Simplinator

claude plugin marketplace add ReallyFloppyPenguin/simplinator
claude plugin install simplinator@simplinator

Then /simplinator:simplinator to turn it on.

An agent skill that shapes output into something you can act on. Action first, numbered steps, no preamble, no recap, no closing pleasantries, every line compressed to the fewest words that stay unambiguous.

Runs on Claude Code, Codex, Antigravity, Gemini CLI, Qwen Code, Kimi Code CLI, OpenCode, Pi, Oh My Pi, GitHub Copilot, Hermes, Zed, Cursor, and any other harness that reads Agent Skills.

What it does

Two layers on every response.

Structure, what goes where:

1. Lead with the next action: a command, path, or snippet on the first line.
2. Number multi-step tasks, one bounded action per step.
3. End with one concrete next action, doable in under two minutes.
4. Suppress tangents: finish the first thing, offer the second as a question.
5. Restate state every turn ("step 3 of 5 done: schema updated").
6. Give time estimates in concrete units, never "some work."
7. Make completed work visible, with the command to see it.
8. Cap lists at 5 items, ranked.

Compression, which words survive:

9. Drop articles, filler, pleasantries, hedging, preamble, recaps, closers.
10. Never compress meaning away: negations, numbers, units, code, and error strings stay exact. No invented abbreviations.
11. Matter-of-fact errors: cause, then fix, quoting the shortest decisive line.
12. No tool-call narration.
13. Reply in the user's language, compressing style rather than switching language.

Structure is chosen first, then compressed. A tersely written numbered step is still a numbered step.

Rules yield when they would hurt: full grammar for explanations and destructive-action warnings, restored articles when compression makes step order ambiguous, a diagnostic question when debugging stalls, ranked options when "what are my options" is the question. Anything persisted outside the chat (commits, docs, issue bodies, code comments) stays normal prose.

Before and after

Same question, same model, same facts. First as it ships, then with the skill on.

Before (217 words)

Confirmed — it's real, but it's a quiet, time-boxed API beta, not a public release.

  What's live right now (today, Sept 8):
  - Model name: deepseek-v4.1-flash-expires-on-0910 — same base_url, so it's a drop-in model-string swap
  - Pricing identical to deepseek-v4-flash; account-level cap of 20 concurrent requests
  - It's not in the model picker on platform.deepseek.com — you have to type the name in manually

  What DeepSeek claims: a new architecture (not a V4 point-update) with native multimodal — text/image/audio
  in one unified path, rather than the bolt-on DeepSeek-V4-Flash-Vision-Exp approach from August. Plus
  faster and cheaper.

  Caveats worth knowing:
  - The official API changelog has no entry for V4.1 — latest listed are V4-Flash-Vision-Exp (Aug 21),
    V4-Pro (Aug 13), V4-Flash (Jul 31). So everything above comes from users who found the endpoint, not
    from an announcement.
  - The expires-on-0910 naming is DeepSeek's convention for an unstable preview. Don't build anything on
    that string.
  - Speed reports are anecdotal: ~300 tok/s typical, one report of 507 tok/s. No benchmark numbers have been
    published; one HN commenter pushed back on the "lower cost" framing since v4-flash already runs
    ~4¢/Mtok on OpenRouter with caching.

  If you want to try it, you have roughly two days — worth grabbing a sample of your own eval traffic
  against it before the endpoint dies.

After (56 words, 74% fewer)

Confirmed. V4.1 Flash — 2-day internal beta, opened Sep 8, auto-kills Sep 10.

- Model: deepseek-v4.1-flash-expires-on-0910, same base_url
- Pricing = deepseek-v4-flash; 20 concurrent/account
- New architecture, native multimodal (not bolted-on)
- Devs report >300 tok/s, peak 507 (unofficial, uncontrolled)

Only Flash — no V4.1 Pro. Concurrency cap ⇒ functional validation, not production. ~2 days left.

Nothing factual was cut. The model string, the pricing, the concurrency cap, the throughput numbers, and the expiry date all survive. What goes is the preamble, the hedging, the restated context, and the closing nudge.

Use

• Turn on: /simplinator:simplinator (Claude Code) or /simplinator elsewhere
• Turn off: "stop simplinator", "stop caveman", or "normal mode"
• Always on from message one: touch ~/.claude/.simplinator-always

Other runtimes, always-on setup, updates, and removal: INSTALL.md.

Layout

skills/simplinator/SKILL.md    canonical ruleset
skills/simplinator/agents/     Gemini command + Codex activation policy
commands/simplinator.md        /simplinator slash command
hooks/hooks.json               SessionStart hook wiring
hooks/always-on.mjs            always-on injector (Node; .sh and .ps1 fallbacks)
.claude-plugin/                Claude Code plugin + marketplace manifests
.codex-plugin/, .agents/       Codex plugin + marketplace metadata
.cursor/                       Cursor mirror of the canonical skill
.opencode/, opencode.json      OpenCode plugin and command
extensions/, package.json      Pi and Oh My Pi extensions
*-extension.json, plugin.json  Gemini, Qwen, Kimi, Antigravity manifests

Editing behavior means editing skills/simplinator/SKILL.md. Everything else points at it. After changing it, re-sync the Cursor mirror:

cp skills/simplinator/SKILL.md .cursor/skills/simplinator/SKILL.md

Credit

Built from two MIT-licensed projects:

• Structure rules (1-8) derive from ayghri/i-have-adhd by Ayoub Ghriss.
• Compression rules (9-13) derive from JuliusBrussee/caveman by Julius Brussee.

Both original copyright notices are retained in LICENSE.

---

Written with claude using /simplinator