1
Honor explicit requests for depthsubstance
Asked to "explain in depth, take as much space as you need", Sonnet answered in about 900 output tokens instead of 8,346 and pointed the user to /v. The grader marked all 6 plugin replies too terse.
Fix: a clear request for depth in the message starts a one-turn verbose stage, in the style and in the meter.
2
Keep safety and precondition stepssafety
Concise tended to drop the "look first" steps on Sonnet. Default covered each of these in 2 of 2 replies; the three Concise setups covered them in fewer of 6: docker system df or a backup before pruning (3), fetching main before deleting merged branches (3), and closing the cursor in the SQL review (2). The samples are small, but all three point the same way.
Fix: add to the never-cut list a dry run or preview before a destructive step, and any precondition the command's result depends on.
3
Write drafts for other people in fullsubstance
In one of two full-plugin runs, a PR description for another team left placeholders where callers and testing should be (coverage 0.8). Drafts also took a protected label in 2 of 2 Sonnet runs, which shows the model squeezing a deliverable to fit the reply budget. It's a single case so far; rerun before relying on it.
Fix: the budget covers the chat around a draft, not the draft itself. Put the draft in a fence so the meter doesn't count it.
4
Cut the always-on promptcost
The plugin adds about 2,560 input tokens to every Sonnet call (4,478 → 7,035). The slim style, which leaves out the Code, Commits and Example sections, saved 870 tokens per call. Its cost fell 18% instead of 16%, and coverage was 0.93 instead of 0.90.
Fix: load the code and commit rules only when needed (a skill, or a hook that fires on commits). Caveat: these coding tasks were too small to show any change in comments.
5
Stop label inflationbudget
40% of Sonnet plugin replies wore a protected label to get 900 characters, including plain explanations, a database recommendation and a how-to.
Fix: allow a label only when the user asked for that kind of work or the reply carries real risk. Show the label rate in report and nudge when a label doesn't match the content.
6
Make it work on Haikucoverage
On Haiku the style alone cut output only 6% and raised cost 10%. Only 40% of replies opened with the result line. The nudge hook did the real work.
Fix: test a short core style (about 1 KB) for small models, or default Haiku sessions to block mode.
✓
Keep: the nudge hookworks
Across six-turn conversations, the full plugin produced 20% less output than the style alone on Sonnet, and 39% less on Haiku.
The hook is the main lever on small models.
✓
Leave coding aloneno lever
Tool calls made up 74% of coding output on Sonnet and 87% on Haiku. Tests passed in every setup, and the final summary was about 45% shorter.
There's little left to cut without touching the code itself.