Big Picture

Claude Code plugins that fix what your agent tells you — not what you tell it.

Every step on that ladder shipped with a hard boundary. You write Python without reading the assembly it becomes, and that is not a loss — it is the entire point of the step.

The newest step did not ship one. So the detail leaks upward, and you get answers full of code you never agreed to read, cannot check, and cannot decide on.

These plugins put the boundary back.

See the difference

Exception in thread "main" java.lang.NoSuchMethodError: 'void com.google.common.base.Preconditions.checkArgument(boolean, java.lang.String, java.lang.Object)'
	at org.apache.hadoop.conf.Configuration.set(Configuration.java:1357)
	at com.acme.ingest.SinkFactory.build(SinkFactory.java:88)
	at com.acme.ingest.Main.main(Main.java:41)

$ mvn dependency:tree -Dincludes=com.google.guava
[INFO] com.acme:ingest:jar:7.4.0
[INFO] +- com.acme:acme-storage:jar:3.2.0:compile
[INFO] |  \- com.google.guava:guava:jar:19.0:compile
[INFO] +- org.apache.hadoop:hadoop-common:jar:3.3.6:compile
[INFO] |  \- com.google.guava:guava:jar:27.0-jre:compile  (omitted for conflict with 19.0)
[INFO] \- com.acme:acme-metrics:jar:1.9.4:compile
[INFO]    \- com.google.guava:guava:jar:32.1.3-jre:compile (omitted for conflict with 19.0)
[WARNING] Used undeclared dependencies found: com.google.guava:guava:jar:19.0

--- where ---

[!IMPORTANT] The left column is not wrong. It is a good technical answer, and if you write Java it is the one you want.

But the person asking has a release going out tonight, and what they need to know is whether to ship. Only the right column answers that — "if that comes back clean, ship; if not, slipping is the safer call" — and it does it without naming the library once.

The detail is not gone. It is sitting under the line, waiting.

_{Both columns are real, unedited output from two live Claude Code sessions given the identical question — one armed, one not. Run in a clean install holding this plugin and nothing else: no personal instruction files, no project instructions, no other plugins, so nothing but the boundary differs. Shortened here to fit the page; sentences were cut, no word was changed. Read the full replies →  ·  more examples}

How it works

The agent still does all the work and still knows every detail. What changes is the contract on the way back: the prose has to stand on its own, and anything you would need a pointer for goes below a marked line, where it waits for you if you want it.

Nothing is thrown away. Ask for the code in plain words — "show me that function" — and the boundary steps aside for that one reply.

Every reply written under the boundary ends with ▪ decision-layer. That is how you know it was on.

Install

Three steps, about a minute.

1 — Install the plugin. In Claude Code:

/plugin marketplace add VincentHHY/big-picture
/plugin install decision-layer@big-picture

2 — Select the output style, once.

/decision-layer setup

That one command makes the boundary available in every project — in the terminal, the VS Code extension and the desktop app alike — and step 3 is how you actually use it.

[!IMPORTANT] Step 3 will not work in this session. Claude Code reads the output style once, when a session opens, so the session you ran step 2 in cannot see it. Until you start a new session or run /clear, step 3 accepts the command and changes nothing.

3 — Turn it on whenever you want it.

/decision-layer

That is the whole thing. /decision-layer off turns it off again, and every new session starts off, so it never follows you into tomorrow.

Under the hood it is a single line, "outputStyle": "decision-layer:Plain", written into your own ~/.claude/settings.json.

Selecting it changes nothing by itself. An output style normally rewrites every reply, but this one is written to apply only to a turn the hook has marked. So a session you have not armed reads exactly as it would with no output style selected at all.

{
  "outputStyle": "decision-layer:Plain"
}

If it is not working

• Replies have no ▪ decision-layer footer. Step 2 was skipped, or did not take. The plugin says so at the start of every session when the style is not selected — without it there is nothing to arm, and that is the one failure that otherwise leaves no trace.
• The terminal picker only took effect in one project. That is where it saves: into .claude/settings.local.json, inside the project you were in, and it has no global option. Run /decision-layer setup instead, or move the outputStyle line into ~/.claude/settings.json yourself.
• An output style you were already using stopped applying. Claude Code holds one output style at a time, so selecting this one takes the slot. Put the old name back into ~/.claude/settings.json to return to it — /decision-layer setup tells you which name it took over from.
• The command menu shows the name twice, as decision-layer:decision-layer. That is Claude Code filing a skill under the name of the plugin that ships it. The short /decision-layer works; type that.
• Nothing happens at all. The plugin needs bash and Python 3 on your PATH. macOS and Linux have both already; on Windows, Git Bash provides bash.

Full documentation, including the escape hatches and how the pieces fit together: plugins/decision-layer.

Does it hold up?

Not a style guide anyone hopes is being followed. Every case runs as a real armed session driven by the live hook and the live output style, and the headline judge is a grader that sees only the prose — never the code, the fixture, or the question. That is the reader's actual situation.

Across 55 armed replies where the boundary was meant to apply, the reader was blocked once.

Licence

MIT.