Brevity is not the product.
The fast-rising i-have-adhd repository packages an output style for coding assistants: lead with the next action, number multi-step work, suppress tangents, make completed work visible, and end with one concrete next step. Its README demonstrates the change as a before/after interface, not a smarter underlying model.[1]
The useful idea is bigger than a diagnosis and narrower than “write less.” A five-line answer can still be useless if line one is throat-clearing, the command is buried at line four, and the failure handoff says only “let me know.” Compression without sequencing is just smaller fog.
Do not tune the word count. Tune the handoff.
Context has two consumers.
Anthropic’s Claude Code guide says the context window contains the conversation, read files, and command output; it recommends managing that context aggressively as it fills. The same guide also says to give the agent a check it can run—tests, a build, a linter, a fixture diff, or a browser screenshot—so “looks done” is not the only stopping signal.[3]
That is model-side attention. The operator has an attention window too. A dump of every command can bury the one failing assertion. A summary with no receipt forces blind trust. Good output allocates both windows: enough state to steer, enough evidence to verify, no scenic route through the agent’s autobiography.
Optionality belongs behind a gate.
Hardik Pandya’s “Tyranny of optionality” is about career attention, not coding agents. Its relevant metaphor is that constantly evaluating alternate paths consumes the attention needed to inhabit the current one.[2] Treat this as an analogy, not evidence about interface performance.
Agent output has the same surface failure: four alternate implementations, three “you may also want to” detours, and a new framework suggestion while the requested test is still red. Exploration is valuable when exploration is the gear. During repair, optionality is wheelspin.
Install a gearbox, not a muzzle.
- Action: lead with the command, path, or decision. Keep the first move executable.
- Build: report changed surfaces, exact verification, remaining risk, and one handoff.
- Explain: preserve mechanism and causality; use headings so detail stays navigable.
- Explore: rank a small option set, name trade-offs, and recommend one route.
- Carry four interlocks: next action, current state, proof, and tangent brake.
This does not mean every response should be tiny. It means long answers must earn their length, and short answers must still preserve the controls. The goal is not silence. It is traction.