Gist. agent-written specs decoded, gist first, plain English after, one unskimmable question, every caveat kept

Bottom line. Your agent writes a 1,600-word spec, and you approve it after a skim. decipher rewrites any agent-written document so its first three parts carry the whole idea in plain English, in about 120 words. The rest waits below as reference.

Self-check. You approved your agent's last spec. Could you name its riskiest assumption without opening it?

Install

decipher is listed in Claude's plugin directory and on cursor.directory.

Claude Code gets the skill and an always-on output style.

/plugin marketplace add amirimi/decipher
/plugin install decipher@decipher

Codex, OpenCode, and Cursor each have their own install in INSTALL.md. So does any other agent that reads SKILL.md.

Tested with Claude Code 2.1.307, Codex 0.149.1, OpenCode 1.19.3, and the Cursor CLI 2026.10.19.

Use

Point it at a document, or say "decipher" after a reply that lost you.

/decipher:decipher specs/govern-every-model-call.md

A clone install shortens the command to /decipher, and Codex uses $decipher.

Say	Get
decipher, simplify this	The latest reply or a named document, rewritten
decipher, again on the same document	A new angle with none of the earlier sentences
walk me through a spec, PR, or report	Three to five parts, one per turn, with a bookmark
continue, back, all at once, stop	Moves through the walkthrough
I'm lost	The goal again, from a different angle

Specs and design docs are home turf. PRs, reports, incident write-ups, READMEs, and agent replies work too.

Every answer, same shape

Gist.                a few words that carry the whole idea
Bottom line.         what it does, where it stands, what backs it
Self-check.          one question you answer to yourself
-------------------  most readers stop here
Terms.               only when the jargon would lose you
Needs you.           decisions, blockers, risks
How it works.        the steps, in order
Make sure you know:  a foundation you may be missing, only when your questions show it

The Self-check is the point. It comes from the document's own premise and asks for your judgment, so scrolling back won't answer it. Can't answer it? You've stopped steering. Take the wheel back before you approve anything.

When a follow-up shows you are missing a concept the document never explains, the last line names it. Learn that first, and the document reads differently.

"Can't I just tell Claude to be concise?"

Go ahead. A concision rule shapes the next thing your agent writes. The 1,600-word spec already in your repo stays 1,600 words, and so does the PR a teammate's agent opened. decipher works on what already exists. It defines the jargon from the source, keeps what the document claims apart from what anyone verified, and checks that you still follow.

Always on

Always-on mode puts every reply in plain English, and decipher's format still waits until you ask for an explanation.

Claude Code turns it on with the plugin's output style. Select it in ~/.claude/settings.json. A plugin install names it decipher:decipher, and a clone install names it decipher.

{ "outputStyle": "decipher:decipher" }

Every other agent turns it on with this block in the instruction file it reads every session.

## decipher
Use decipher's always-on mode. Before your first reply in a session, load the decipher skill and follow its rules for always-on mode. Explain a source in decipher's format only when I ask. Load the skill again whenever its rules leave your context.

Agent	Instruction file
Codex	~/.codex/AGENTS.md
OpenCode	~/.config/opencode/AGENTS.md
Cursor	.cursor/rules/decipher.mdc in each repository, with alwaysApply: true in its front matter

While decipher's rules apply, its own limits replace brevity budgets from other output styles, CLAUDE.md, and AGENTS.md. Your accuracy and punctuation rules still win.

A real run

decipher on a 2,601-word agent-written spec for litellm-brownout, a LiteLLM plugin that switches a team to a cheaper model once its spend crosses a set line. The reply runs 354 words, and the three parts you read first run 129.

Full reply

Gist. checks extend beyond chat, Claude Code now downgraded, embeddings never swapped, unlisted calls refused, assistants endpoints escape

Bottom line. Brownout used to check and downgrade only /v1/chat/completions. As a result, Claude Code on /v1/messages and Codex on /v1/responses were never downgraded, and their unpriced models were served at $0. The spec sorts every LiteLLM call type into five kinds, checks team and key on all of them, and swaps models only for text generation. It shipped in commit 7822bbf, and the spec reports lab runs on LiteLLM 1.101.6 to 1.104.2 that downgraded all five text endpoints.

Self-check. Brownout refuses any call type that a future LiteLLM release adds until someone reviews it. Would your governed teams rather hit that 503 than have a new endpoint skip the downgrade?

Needs you.

• Decide whether to block the assistants, threads and runs endpoints inside LiteLLM, because LiteLLM runs no plugin hook there and brownout cannot see or refuse them.
• Check that every fallback deployment that uses an openai/ model implements the Responses API, because LiteLLM forwards /v1/messages there as a Responses call. A server without that API answers 404 and LiteLLM then puts the deployment on cooldown.
• Weigh that images, speech, transcription, moderation, rerank, OCR and video skip the price check, because an unpriced model on those endpoints records $0 spend.

How it works.

1. The pre-call hook looks up each call's LiteLLM call_type in one table in bridge.py and gets back one kind.
2. Every allowed call gets the team and key checks. A call that names a model also gets the model checks and the spend-tier decision.
3. Only text calls switch to the team's cheaper fallback once spend passes degrade_at. These are chat, text completion, /v1/messages, /v1/responses and Gemini generate_content.
4. Embeddings and /v1/responses/compact keep the model the caller named. Another model would produce incompatible vectors or an unreadable summary.
5. Calls that name no model, such as MCP tool calls and file reads, get only the team and key checks.
6. Provider pass-through, realtime, batches, fine-tuning and every unlisted call type get a 503 that names the call type.

Feedback

decipher left you lost? That's a bug. Open an issue with the original text and the reply.

Anything else, such as a reply that helped or a section you never read, goes in Discussions.

If you find decipher useful, please star the repo. It helps other engineers who review agent-written specs find it.

License

MIT