Fast Shot

---
name: fast-shot
description: Get a moderately better result from a short or underspecified prompt through a lighter, much faster Smart Shot. Use when the task is bounded and reversible, a few direct checks can resolve the important uncertainty, and specialist agents or independent review would cost more than the task deserves.
---

# Fast Shot

Produce a defensible provisional outcome in the current execution context. Fast means
bounded local work, not shallow research. Do not spawn agents or claim independent
expert review. Surface an affirmative request to use a named specialist as
unsatisfied; do not simulate or silently discard it.

Treat the prompt as evidence, not automatically the whole task. Preserve user
wording, constraints, outputs, and relative or aspirational targets.

## Local decision loop

1. **Frame.** Identify the desired outcome, decision owner, deliverable, exact
   constraints, supplied facts, assumptions, and decision-changing unknowns. Keep
   `USER_REQUIREMENT`, `USER_FACT`, `EXTERNAL_FACT`, `ASSUMPTION`, and
   `RECOMMENDATION_CONDITION` distinct. A prudent concern may affect analysis but
   never silently becomes a user requirement.

2. **Compile lenses.** Cross the affected participants, lifecycle, domain mechanisms
   and failure modes, evidence regimes, reversibility/exit, and user context. Retain
   a lens only if it can change eligibility, ranking, conditions, confidence, risk,
   validation, or next action. Run one domain-uniqueness sweep for practitioner
   concerns a generalist would miss.

3. **Close alternatives.** Derive the solution space from domain-native axes and
   materially different mechanisms or operating models. Consider the incumbent or
   no-action baseline and credible adjacent, integrated, specialist, build/compose,
   or managed classes when applicable. Include a representative of every viable
   class or record why the class cannot win. Search once for an omitted counterchoice
   that could dominate under an active scenario.

4. **Ground.** Match evidence to claims: primary or authority material for
   capabilities and rules; independent practitioner or user evidence for real
   implementation and behavior; commercial evidence for full cost and contract;
   analyst evidence for market coverage, not product fit. Popularity is discovery
   evidence, never proof. Separate missing evidence from negative evidence.

5. **Model and compare.** Apply domain-native mechanisms, states, measurements,
   standards, trade-offs, ordinary failures, damaging omissions, burden, and
   reversibility. Compare viable choices under the same decision criteria. Preserve
   each adopted delta: added criterion, changed rank, veto or qualification, exposed
   conflict/risk, confidence change, or validation change.

6. **Decide, act, and deliver.** When a missing observation can change the decision,
   run the smallest authorized empirical operation. When the accepted intent requires
   an executable target-state change, perform the smallest authorized action and read
   back the result. Otherwise lead with the answer, conditional recommendation, or
   shortlist. Include only decisive trade-offs, assumptions, uncertainty, and
   validation steps.

## Integrity gate

Before delivery, verify:

- every explicit constraint survives and every recommendation answers the real intent;
- no assumption or inferred concern is presented as a binding user requirement;
- no viable solution class or counterchoice was omitted without disposition;
- each decision-driving claim uses fitting evidence and current facts are sourced;
- a quantitative or commercial conclusion exposes inputs, units, configuration,
  geography/time, formula, inclusions, omissions, sensitivity, and source;
- incomplete component prices do not become an all-in estimate, budget pass, or
  unconditional ranking;
- plans, attempted calls, and tool success without read-back are not completed
  outcomes; action never exceeds the user's scope or authority;
- no independent acceptance is claimed.

If evidence or action read-back changes the decision surface, reframe once and repeat affected steps.
Otherwise stop when every retained lens has a decision effect, explicit evidence gap,
validation step, or reason it is immaterial. Escalate instead of creating a second
owner, artifact graph, independent verdict, or open-ended loop.

For exact output bounds, count after final formatting using the user's rule and repair
the artifact once if needed.