AmplifierSkills and tools for thoughtful work

Smart tool

Fast Decisions

Find the agent decisions worth making cheaper.

For developers tuning agent cost and latency. Inspect routing, try bounded advisory choices and compare recorded runs with task outcomes, so an attractive estimate can be checked against what actually executed.

Use Fast DecisionsWatch the 67-second filmWalk through an example
What you getFictional illustration
Question: should this bounded read use a fast path?
Offered candidates: README.md, docs/architecture.md
Returned choice: one offered ID, or abstain
Host checks: offered ID, current state, native permission

Comparison receipt
Predicted cost difference: estimate only
Executed model calls: collect from both runs
Task outcome: check independently
Observed saving: compute only after those records exist

Fictional routing and measurement outline. It contains no measured speed, cost or task-quality result.

See the work behind the bill

Inspect recorded decisions, model calls and receipts before choosing an optimization.

Keep small choices bounded

Offer eligible candidates and let the host validate the returned choice or accept an abstention.

Measure the tradeoff

Compare matched baseline and enabled runs, including quality, cache effects and cases that cost more.

The film

The cheaper ride.

In 67 seconds, a cheaper model begs for every job, then needs four trips where the usual one needs one. Fast Decisions makes the call once, before the session starts: it checks the price and the project, prints its reasons, and keeps large projects on your usual model. A prediction is not a bill, so you compare matched runs. Captions are on the player; a transcript is below.

Download the film 15-second cut Captions
Read the transcript

Every session starts with a choice. Your usual model, or a faster, cheaper one.

Cheaper looks obvious.

But a cheaper model can make more trips to finish the same job.

Fast Decisions makes that call once, before the session starts.

It checks the price. It checks the project. Then it prints its reasons.

A project with more than three hundred files stays on your usual model.

A prediction is not a bill. So you run both, on the same task, and compare.

Your own matched runs answer that. Nothing here is a promise.

On screen: Check the bill, not the price tag.

Fast Decisions. Find the agent decisions worth making cheaper.

The tasks and prices are a fictional illustration. Savings on your own work are not measured; the film shows how the tool behaves and reports no results.

See the work take shape.

Fictional routing and measurement outline. It contains no measured speed, cost or task-quality result.

Inspect your setup

Use deterministic doctor and offline diagnosis before configuring an optional scoring service.

Fictional example

Check a proposed fast path

The brief Use Fast Decisions to inspect the installed CLI and run its deterministic doctor. Explain what is available without a scoring provider. Propose one bounded read/list choice and a matched comparison that would test whether it helps; do not enable a hook or start a model service.

Before

A router claims savings from a suggested cheaper choice.

After applying the method

Matched receipts show which calls actually executed and whether the task passed.

Advisory scores alone cannot establish execution or cost avoided.

A clear job.
A useful stopping point.

Select a read/list candidate, diagnose routing, or compare observed execution.

Use it when

  • You have agent traces and want to identify expensive repeated decisions.
  • A small read/list choice has explicit eligible candidates.
  • You can compare enabled and baseline runs on the same task.

Choose another route when

  • A guaranteed speed or savings improvement.
  • Letting a routing suggestion bypass the host’s approval checks.

Inspect

A router or model service may be misconfigured.

Native manifest, diagnose and observed execution counts.

Select

The caller supplies bounded eligible candidates.

A native advisory choice or explicit abstention.

Compare

Matched baseline/enabled records exist.

Recomputable outcome, cost and latency differences.

Read the canonical method

Fast Decisions

Use the actual amplifier-fast-decisions CLI or amplifier_fast_decisions library. This is a Smart Tool with deterministic manifest/describe/diagnose/measure/compare operations; it is not a request to let a prompt execute arbitrary tools. Start with amplifier-fast-decisions --help and <capability> --help for the installed version.

Start a session: use doctor for a deterministic setup check, then decide once before a session begins. The shipped rule decider uses the native scope and price gates without a model call. A predicted saving is a prediction, not an observed invoice result. launch --harness claude|codex|copilot --dry-run shows the proposed startup command without starting that host; actual launch starts a process only within the user's authorized task. Avoid mid-session model switches. Opt-in judges require the documented credential and external-state consent before task text leaves the machine.

Inspect: run manifest, describe and diagnose --offline when a live service probe is unnecessary. Keep optional Jev/Laya/local providers separate from deterministic installation. A missing scorer is a setup finding, not a zero-quality result.

Select: supply only eligible relative read/list candidates and a bounded task. Selection is advisory; the host validates the returned offered ID, state freshness and its own approval boundary before executing anything. An abstention returns control to normal reasoning. Search may read bounded eligible sources; CUA only proposes an action.

Compare: use measure and compare on recorded matched runs with explicit task outcomes. Distinguish predicted savings from observed executions and invoiced cost. Local scores are not calibrated correctness. Read references/PRIVACY.md before enabling external scoring; existing user authorization governs whether particular data may leave the machine.

The retained native orchestrator, policies, receipts, observatory, benchmark records and CLI aliases remain available. Do not activate a host hook or stop an existing viewer as a side effect of installing the portable skill. Goal numbers in historical source records are targets, not results for this consolidation.

Limits and prerequisites
  • No speed, quality or savings improvement is claimed by packaging or offline replay.

Your first useful run.

Inspect one trace, choose one narrow intervention and compare it with an unchanged baseline.

The download contains the complete skill folder. Extract it into the skills directory supported by your agent, keeping its references and license together. Or use the source installer below. Python 3.11+ is required for that installer.

Download the portable skillSource on GitHub

Start from the public source

git clone https://github.com/michaeljabbour/amplifier-smart-tool-fast-decisions.git
cd amplifier-smart-tool-fast-decisions
python3 scripts/capability.py install --dest ~/.agents/skills
Native tools and optional execution
python3 -m pip install .

These commands are for the optional native surface. Follow its installation guide and prerequisites; portable method installation is separate.

Keep the complete skill folder together, including its references and license. The host applies the method in your context; the skill does not start a background service.

  • Python 3.11 or later is required by the portable install/build helper. The installed method runs through your agent host.
  • The native command is amplifier-fast-decisions; the Python import is amplifier_fast_decisions.
  • Install the portable skill alone for guidance, or install the native package in a virtual environment for the CLI.
  • Deterministic inspection works without a model key. Optional Jev, Laya or local scoring needs its own setup and data-sharing decision.

Try this with your agent

Use Fast Decisions to inspect the installed CLI and run its deterministic doctor. Explain what is available without a scoring provider. Propose one bounded read/list choice and a matched comparison that would test whether it helps; do not enable a hook or start a model service.

Use after installing in a host that loads skill folders. Supply your own task files and review the result.

Keep the guide nearby

Read the guide · License · Changes · Release metadata · CONSOLIDATION

Know what you can rely on.

The reconciled native suite and clean-wheel deterministic rule decision passed. This release does not establish a speed, savings or quality improvement on another person’s workload.

Read the checks and their scope
  • Functional acceptance — passed

    1650 native tests passed, 70 skips before metadata rename; 18 Smart Tool regression checks passed after rename.

  • Native host acceptance — not-run

    No live model, real desktop transaction, invitation, deployment or production data migration was performed by this functional baseline.

Source and conformance checks are separate from demonstrated benefit. Comparative effectiveness and outside-user outcomes are not established by this site.

Fictional examples explain the capability. Tests check the stated behavior. Outside-user outcomes need their own evidence.