The "≈10 LLM calls, 5–10×" framing counts calls, not tokens, and so undercounts real cost: each divergence branch is an isolated context, so the base substrate (CLAUDE.md + tool context in a Claude Code session) is re-loaded once per branch before any novel token is generated. - when-to-use.md: replace the simple "10 calls" line with the honest formula `N × (base + branch) + critic + K × deepen`, and split out the library (small substrate) vs skill-in-agent (large substrate) regimes. - paper (docs/index.html) §Cost: same substrate-multiplied accounting; §Implementation now points to it. Drop the single hardcoded "$0.30 / $50k" figure in favor of a per-deployment computation, since the real number depends on N, substrate, and current pricing. Co-authored-by: Udit Raj <researchudit@gmail.com> Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
4.2 KiB
When to use ADHD (and when not to)
Use it for
- Architecture & design decisions (storage layer, sharding, auth model, queue topology, retry strategy)
- API / SDK / CLI surface design
- Fuzzy debugging — generate hypothesis classes you haven't considered
- Migration & refactor planning
- Naming — functions, products, services, env vars
- Code review widening — what could go wrong here, beyond the checklist
- Strategy, positioning, pricing — anywhere you'd say "give me a few ways to…"
- Inside agent loops at decision points where the cost of premature convergence is high
Don't use it for
- Lookup questions
- Bug fixes with a known root cause
- Anything where the right answer is one Google away
- Inner-loop / tight latency / per-keystroke use
- Single-correct-answer problems
One-sentence test: If a junior would Google it and find the answer, baseline wins. If a senior would say "hm, let me think about this differently for a minute" — that's the moment ADHD replaces.
Why it shines on creative and interdisciplinary work
Creative and cross-domain work is exactly the regime where premature convergence costs the most.
- The right answer is often not in any one domain's playbook — you need to transplant a mechanism. ADHD's cross-domain frames (biology, logistics, game design, markets) do this on purpose.
- The textbook answer is usually a trap — it looks right because it's familiar. ADHD's separate critic pass flags traps with named reasons, not just "could be risky."
- The interesting ideas live in the awkward middle — past the first 3, before the absurd. Single-pass generation never gets there because each token is biased by the previous one. Parallel isolated branches do.
- You don't always know what good looks like yet. ADHD's cluster pass surfaces the shape of the design space so you can argue at the angle level, not idea-by-idea.
In one line: ADHD is what to reach for the moment a single-pass agent would give you a competent, forgettable answer.
Cost & speed
Honest numbers. A default run is roughly:
- N parallel divergence calls (default 5; can be increased)
- 1 scoring call
- 1 clustering call
- K deepen calls (default 3)
That is ≈ N + K + 2 calls (≈10 at defaults). But call count is the wrong unit — token cost is what you pay, and it is dominated by context that gets re-loaded on every branch, not by the novel output.
The honest cost formula
Each divergence branch is a fresh, isolated context (that isolation is the whole point — see how it works). So the base substrate that prefixes every call — your CLAUDE.md, state files, and tool context inside a Claude Code session — is paid once per branch, before a single novel idea token is generated:
cost ≈ N × (base_context + branch_output) ← divergence
+ critic_context ← score + cluster see all N×k ideas
+ K × deepen_context ← focus passes
The N × multiplier on base_context is the part the simple "10 calls" framing hides. If the base substrate is ~26K tokens, five branches re-load ~130K tokens of substrate before any divergence — that is the real floor, and it scales with N, not with how much the model actually says.
Substrate matters more than call count
- Standalone library (
adhd "..."): minimal substrate. Each branch carries only the problem + frame prompt, sobase_contextis small and the premium is modest — close to the naive 5–10× figure. - Skill inside a Claude Code session: large substrate. Every branch re-loads
CLAUDE.md+ tool context, so the premium is meaningfully higher than the library and grows with your session's base context. Budget forN × base, not1 × base.
Rule of thumb
Frame it as: a few cents to a few dollars to widen a high-stakes decision — the exact figure depends on N, your base substrate, and current API pricing, so compute it for your own setup rather than trusting a single headline number. The mental model holds regardless: cheap relative to shipping the wrong obvious answer. Don't run it on every keystroke. Run it at decision points.