Think Short — terse mode toggle

sonnet session+task

Caution

Output-token bloat kills long workflows. Preamble (“Let me…”), filler phrases (“Great question!”), unsolicited alternatives, and chain-of-thought verbalization accumulate fast across many sub-agent turns. In a brewcode run with 10+ agents, this can waste thousands of tokens per phase and obscure signal with noise.

Tip

Three profiles, two injection points. light / medium / aggressive profiles are injected once per session via SessionStart (cached, no repeat cost) and prepended fresh to every sub-agent prompt via PreToolUse:Task. Agents on the blacklist (debate, docs-writer, architect) are skipped automatically — they need full expressive output.

Quick reference

Command	Argument	Action
`/brewtools:think-short on`	`[--scope global\|project]`	Enable terse mode, default scope = project
`/brewtools:think-short off`	—	Disable terse mode
`/brewtools:think-short profile <name>`	`light \| medium \| aggressive`	Set active profile
`/brewtools:think-short status`	—	Print effective state, sources, last 10 log lines
`/brewtools:think-short blacklist add <agent>`	`<agent>` name	Add agent to injection skip-list (e.g. `docs-writer`)
`/brewtools:think-short blacklist remove <agent>`	`<agent>` name	Remove agent from injection skip-list

Default scope for all mutations: project. Pass --scope=global to write globally.

When to use

Output-heavy brewcode runs with many sub-agent spawns — reduces noise per phase
Long single sessions where preamble accumulates across compacts
Cutting verbose boilerplate from agents that over-explain before acting
Enforcing ASCII-only output and no smart-quotes in structured data pipelines

Examples

Enable with default profile (medium):

/brewtools:think-short on

Switch to aggressive:

/brewtools:think-short profile aggressive

Natural language — RU combo (resolves to on + profile aggressive):

включись максимально

Natural language — level shorthand:

уровень 3
агрессивный
макс
level 2

Check current state:

/brewtools:think-short status

Expected status output:

think-short: ENABLED (source: project-state)
profile: medium (source: project-state)
blacklist: [debate, docs-writer, architect]
state files:
  project: .claude/brewtools/think-short.json (exists, updated 2026-04-20T12:34:56Z)
  global:  ~/.claude/plugins/data/brewtools-claude-brewcode/think-short.json (missing)
plugin.json defaults: enabled=false, profile=medium
env override: THINK_SHORT_DEFAULT=(unset)
recent log:
  think-short: NL-prompt "включись максимально" → resolved as on + profile aggressive

Disable:

выключи
/brewtools:think-short off

Profiles

Profile	Directives	Approx tokens	Typical use
`light`	Be terse. Results first, reasoning only if asked.	~10 tokens	Light reduction, keeps reasoning visible
`medium`	Be terse. No preamble, no filler. Skip AI phrasings. Direct answers. No sycophancy, no unsolicited alternatives.	~35 tokens	Balanced — recommended default
`aggressive`	Be terse. ASCII only. No preamble, closing fluff, sycophancy, restatement, unsolicited alternatives, “as an AI” framing. Results first. Prefer Edit over Write.	~60 tokens	Maximum suppression — long automated runs

How it works

SessionStart — full profile injection
When a new session starts, the SessionStart hook reads the effective state. If enabled, it injects the full profile text into additionalContext. This happens once per session — the profile is cached by the model context, so no repeat token cost on subsequent turns.
PreToolUse:Task — per-agent prepend
Before each sub-agent spawn (Task tool), the PreToolUse hook prepends the first 2 lines of the active profile to the agent’s prompt. This ensures fresh injection even for agents spawned mid-session after context shifts. Agents in the blacklist are skipped entirely.
Blacklist filter
Agents named debate, docs-writer, and architect are excluded from injection by default. These agents need full expressive output — injecting brevity directives breaks their output quality. The blacklist is configurable via blacklist add/remove.
State merge — fallback chain
Effective state is computed by merging all sources in priority order. The skill reads and writes through helpers/state.mjs and helpers/safe-write.mjs (atomic write, O_NOFOLLOW, mode 0600). Project-scope state wins over global; env override wins over all.

State files

Scope	Path
Global	`$CLAUDE_PLUGIN_DATA/think-short.json` (fallback: `~/.claude/plugins/data/brewtools-claude-brewcode/think-short.json`)
Project	`<cwd>/.claude/brewtools/think-short.json`

Project scope overrides global field-by-field — enabled, profile, and blacklist can each come from different scopes.

State schema

{
  "version": 1,
  "enabled": false,
  "profile": "medium",
  "blacklist": ["debate", "docs-writer", "architect"],
  "updated_at": "2026-04-20T12:00:00.000Z"
}

Fallback chain

THINK_SHORT_DEFAULT env var (on / off / profile name) — highest priority
Project state: <cwd>/.claude/brewtools/think-short.json
Global state: $CLAUDE_PLUGIN_DATA/think-short.json
plugin.json config (config.think_short.default_enabled, config.think_short.default_profile)
Hardcoded defaults: enabled=false, profile=medium, blacklist=[debate, docs-writer, architect]

Injection timeline

SessionStart profile injection fires once and is cached in the session context — no repeat cost. PreToolUse:Task injection fires fresh on every sub-agent spawn, so agents created later in a long session always receive the current profile. If you toggle or change profile mid-session, the new state applies to the next sub-agent spawn immediately; the main-conversation context retains the earlier SessionStart injection until the next session.

Logging

Note

All state mutations and NL-prompt resolutions are logged at INFO level to .claude/brewtools.log. Filter with grep 'think-short' to see only think-short events. The status command tails the last 10 matching lines automatically.

⚙️

Updating plugins

Use /brewtools:plugin-update to check and update the brewcode plugin suite in one command. See the FAQ for details.