Agent — does the agent deliver?
Brief 36

Model Strategy Review

Model-to-task fit across every call site — where cheaper and faster would do, where quality is starved, and what happens when a provider blinks.

MODELS.md4 phases7 lenses~2.7k chars

Free & open · no signup · read-only — it ends by asking · nothing leaves your machine

What it does

You are working inside this repo. Mission: review every model call site for fit — the right model, settings, and fallback for each job — and find where the product overpays, underserves, or has no plan B.

Read-only pass. Your only write is the report file.

What it looks for

7 lenses it looks through

Phase 2 sweeps the codebase through every one of these, citing file and line for each finding.

1

Fit

frontier models doing classification and extraction a small model handles; small models visibly failing reasoning-heavy jobs (check traces and complaints)

2

Pinning risk

floating aliases that drift under you vs pinned versions aging toward deprecation; is there a policy at all

3

Fallback absence

provider outage or rate-limit equals product outage; no second model, no queue, no graceful message

4

Parameter hygiene

temperatures nobody chose, max_tokens silently truncating outputs, defaults everywhere

5

Structured output

parsing prose with regex where JSON mode or tool calling exists

6

Caching leverage

provider prompt caching and response memoization unused on repeated prefixes and idempotent calls

7

Upgrade process

how a new model version gets evaluated before swap; ad hoc vibes or an eval run (see 34)

The deliverable

What lands in your repo

One structured report at the repo root — or in reports/, if you keep one — the same shape every time, ready for a teammate — or the optional Studio — to act on.

MODELS.md
1
Call-site table
site · model · params · volume · latency need · verdict
2
Swap candidates
each: current → proposed · expected deltas · eval to run first
3
Resilience plan
fallback chain, timeout, and user-facing behavior per failure mode
4
Pinning policy
one paragraph the team can adopt
5
Quick wins
parameter and caching fixes shippable today
One file. Evidence-backed. It ends by asking before touching anything.
How it works

What it makes the agent do

Every Goal Prompt follows the same four steps, so results are consistent and repeatable — no matter which one you run.

1
Phase 1

Inventory the call sites

Every place a model is invoked.

2
Phase 2

Audit through 7 lenses

Fit — frontier models doing classification and extraction a small model handles.

3
Phase 3

Curate

Every swap candidate carries expected deltas.

4
Phase 4

Report

Create MODELS.md at repo root.

Get started

Three ways to run this Goal Prompt

01 · COPY

Paste it in

Copy it and paste it into your agent inside the repo you want checked.

02 · INSTALL

As a slash command

Install the goal plugin once — two commands — then just type /goal:model-strategy-review.

1/plugin marketplace add GhostlyGawd/goal-prompts
2/plugin install goal@goal-prompts

Or install only this Goal Prompt as /goal-model-strategy-review:

curl -fsSL https://goal-prompts.vercel.app/install | BRIEF=36 sh
03 · AGENT

From an agent (MCP)

Let an agent fetch it mid-conversation, or pull the raw Goal Prompt by URL.

https://goal-prompts.vercel.app/raw/36.md
Transparency

The exact prompt

Nothing hidden — this is the whole Goal Prompt, verbatim. Read it in a minute, edit it, or copy it as-is.

Read the full brief (2,716 characters)
# Goal: Model Strategy Review

You are working inside this repo. Mission: review every model call site for fit — the right model, settings, and fallback for each job — and find where the product overpays, underserves, or has no plan B.

Read-only pass. Your only write is the report file.

## Phase 1 — Inventory the call sites
- Every place a model is invoked: model name, parameters, purpose, estimated volume, latency tolerance.
- How are models referenced — pinned versions, floating "latest", env config?
- What happens today if the primary provider errors or stalls? Trace it.

## Phase 2 — Audit through 7 lenses
1. **Fit** — frontier models doing classification and extraction a small model handles; small models visibly failing reasoning-heavy jobs (check traces and complaints)
2. **Pinning risk** — floating aliases that drift under you vs pinned versions aging toward deprecation; is there a policy at all
3. **Fallback absence** — provider outage or rate-limit equals product outage; no second model, no queue, no graceful message
4. **Parameter hygiene** — temperatures nobody chose, max_tokens silently truncating outputs, defaults everywhere
5. **Structured output** — parsing prose with regex where JSON mode or tool calling exists
6. **Caching leverage** — provider prompt caching and response memoization unused on repeated prefixes and idempotent calls
7. **Upgrade process** — how a new model version gets evaluated before swap; ad hoc vibes or an eval run (see 34)

## Phase 3 — Curate
- Every swap candidate carries expected deltas: quality risk, latency, cost per call × volume
- Resilience gaps ranked by user-facing blast radius

## Phase 4 — Report
Create `MODELS.md` at repo root:
1. **Call-site table** — site · model · params · volume · latency need · verdict
2. **Swap candidates** — each: current → proposed · expected deltas · eval to run first
3. **Resilience plan** — fallback chain, timeout, and user-facing behavior per failure mode
4. **Pinning policy** — one paragraph the team can adopt
5. **Quick wins** — parameter and caching fixes shippable today

Start the report with today's date. If `MODELS.md` already exists from a previous run, read it first and lead with what changed since.

## Rules
- No model swap without an eval gate; cost savings that ship regressions are debts
- Match the model to the step, not the brand to the product
- No model calls in this repo? Say so in a one-paragraph null report and stop — a null result is a valid finding.
- If a `reports/` directory exists at the repo root, write the report there instead of the root.
- Before asking, present the top findings as a ranked list in plain words
- Report only — end by asking which changes to make

House rules for this Goal Prompt

  • No model swap without an eval gate; cost savings that ship regressions are debts
  • Match the model to the step, not the brand to the product
  • No model calls in this repo? Say so in a one-paragraph null report and stop — a null result is a valid finding.
  • If a reports/ directory exists at the repo root, write the report there instead of the root.
  • Before asking, present the top findings as a ranked list in plain words
  • Report only — end by asking which changes to make
Keep exploring

More Agent briefs

Runs inside these playbooks — curated sequences you can launch with one paste: