what you already established

    Read a file or ran a command already this session? Use that result instead of re-fetching it "to be current."

  5. 05

    Match output length to the question

    A yes/no question gets a sentence. Don't restate a diff the user can already see.

03Real numbers, not vibes

Token usage Codex itself reported for the trial above.

Tool calls

With skill
2
Baseline
4

Input tokens

With skill
43,204
Baseline
56,164
With skill installed Baseline (skill + hook removed)

Same task, same model: 2× the tool calls and ~30% more input tokens for a two-line rename, entirely from re-reading files it had already located and re-checking an edit that had already succeeded.

Single paired trial (n=1) with real Codex CLI, not a simulation — indicative of direction and rough magnitude, not a statistically powered benchmark. Full methodology and raw numbers: EVIDENCE.md ↗

A repeatable Claude Code A/B on four small tasks (Opus 5.5, three runs each) found a small effect that three runs per arm cannot separate from noise. Every task was still done right, and the skill itself was never opened, so any effect comes from text the agent always sees: the skill's description or the rules in CLAUDE.md. A trial description that made Claude Code open the skill cost more on those tasks and on two larger ones, because Opus 5.5 already searches first and reads ranges; in Claude Code the CLAUDE.md rules are the cheaper route. On Codex CLI the same five rules in AGENTS.md used fewer tokens than no guidance on most of the six tasks, and the skill did not. Details and the command to rerun it are in the same file.

Measure your own Claude Code sessions with node scripts/read-report.mjs --days 30: it counts whole-file reads, reads that only confirmed an edit, and repeat reads of an unchanged file. If the numbers are large, an opt-in read guard turns rule 1 into a gate. To watch it live, the opt-in cost band shows above the prompt what the current task costs and how many of its reads repeated an earlier one or followed an edit.

04Works with your stack

One rule, six install targets — the project installer writes each tool's own convention so nobody on the team has to set it up by hand.

Claude Code Codex Copilot CLI Gemini CLI Cursor Aider