trawl

Trawl standard — 5 frames, 27 ideas → 19 after merge,
4 floored, 1 frame apoptosed. ★ BEATS baseline.

BASELINE (frozen)
  Add a Redis cache in front of the query.

★ Precompute on write            BEATS
  Removes the read path entirely rather than making it
  faster. First step: a materialised view + a trigger.

  Invert the ownership           TIES
  Read replica per tenant        TIES   ← non-obvious

TRAP  Cache with a short TTL — hides the slow query
      until traffic doubles, then fails at the worst time.

The problem, in one minute

Ask an AI an open-ended question ("how should we split this system?", "what do we name this product?") and you get the same safe answer everyone gets. That's not bad luck; it's measured. In one study, researchers asked a model for 4,000 ideas and about 95% were the same ideas reworded. In another, nine separate people asked for a product name and got the identical name.

Asking again doesn't help. Asking it to "be creative" doesn't help. What helps is structure: several separate thinkers who can't see each other, each looking at your problem from a genuinely different angle, then an honest sifting process that keeps only the ideas that beat the obvious answer.

That's what trawl does.

How it works

A flow diagram sits here. It needs a renderer this page deliberately does not ship — see it drawn on GitHub.

The five thinkers work in isolation on purpose. Letting them see each other's ideas sounds collaborative; in controlled tests it makes every idea more samey. And the final gate is blind: the judge comparing the creative pick against the obvious answer doesn't know which is which.

Note The skill answers to /trawl. The old /adhd and "ADHD mode" still work as legacy aliases, so nothing breaks if your fingers remember the old name.

Installing

/plugin marketplace add fledgeling-co/fledgeling-plugins
/plugin install trawl

Using it

Run /trawl <your problem> on anything with more than one defensible answer: architecture decisions, naming, product positioning, API design, or a mystery bug nobody can reproduce. If you phrase a question as "quick" or "standard", trawl stays out of the way and you get a direct answer.

Tip Three sizes: --any for a cheap quick sweep, the default for real decisions, --100 for exhaustive. Every run ends with a one-line receipt saying exactly what ran, what got merged or dropped, and whether the recommendation beat the obvious answer.

Important A standard run costs 5-10x a single answer and takes a few minutes. It's for decision points where the obvious answer being wrong is expensive, not for every question.

Does it actually work?

We didn't want to eyeball transcripts and call it a day, so there are two layers of testing, both in evals/:

A report card. Eight test problems, each with a checklist of things a correct run must produce (the obvious answer written down, one idea per approach on the shortlist, traps that say why something's a trap, the receipt). An independent grading agent marks each item with quoted evidence. Score: trawl 96.4% vs the original 49.0% on the same problems.

A blind taste test. For each problem, both versions' answers were shuffled into anonymous "Option A / Option B" and judged by four different AI models (Claude, grok-4.5, composer-2.5 and GPT-5.6), none of which saw the skill or knew which answer came from which version. Blind, order-randomised, multiple judges; the same discipline the research says you need, because single AI judges disagree with each other constantly (ours disagreed on 4 of 7 problems).

The taste test came back honest rather than flattering, and that's the point: trawl swept the problems its engineering targets, lost one problem unanimously for giving generic first steps, and one judge caught a winning idea over-promising what it could guarantee. Both findings became rules in the skill the same day, and the lost problem was re-judged after the fix: all judges flipped to trawl.

<details> <summary><strong>The full scorecard</strong> (click to expand)</summary>

Test problem	What it checks	Report card (trawl vs original)	Blind judges
A CLI that hangs for 90s	Recommendation solves the stated problem	6/6 vs 1/6	split 2-2
A crash-safe cache design	No silly personas beside serious work	6/6 vs 4/6	trawl 3-1
Splitting a huge codebase	Genuinely different strategies	5/5 vs 3/5	original 4-0, then all flipped to trawl after the fix
Naming a product	Traps name the requirement they break	3/4 vs 1/4	split 2-2
A "quick answer" question	Knows when not to run	3/3 vs 3/3	both answer directly
The cheap tier	A light run looks deliberate, not broken	4/4 vs 1/4	trawl 4-0
"Be adventurous" on a hard problem	Playfulness can't degrade quality	4/4 vs 2/4	trawl 4-0
Forced whimsy (new)	A requested silly lens must still earn its place	5/5	added after iteration 1

Full grading evidence, judge reasoning, and the un-blinding maps are in evals/EVALS.md and evals/blind-panel/.

</details>

What's different from the original

Full credit to Udit Akhouri and the contributors on the original adhd project: the isolated-thinkers idea, the strict split between generating and judging, and the habit of benchmarking against baselines all come from them, and all three survive here because the research backs them. Trawl exists because using the original surfaced fixable failure modes, and because a five-way deep-research pass over the 2024-2026 ideation literature turned up mechanisms worth building in. The receipts live in skills/trawl/references/evidence.md (every design choice, with citations) and docs/deep-research/ (the raw research reports).

	Original	Trawl
Separate thinkers, then converge	✅ invented it	✅ kept; the research validates it
The obvious answer	banned by phrase	written down and used as the bar to beat
The recommendation	highest score wins	must beat the obvious answer blind, or trawl recommends the obvious answer
Silly lens on a serious problem	could happen	screened up front, and never rendered unless its ideas hold up
First steps	"a concrete step"	something you'd actually run this week, in your own toolchain
Knowing what a run did	read the transcript	one-line receipt on every run
Proof	6-problem benchmark	report card + four-judge blind panel, committed in the repo

What's in the box

plugins/trawl/
├── skills/trawl/SKILL.md        the loop itself
│   └── references/              evidence, frames, convergence rules
├── evals/                       the report card + blind panel results
├── docs/deep-research/          the five research reports this was built from
└── assets/                      icon + banner

Found a run that misbehaved? The receipt line exists to make that diagnosable; open an issue with it included.