positioning

docs/positioning/work/
  Product truth ................ 4 shipped · 2 designed
  Research panels .............. 2 dispatched · 7 members · $20.00
  Claims ....................... 160 · 38 independent domains
  Candidate positions .......... 6 generated via trawl → 3 shortlisted

DECISION INSTRUMENT
  Veto ......................... ML matching engine excluded (unbuilt)
  Consequences ................. natural units (conversion %, CAC, sales cycle)
  Dominance .................... Territory A dominates status quo
  Rank stability ............... A leads in 44.7% of 231 weight vectors

Recommendation: "Territory A — Audit-ready reconciliation" (promising hypothesis)
→ docs/positioning/positioning-report.html  (design-review: clean)

Work out what your product should stand for, and be able to show your working.

Positioning is the sentence you lead with; the word you want to own; the enemy you name; the one customer you're for before you're for everyone. Getting it wrong is expensive and slow to notice, because a bad position doesn't fail loudly. It just quietly makes every ad, every demo and every sales call about 20% harder, for years.

Most AI positioning help gives you a confident answer with nothing underneath it. This one runs the research itself, checks the sources, and refuses to put a promise in your headline that your product can't keep today.

/plugin install positioning@fledgeling-plugins

Then: /positioning:positioning with a product to position.

What's different from the skill it replaces

This is a rebuild of positioning-pipeline by DiologIR. That skill's grounding in four positioning books, and the shape of its territory template, were good enough to keep; they're carried forward here with credit. What changed is where the quality controls live.

	positioning-pipeline	positioning
The research	Writes you two prompts and a page with copy buttons. You paste them into Gemini, wait half an hour each, and come back with the output	Runs the research itself across several AI research providers at once, then merges them
Checking the sources	Nothing checks them	Every link is resolved; anything going into a headline gets a model to read the page and confirm it actually says that
Counting agreement	Not counted	Counted in independent sources, not in how many AI tools agreed. Four tools quoting one Reddit thread is one source
Claiming things you haven't built	A written rule saying don't	A table of what actually ships, and a command that fails if a headline rests on something that doesn't
How you choose	A slider scorer that multiplies and sums	Deal-breakers first, then a plain table in real units, then which option wins across every reasonable weighting, not just the default one
How many options	Always exactly three	Generated wide, shortlisted, and your current position is always carried as one of the options to beat
The write-up	Three markdown files and one HTML page	Nine templated documents plus one designed, interactive decision page
Before you see it	No check	The page goes through a full design and accessibility review first

Why the slider scorer had to go

This is the part worth reading even if you never install this.

The old skill's centrepiece was a lovely interactive thing: seven sliders for what matters to you, move them around, watch the three options re-rank live. It feels rigorous. It's the most common way strategy tools are built.

We commissioned two research panels across seven different AI research providers to check it. All four members of the first panel independently came back with the same answer, from the academic literature: that family of tool is documented as unsafe for exactly this job. Four separate problems, each measured:

• Adding a bad option can change which good option wins. Proved in 1983, and re-derived for slider-style scoring as recently as 2023. It gets more likely when you have few options with close scores, which is every positioning shortlist ever made.
• Whoever writes the list of criteria picks the winner. Split one criterion into three sub-criteria and its total weight jumps from about 0.25 to between 0.40 and 0.48. Nobody has to move a slider for this to happen.
• The sliders start somewhere, and people don't move far from it. An experiment across five different weighting methods found all five pulled towards an even split, which flattens exactly the differences you built the tool to see.
• A high score somewhere trivial can outvote a fatal problem. "Founder excitement: 5" should never rescue an option the company physically cannot deliver.

So the scorer is gone. In its place: deal-breakers that eliminate rather than deduct, a table in real units (conversion, dollars, days, months) instead of scores out of ten, a check for options that are simply beaten on everything, a direct "what would you trade for what" question, and then, instead of a ranking, how often each option wins across every weighting you'd be willing to defend. An option that wins only at the default slider positions is an artifact of the sliders, and the page says so in those words.

It can also answer "no decision", and name the one experiment that would break the tie. A tool that always produces a winner produces winners from noise.

How it works

Seven phases. The first two happen before any money is spent, and that ordering is the whole trick: a research panel asked "what should our positioning be" gives you a survey, and one asked "here are four candidates, find what separates them" gives you a decision.

0  Product truth      what actually ships, as a table with ids
1  Candidates         generate wide under different personas, shortlist on distinctness
2  Research           decide what to buy, buy it once, verify it
3  Territories        one file each, every claim bound to an id
4  Reports            nine templated markdown documents
5  Decision page      one designed HTML surface, design-reviewed before you see it
6  Decision           the recommendation, what it costs you, and what to test

Phase 0 reads your product, not your pitch. Running code and passing tests first, the live site second, the plans third, the founder's ambition fourth. Every capability lands as shipped, designed or aspirational, with the file or URL that proves it.

Phase 1 generates candidates before the research runs, using /trawl:trawl with five positioning-shaped personas: the founder who repeats the pitch forty times a week, the buyer with no budget line for your category, your strongest competitor's head of product briefed to take your position first, a mechanism borrowed from outside software, and one deliberately odd seat.

Phase 2 decides whether to buy research at all. Four gates: is it already in your repo, would the answer change the decision, is the free lane enough, and only then is a paid panel worth it. If it commissions one, it tells you the worst-case cost before spending it.

Phase 5 builds the decision page through /design-craft:design-craft and /ux-craft:ux-craft, using your project's DESIGN.md if you have one and writing you one if you don't. GSAP where scrolling actually carries the argument. Three.js only where a strategy canvas genuinely needs a real volume, and it says so when it didn't. Diagrams are mermaid, never generated images, because an image of a chart is a chart nobody can fix. Then /design-review:design-review runs on the rendered page before it reaches you.

The two commands that make the honesty real

Everything above is prose, and prose is what the old skill already had. These are the difference:

python3 scripts/claim_ledger.py check docs/positioning/work --require-move hero ...
python3 scripts/positioning_lint.py docs/positioning --html .../positioning-report.html

claim_ledger.py check fails when a headline rests on capability that isn't shipped, when a claim's sources were never verified, or when you've called something high-confidence on fewer independent sources than that deserves.

positioning_lint.py fails on all-in-one framing anywhere in the deliverables, on a number with no source, on two "different" territories that share the same word or enemy or category or beachhead, on an owned word that's an abstraction nobody could contest, on a Blue Ocean "Eliminate" row that eliminates nothing, and on a page whose motion has no reduced-motion fallback.

Both gates were checked in both directions before shipping: 41 failures on a deliberately broken example, zero on a clean one. A gate you've only ever seen pass is a gate you haven't tested.

What it won't do

Worth knowing before you install it.

• It will not tell you a position is right. The four books make a position coherent. None of them predicts which one works, and the research is blunt about that: no named positioning framework has shown a repeatable, causal ability to pick winners. A desk-research run of this skill produces "promising hypothesis" at best, and it labels itself that way rather than saying "recommended".
• It won't hide a disagreement. Where two research providers contradicted each other, both positions stay in the evidence file. There's one in this very build: on how badly surveys overstate what people will pay, one meta-analysis says 21% and another says nearer three times. The direction is certain and the size isn't, so any pricing number carries that caveat.
• Paid research costs money. Roughly $4 to $9 per provider per panel. It always tells you first, and the free lane is genuinely useful on its own.
• It can't fix bad inputs. Better decision structure makes your dependence on guesses visible; it doesn't remove them.

Note: the old Gemini workflow is still in here as a lane, not the default. If you'd rather run Deep Research yourself and prune the plan as it goes, the prompts and the launcher page are still generated. The README for that lane lists exactly what you give up by taking it.

Does it actually work

See EVALS.md for the report card: what was tested, what the old skill scored on the same prompts, what's still untested, and the eval this one didn't win.

Credit

positioning-pipeline:positioning-pipeline by DiologIR is the predecessor. Its distillation of Ries & Trout, April Dunford, Blue Ocean and David C. Baker, its territory template, and its product-research persona are carried forward here largely intact. The research behind this rebuild is committed in docs/deep-research/, so every claim above stays checkable.