Your coding agent did the work. Plain English makes the answer usable.
One open Agent Skill for Cursor, Claude Code, Codex, and Grok. It simplifies language, not evidence.
Install · See the difference · Why it is different · What's new in v3
npx skills add Tradelord223/plain-english -gThe installer finds the plain-english skill and lets you choose the agents that
should use it. The skill is Markdown: no runtime, API key, account, or telemetry.
Then ask your agent:
Explain what changed in plain English. Keep the exact file names and validation results.
Prefer a manual install? Use the guide for Cursor, Claude Code, Codex, or Grok Build.
Without Plain English
In
auth.ts, the authentication middleware implementation now facilitates the rejection of expired JWTs prior to downstream request handling.
With Plain English
auth.tsnow rejects expired JWTs before the app handles the request.
Same fact. Fewer obstacles.
The skill also changes the explanation for the reader without changing the underlying behavior:
| Reader | Explanation |
|---|---|
| Backend engineer | The endpoint is idempotent: the idempotency key returns the first payment response instead of creating a second charge. |
| Store owner | Retrying with the same request key returns the first result instead of charging the customer twice. |
v3 treats plain English as meaning-preserving compression. It makes an answer shorter or easier to read only after identifying the details that must survive.
- Truth anchors: outcomes, evidence, certainty, obligations, scope, conditions, numbers, risks, blockers, and exact tokens stay intact.
- Operational-state safety: planned, attempted, completed, tested, verified, live, and customer-visible are no longer treated as interchangeable.
- Mixed-audience layers: a practical outcome comes first, followed by the exact mechanism or evidence specialists need.
- Compression floor: a short answer cannot hide a failed check, uncertainty, material risk, or required action.
- Precision reference: difficult claims route to a focused preservation contract instead of making every ordinary invocation longer.
- Human voice without theater: answers match the reader and a supplied writing sample, cut templated AI habits, and never add fake quirks or promise to beat detectors.
- Broader behavior contract: twenty adversarial cases now cover state, quantities, negation, causation, mixed audiences, voice matching, templated structure, and unsupported next steps.
This is a major release because both the behavior contract and package layout changed. Supporting references now ship and synchronize with the canonical skill on every platform.
Most “plain language” prompts are word-swap lists. That is dangerous in technical work because the easy synonym is often wrong.
| Blind replacement | What Plain English does instead |
|---|---|
serialize → “turn into JSON” |
Names JSON only when the format is known to be JSON |
deploy → “put live” |
Names the real target: local, staging, or production |
idempotent → “safe to run twice” |
Explains the actual retry and duplicate-effect guarantee |
should → must |
Preserves whether something is advice or a requirement |
| “likely cause” → “cause” | Keeps uncertainty when the evidence is incomplete |
Plain English simplifies the explanation, not the evidence.
It also treats “human” as a communication problem, not a banned-word list. The agent can keep a writer's directness, contractions, rhythm, stance, and honest uncertainty while removing throat-clearing, inflated significance, forced groups of three, repeated conclusions, and mechanical formatting. It does not add typos, random fragments, fake opinions, or detector guarantees.
The human-voice layer was informed by the public pattern work in blader/humanizer, then re-authored for technical agent communication. Plain English keeps exact tokens, uncertainty, requirements, validation limits, and operational state even when a more casual rewrite would sound smoother.
- Change summaries that lead with the user-visible result
- Status updates that separate done, in progress, and blocked
- Error reports that distinguish a confirmed cause from a likely one
- Handoffs that keep exact files, commands, decisions, and open work
- How/why answers that match the reader's technical background
- Walkthroughs with clear actions, checkpoints, and recovery steps
It does not rewrite code, commands, raw logs, or load-bearing technical specifications. It does not invent facts, hide failed checks, or remove caveats to make an answer shorter.
Before sending a user-facing answer, the skill makes the agent:
- Extract the truth anchors that must survive compression.
- Infer whether the reader is expert, mixed, nontechnical, or unknown.
- Choose the smallest response shape that fits the job.
- Explain behavior instead of applying a jargon dictionary.
- Match the room and remove response-template habits without inventing a personality.
- Compare the draft with the source and reject any semantic promotion.
The canonical instructions live in
skills/plain-english/SKILL.md.
- One source of truth: platform copies are generated from one canonical skill.
- Cross-platform drift checks: CI verifies Cursor, Claude Code, Codex, and Grok skill files and supporting references stay synchronized.
- Precision contract: a focused reference protects operational state, modality, scope, conditions, quantities, causation, and high-stakes boundaries.
- Human-voice guide: a focused reference covers voice matching, natural rhythm, concrete writing, templated structures, and detector-gaming boundaries.
- Behavior contract: twenty adversarial cases cover the main workflow, material edges, mixed audiences, human voice, and raw-artifact boundaries.
- Open source: MIT licensed, with no service dependency.
Read the behavior cases, precision contract, human-voice guide, or changelog.
| Agent | Skill path |
|---|---|
| Cursor | skills/plain-english/ |
| Claude Code | .claude/skills/plain-english/ |
| OpenAI Codex | .agents/skills/plain-english/ |
| Grok Build | .grok/skills/plain-english/ |
The open Agent Skills installer can install the canonical skill into supported hosts. Platform copies remain in the repository for native and manual installs.
If Plain English makes one handoff, error report, or code explanation easier to use, star the repository. Stars help other builders find useful open Agent Skills.
Issues and focused pull requests are welcome. See
CONTRIBUTING.md before changing the skill contract or
platform copies.
MIT © 2026 Raqz