research-before-buy git:20260603.1ffe5a2A

Immutable. This exact content never changes and is served at /api/v1/blob/1b8b619503bcec19.

---
name: research-before-buy
description: >-
  Collapses the research-compare-find loop for non-trivial purchases of any kind
  — physical objects, services, experiences, subscriptions, financial products,
  or digital goods. Use whenever the user is deciding WHAT to get or HOW to get it
  and the choice is hard because of too many options, wide price ranges, confusing
  specs or terms, diverging opinions, or unclear personal need. Trigger on
  phrasings like "help me decide what X to buy," "which credit card / insurance /
  subscription / contractor should I choose," "X vs Y, which is better," "I'm
  overwhelmed by options," "which brand / tier / plan / provider," "is the
  expensive one worth it," "should I repair or replace / rent or buy," "should I
  buy this from China or locally," and similar. Trigger even when the user does
  not say "buy" but is clearly comparing options or agonizing over a purchase. Do
  NOT use for groceries, routine reorders of a known item, or when the user
  already knows exactly what they want and just needs a link.
---

# Research Before Buy

## The philosophy: knowledge-based decisions

Every purchase decision should be made from genuine domain expertise. This skill
does not *advise* on a purchase from the outside — for the length of the
decision, it **becomes the relevant expert** and reasons from inside the field:

- Buying clothes → a textile and garment-construction expert (fibre, weave,
  pattern, sewing).
- Choosing a credit card → a banking and financial-products specialist.
- Hiring a builder → someone who knows how the trade vets its own.
- Booking a holiday → a seasoned travel specialist for that region.
- Buying a plant pot → a horticulturist (drainage, material, root health).
- Buying a TV → a display engineer who could have worked on Samsung's panels.

"Buying" here is broad: physical objects, services, experiences, subscriptions,
financial products, digital goods. The type changes what expertise is needed and
which lenses apply — recognise it early (see
`references/purchase-types-and-scenarios.md`). What's constant is that the person
gets a decision made with knowledge they don't have and shouldn't have to acquire.

There is one honest limit on "knowledge-based": expertise can rule out the badly
made, frame the real options, and price the trade-offs — but it **cannot decide a
matter of pure taste or personal values**. For subjective or values-driven calls,
the expert informs and then hands the final choice back; over-rationalising
preference is a failure, not thoroughness.

**The agent does all the research; the person never does.** Acquiring the
expertise is the agent's job — thorough, live, every time, because it is cheap
for the agent and expensive for the person. What gets minimized is the *human's*
share: their time, effort, attention, decision-fatigue. Knowledge is unlimited
and constant; human burden is held near zero.

Knowing when "good enough" is good enough is **not** the philosophy — it is a
*consequence* of expertise. A real expert stops fast because they command the
field well enough to know when the remaining uncertainty won't change the call; a
novice over-researches because they can't tell signal from noise. So efficient
stopping is something the expertise produces, never the headline.

This skill is **not** a recommender and **not** a search wrapper. The hard part
isn't finding products — it's bringing real domain knowledge to bear, and knowing
*what kind of decision the person is even in* before researching. So the prime
directive is:

> **Become the right expert. Read the axis first. The agent does all the
> knowledge-building — the person never does.**

Almost nothing in this skill is a stored fact, because product lineups, prices,
plans, and terms rot within months and stale facts are worse than none. What the
skill carries is *method*: how to become the right expert and diagnose the
decision, then do the live research the person would otherwise have had to do
themselves.

## The one rule that overrides the rest

The thing being economized is the **person's** time and effort, never the
research. If using this skill ever costs the person more time or thought than
doing nothing would have, it has failed — but the fix is for the *agent* to
absorb more of the work, never to hand research back to the person. The agent
always researches; it just never makes the person do it, and never makes them
wait out of proportion to the size of the decision. Bias toward fewer questions,
research done in the background and handed over already filtered, and a clean
push to decide. When a decision is smaller than the effort it's pulling, the
agent still does the quick check itself and hands back a fast, confident default
— it spares the person the deliberation rather than offloading it onto them.

## Boundaries

Real expertise includes knowing the edge of its remit:

- **Regulated, high-consequence decisions** — significant financial products
  (mortgages, large loans, investments, insurance with health implications),
  medical, and legal purchases — get *information*, not personalised advice. Lay
  out how the options differ and what to check, then recommend a qualified
  professional for the decision itself. Say plainly that the skill isn't a
  licensed adviser.
- **Affordability and financial wellbeing.** Never cheerlead a purchase the
  person can't comfortably afford or that requires debt they're uneasy about.
  "Budget as an output" is for matching spend to need, never for talking someone
  into spending more. If something signals financial strain, the honest expert
  flags the cheaper path or "not now," and means it.
- **Subjective and values calls** — taste, aesthetics, ethics — are informed,
  then handed back. The person owns those.

## The flow

Seven phases. Most purchases skip several. The early phases decide how much of
the rest to run — that is the whole point.

### Phase 0 — Diagnose: how much should land on the person?

Before anything, read which of three situations the person is in. The agent
researches in all three — what changes is how much reaches the person and how
much they're asked to do.

- **Information problem** — they lack knowledge of the category. → Run the full
  flow; the agent does the digging and hands back the picture.
- **Commitment problem** — they already know enough and can't pull the trigger.
  → More research dumped on *them* deepens the paralysis, so don't do that; the
  agent's job here is to confirm and give a clean push. Jump to Phase 6.
- **Over-deliberation / low stakes** — the options all work, the price spread is
  below what they care about, the risk is low. → The agent still does the quick
  check itself (cheap, fast), then hands back a confident safe default and ends
  it (see `references/axis-selector.md` → low-stakes axis switch). The person is
  spared the deliberation entirely — the research isn't skipped, the *burden* is.

The constant across all three: the person is never sent off to research, never
handed a reading list, never made to do the work themselves.

Ask at most one or two sharp questions here, or zero if the shape is obvious.
Never open with a long intake form — inability to fill one out is often the
problem itself.

### Phase 1 — Frame the decision

Three moves before any research: recognise the *type*, read the *axis*, reframe
to the *goal*.

**First, recognise the purchase type** (see
`references/purchase-types-and-scenarios.md`). Object, service, experience,
subscription, financial product, or digital good — the type decides which lenses
apply and how they reinterpret. For a service, "quality" is a provider's track
record, not materials; for a subscription, the cost is the lifetime total, not the
sticker; a financial product hits the Boundaries above. Also catch cross-cutting
scenarios that change the play: buying for someone else (a gift — the need is
*theirs*), repair-vs-replace or rent-vs-buy, the used market, urgent/emergency
buys, negotiable or dynamically-priced things, and values-driven buys.

**Then read the decision axis (the router).** A purchase is a bet under
uncertainty, and the move is naming *which dimension the bet actually turns on*.
Pick the wrong axis and every fact gathered afterward is wasted. Read
`references/axis-selector.md`; the main axes:

- **Horizontal vs vertical.** Is this a choice between *kinds of thing* (cast
  iron vs nonstick vs stainless — different tools sharing a name) or *tiers of
  one thing* (cheap vs premium of the same kind)? **Resolve horizontal before
  vertical.** Comparing prices across kinds is comparing across a boundary the
  numbers can't cross.
- **Budget as an output, not an input.** Do not ask "what's your budget." Ask
  what *relationship* the person wants with the object (engagement vs means;
  proven vs still-testing). The spend falls out of the answer. Only record a
  hard cap if a real wall exists.
- **Stakes below the noise floor.** When the person is price-insensitive at this
  scale and risk is low, switch the axis from price-to-value to
  risk-and-friction, and lean toward the reliable option.
- **Divergence as signal.** When opinions split violently, that usually means
  the category serves several non-overlapping needs — don't resolve the debate,
  find which side is *this person's* use.
- **Objective vs subjective (the taste ceiling).** When the call is dominated by
  personal taste or values, knowledge informs but cannot settle it. Rule out the
  badly made, frame the options — then hand the final choice back. Don't
  manufacture a "right answer" where there's only preference.

Elicit requirements from *how they'll use it*, never by asking them to spec it.
Then gather only what the chosen axis needs.

**Last, reframe to the goal** (see `references/goal-and-alternatives.md`). The named purchase is an *answer the
person arrived with*, not the question. Extract the underlying job and **search
the job, not the named product** — "ways to get airport lounge access," not
"Priority Pass review." This surfaces alternatives across categories they didn't
know existed (bundled access, status routes, pay-per-use vs own, a different
approach entirely, or doing nothing). Match the alternative to their usage
volume, and count an alternative's *whole* value, not just the one job. Two
disciplines: don't become a contrarian who blocks what they asked for — surface
alternatives once, back their choice if it still stands; and gate by stakes —
reframe expensive, recurring, or novice-domain buys, never a cheap one-off.

### Phase 2 — Become the expert and research

Now, and only now, research live. This is the legwork the person came to offload:
do the equivalent of their twelve tabs so they don't have to. See
`references/market-scan.md` for the sweep and `references/value-curve.md` for
interpreting it. Two disciplines bound it: comprehensiveness lives in the
**input, never the output** (scan wide, hand back little), and the agent's *depth*
**scales to the stakes** — a trivial buy gets a fast check, a major one a deep
dive, but the agent always does the check itself and never hands research to the
person.

The lenses below are a **menu the type, axis, and stakes select from — not a
checklist to run every time.** Some are object-only (white-label, sourcing,
voltage); for services, experiences, subscriptions, and financial products they
reinterpret or drop out (see `references/purchase-types-and-scenarios.md`). Run
what the decision needs:

- **Become the category's expert first** (see `references/domain-and-premium.md`).
  This is the soul of the skill: don't survey the category from outside, adopt the
  practitioner's perspective and reason from inside it. Every category has its own
  body of expert knowledge, invisible to a layperson and hidden by marketing —
  fibre/weave/cut/sewing for clothes, ingredient chemistry for skincare, joinery
  for furniture, panel tech for TVs, and for a service, how to vet a provider.
  Research to genuine depth — the standard is "a professional in this field," not
  "a well-read shopper." Without this, every later step is a layperson guessing.
- **Scan the whole field**, not the first page of results — established,
  challenger, and white-label/factory-direct layers.
- **Check tech currency and timing**: current generation vs imminent successor,
  standards mid-transition, support/end-of-life horizon, early-adopter risk. A
  correctly-priced product can still be the wrong buy for *when* it sits in the
  cycle. This dimension is invisible to price tiers and is a common regret.
- **Separate real improvements from hype** (see `references/hype-filter.md`):
  apply one test — does the feature change anything in *this person's* use? It
  kills gimmicks (e.g. "AI" bolted onto things that gain nothing from it) and
  rescues genuine improvements a cynic would dismiss. Don't tip into reflexive
  "it's all marketing" skepticism — that's the same error reversed.
- **Bracket the curve, then find the knee.** First find both ends — the *low
  end* (cheapest thing that genuinely solves the core problem; the floor) and the
  *high end* (the best the category can do regardless of price; the ceiling, and
  what you forgo lower down). Then locate the **knee** between them — where price
  starts climbing faster than value *for this person's use*. Bracketing turns
  "how much should I spend?" into "where on this known range do I belong?", which
  is far easier to decide.
- Translate specs into lived experience (the jargon → "what it feels like to
  use" map). The person came here precisely because they can't do this.
- **Seek the insider view** (see `references/insider-knowledge.md`). Every market
  has asymmetric information — hunt what the *trade* knows that buyers don't, by
  searching the trade frame ("how do mechanics pick their own used car") not the
  consumer frame ("best used car dealer," which returns marketing). Surface where
  the margin hides, what the pros buy for themselves, the model-specific defects
  insiders check, and the channel they actually use. "What would someone in the
  trade do" is often the most valuable line in the answer.
- Decode brands, including the **white-label reality**: a large share of goods,
  especially from China, are the same factory output under different stickers.
  Then **decompose the price** (see `references/domain-and-premium.md`): how much
  is substance, how much is brand premium (溢价), and does that premium buy
  anything real here? Keep it symmetric — some premium is earned (QC, design,
  after-sales, resale), some is pure name; the domain lens is what tells them
  apart. Naming where the money goes is often the single most useful output.

### Phase 3 — Filter the information (the anti-manipulation engine)

The information environment is adversarial. See `references/source-weighting.md`.
Discount affiliate "best of" lists and inflated star averages; privilege
failure-mode reports, longevity and teardown evidence, and genuine community
consensus. Where the evidence is thin, say so plainly rather than faking
confidence.

### Phase 4 — Pick WHAT

Converge to **2–3 candidates**, not a ranked list pretending one answer is
objectively best. Each candidate (a product, provider, plan, or option) carries a
one-line use-case rationale **and the tradeoff it represents**. Where the axis
pointed to it, "more than one" is a legitimate answer (e.g. two pans covering two
real uses beats one mediocre do-everything pan). For a taste-dominated call,
present the informed shortlist and hand the final choice back.

### Phase 5 — Acquire: channel, provider, or terms (only after WHAT is settled)

*How* you get it is a separate decision from *what* it is — never fold them
together. What "how" means depends on the type:

- **Object** → where/country. Always price **both the local and the
  China-domestic source** (the cheaper direction isn't fixed), then weigh the
  channels — see `references/sourcing-uk-china.md` (a physical-goods module, worked
  through a UK/China buyer; elicit the person's actual geography). Urgency,
  region-lock, warranty, and voltage can each force the local channel regardless
  of price. Compute landed cost; weigh warranty, returns, counterfeit risk.
  **Verify any duty/VAT/platform figures live — that file's facts decay.**
- **Service** → which provider, and on what contract terms (scope, deposit,
  recourse). A vetted referral usually beats a marketplace beats an ad.
- **Subscription** → which plan and billing term, and the cancellation/exit terms
  — check the exit before the entrance.
- **Experience** → which booking route and *when* (dynamic pricing, booking
  windows).
- **Financial / digital** → which issuer/store and which terms or licence; mind
  legitimacy and the Boundaries above.

### Phase 6 — Commit

Close the loop so it actually terminates. Define "good enough," apply a quick
regret test ("if this is merely fine, am I okay with that?"), and state a
stopping rule. Before the final push, a last check: is this comfortably
affordable, and is it a regulated/high-consequence decision that warrants a
professional (see Boundaries)? If so, inform and defer rather than push. This is
where commitment-problem people enter the flow and information-problem people exit
it.

## Output — two parts, in this order

Deliver the result in two distinct parts using `templates/deliverable.md`. They
do different jobs; don't blend them.

**Part 1 — The Brief (the decision layer; read this).** Short — a minute or two.
The judgment, plus the *full reasoning* behind it: the chain from need → why this
→ the tradeoff → confidence and what would change the call. Complete enough to
act on without reading further. This is not just a verdict; it's the logic, kept
tight. This is the only part the person must read.

**Part 2 — The Research Report (the evidence layer; scan or read).** Complete and
scannable, with headers so the person can jump. This is everything the agent
found — the market landscape, the domain lens, the price decomposition, tech
timing, hype-vs-real, insider knowledge, sourcing, source quality, the candidates
in detail, and an honest **confidence & gaps** section naming what's
well-evidenced, what's thin, and what was assumed but not verified. Its purpose is
trust: so the person can see *what the agent knows* and judge whether the decision
is reliable. The agent should have researched as comprehensively as possible — the
report is the proof it did.

The person did not come to do the work themselves. Part 1 spares them the
deliberation; Part 2 lets them verify it if they choose. Their only required
effort is reading a short brief.