---
name: blind-spots
description: Surface consequential decisions a plan, design, or research brief left silently assumed, by interviewing the user in dependency order. Use when the user says "blind spots", "grill me", "stress-test this plan", "poke holes in this", "what am I missing", "interrogate this design", "what haven't I decided", or needs a loose idea or research brief clarified before work begins. Do NOT use for a clear factual lookup, to explain a previous reply in more detail (use `huh`), to generate ideas when it is unclear what to build (use `ideation`), or to review an already-written plan with agents (use `review-plan`).
argument-hint: '[plan, design, idea, or research brief to probe; blank = take it from the conversation]'
effort: high
---

# Blind Spots

Understand the intended outcome, challenge the framing, then settle the decisions
the user does not know they have left open.

A plan or research brief can feel complete because its gaps were filled in silently.
Make those decisions visible before they steer implementation or research.

## Instructions

### 1. Resolve the target

The argument is what to probe. If blank, take it from the conversation. If no target
is in play, ask for one sentence on what they want to build, understand, or decide.

### 2. Recon before asking anything

An uninformed question wastes a round and spends the user's patience. Before round one,
read the available context: the conversation and relevant supplied documents; for repo
work, the source, `AGENTS.md` / `CLAUDE.md`, and relevant `docs/`. Reuse recon and settled
decisions supplied by a calling workflow. For an unfamiliar topic, a short direct lookup
can establish the terminology or options needed to ask useful questions.

**Finding facts is your job, never the user's.** Anything discoverable from the
filesystem, the git history, or a tool is yours to look up. Ask the user for intent,
experience, priorities, and decisions that the available context cannot establish.
Open empirical questions remain research tasks; the user need not settle
them before research begins. For example, ask whether offline use is required, then
research which candidates support it.

### 3. Build the decision tree, then prune it

**First, check the target is one piece of work.** A brief describing several
independent subsystems needs splitting before it needs questions — refining the
details of something that should be three projects spends a round on the wrong
thing. Say so, propose the split and the order, and probe only the first piece.

**Establish intent before narrowing.** Briefly reflect the intended outcome, who it
serves, and what success means; distinguish the user's statements from your assumptions.
Where those are unclear, ask for a concrete situation in which this would help, or what
would make a technically correct result disappointing. Follow the answer before asking
about implementation details. Reuse concrete examples already supplied.

**Challenge the framing once.** Consider a simpler way to achieve the outcome, a
materially different mechanism, and whether a presumed constraint is actually required.
Use a relevant lens from `ideation` when useful: inversion (what would make this fail?),
analogy (how does another domain solve this?), or removing one assumed constraint.
Generate candidates before judging them. Surface only credible alternatives that could
change the outcome, with trade-offs and a recommendation; do not invent options to fill
a quota. Respect explicit constraints and settled choices unless new evidence challenges
their rationale. Keep this within the target; an open-ended search for what to build
belongs in `ideation`.

Then map the work as decisions, each branching into the decisions that hang off it.
Sweep the categories below rather than working from intuition — a whole category is
easier to miss than a question inside one:

- **Building something** — purpose, constraints, success criteria; then the shape:
  component boundaries and interfaces, where state lives and how data moves, failure
  behaviour, and what counts as tested.
- **Researching something** — purpose, population, region, timeframe, and the outcome
  being compared. A request to learn about a topic need not serve an adoption decision.

Most categories come back already settled, by the context or by a default nobody
would argue with. The sweep exists to catch the ones that do not.

Pruning is what separates a useful round from an interrogation:

- **Keep** a decision that changes the work — different answers lead to different
  user outcomes, workflow, structure, interfaces, candidate pools, evidence to gather,
  or success criteria. A valuable discovery can be cheap to implement; reversibility
  alone is not a reason to discard it.
- **Cut** scope the brief carries that its stated purpose does not require, and ask
  whether it is wanted instead of designing around it. Unrequested scope is a fork
  the user cannot see, and the cheapest one to close.
- **Drop** a decision that is cheap to reverse and constrains nothing downstream.
  Naming, log wording, which of two equivalent helpers to use: decide those yourself
  during the work.

An unspecified detail is not automatically a blind spot. A fork the user cannot see is.

### 4. Ask the frontier

The **frontier** is every kept decision whose prerequisites are already settled — the
questions answerable *now* without guessing at answers you have not heard yet.

Batch independent decisions in one message, ordered by consequence, then **stop and
wait**. Give a question about intent, a surprising answer, or a consequential tension
its own round when the answer could change which other questions are worth asking.

```
❓ **Q1 — <short title>**: <the question, with the options if it is a choice>

➡️ <recommended answer, and the reason in a clause>

---

❓ **Q2 — <short title>**: …

➡️ …
```

Rules for a round:

- **Every decision question carries your recommendation.** "Go with your defaults"
  must be a complete answer. For exploratory questions about the user's experience
  or intent, offer a tentative interpretation to correct, not an invented answer.
- **An asked question stays pending until the user answers or explicitly delegates
  the choice.** Silence, elapsed time, an empty tool result, and a preselected
  recommendation settle nothing. An asynchronous prompt returning only confirms
  delivery; wait for the user's reply. Continue independent read-only work while
  waiting, then yield if none remains. These are required decisions, not optional
  preferences that may fall back to defaults on timeout.
- **A question whose answer depends on another open question belongs to a later
  round**, not this one. This is the whole mechanism: question 12 does not exist as a
  question until question 4 is answered.
- **More than about seven questions means the tree is split too fine.** Re-prune per
  step 3 rather than sending a wall.
- Never block on a lookup. A running exploration is an unsettled prerequisite, so only
  the questions downstream of it wait — ask the rest of the frontier now.
- Where the harness offers a structured choice prompt, use it for the discrete
  either/or questions and prose for the rest.

### 5. Recompute and repeat

Each set of answers settles decisions, pushing the frontier outward and unblocking
questions that depended on them. When an answer reveals a surprising need, conflicting
priorities, or a promising alternative, follow it with a concrete example or trade-off
question before returning to the checklist. Update the framing and decision tree when
that changes the goal. Continue only while the next question could change the outcome
or a consequential decision; a question count is not evidence of depth.

### 6. Close

Stop when the intended outcome is understood, the framing has been challenged,
credible alternatives have been considered, and every consequential decision is
settled. Unknown facts may remain. Existing context can satisfy these checks.

**When called by another workflow:** return a compact brief with the purpose, scope,
settled constraints and priorities, alternatives considered and why the chosen approach
fits, and remaining factual questions. If the context already satisfies the closing
checks, return immediately without an interview. The caller
owns any existing confirmation or approval step; do not add a separate closing
confirmation or hand off to another workflow.

**When invoked directly:** state the settled understanding back — the decisions and
their answers, compactly, in the user's own terms — and ask the user to confirm it
matches what they meant.

On confirmation, hand off rather than sprawling into implementation: `plan` for a
medium task, `longshot` if they are handing it over and leaving, a heavier planning
workflow for an architectural change. For a research brief, return it for the requested
research. Write the understanding to a file only if asked.

## Boundaries

- **Do not build during the interrogation.** No edits, no scaffolding, no mutating
  commands, no implementation subagents. Direct invocation awaits confirmation of the
  closing summary; embedded use follows the caller's existing approval requirements.
  Reading and searching are expected.
- Cap parallel lookups at **three read-only subagents**, one question each.
- Do not answer a decision on the user's behalf to shorten the session. Recommend, then
  wait.
- Do not re-ask a decision already settled, in this session or in the docs.

## Examples

### A loose feature idea

`blind-spots add team workspaces` — recon finds single-tenant auth and no authorization
layer. Round 1 asks the four forks that follow: does a user belong to many workspaces,
is existing data migrated or grandfathered, are invites by email or link, does billing
move to the workspace. Round 2 exists only because round 1 answered "many workspaces" —
role model and per-workspace permission checks are now answerable. Round 3 is empty.
The summary lists nine settled decisions, three of which the user had not realized were
open.

### A plan the user already wrote

`blind-spots` after pasting a migration plan — recon reads the current schema, so the
round asks only about forks the plan left implicit (backfill under load, rollback
window, dual-write duration), never about what the plan already states.

### A coherent brief with the wrong emphasis

`blind-spots add a weekly activity dashboard` — ask what decision the dashboard would
help someone make. The user says missed handoffs are the real problem. Compare the
dashboard with an exception digest and notifications at the handoff itself. Their
answer reveals that interruptions are costly, so probe when someone can act before
choosing a digest. Then settle ownership and escalation. The exploration changes the
solution without adding features or requiring an architectural rewrite.

### A research brief

`research-general research home batteries` — context leaves the purpose unclear.
Ask whether the user wants an explanation or a purchase comparison, recommending an
overview if they are exploring. Only a purchase comparison opens questions about the
installation region and whether backup power or bill savings matters most. Prices and
available products remain research questions. Return the brief to `research-general`.

### A caller already settled the scope

`evaluate-tech` delegates research on export support for a named candidate with the
job and constraints supplied. Research that capability directly. A worker discovering
a scope blocker returns it to the caller; it does not start another interview.

## Troubleshooting

### The user answers "you decide" to everything

Take it at face value: record your recommendation as the decision and move on. If it
happens across a whole round, step 3's pruning was too loose — those were not forks.
Tighten it and shorten the next round.

### A question turns out to depend on an answer already given

It belonged to an earlier round and was missed. Drop it, apply the known answer, carry
on. Do not make the user repeat themselves.

### The frontier keeps growing and the session will not end

An exploratory round may widen the options; use the intended outcome to narrow them
afterward. If rounds keep adding unrelated scope, propose splitting the work into
bounded pieces and probe only the first piece.

### The interview closes quickly but the design feels obvious

Check whether you accepted the proposed solution as the goal. Return to a concrete
use case, test one assumed constraint, and compare a credible alternative. Ask only
about differences that could matter to the user; do not pad the interview.

### The user says a question is trivia

Believe them and drop it. It failed step 3's fork test.
