Scout Targets

A scout maps terrain; it does not take it. This skill is read-only end to end — it nominates and ranks targets for an autoresearch loop, and never edits the code or runs the loop. Because discovery touches nothing, the lanes fan out in full parallel with no collision.

The conserved quantity is the receipt: the exact command that turns a candidate into a number. A nomination without a receipt is a vibe — drop it.

Procedure

1. Frame the hunt. Name the two corpora in concrete paths and sources: the codebase (services, queries, routes, jobs, build) and the work logs (past conversations, PLAN/EXPERIMENTS notes, Substack drafts, issue threads). State the search is read-only.

Done when: both corpora are named with reachable paths/sources and the read-only boundary is stated.

1. Fan out lanes. Launch one independent read-only lane per candidate family. Each lane hunts only its family and returns nominations carrying a receipt. Lanes never write and never read each other's findings. Two axes of families:

• Runtime — query latency (EXPLAIN ANALYZE on a frozen snapshot), endpoint p50/p99 under fixed load, bundle size / lab Core Web Vitals, type coverage (tsc), test/mutation score, cost-per-request, background-job duration, cache hit-rate, error rate, throughput.
• Artifact — prompts, skills, agent-instruction files, templates with schema checks, and retrieval pipelines (precision / recall / latency). These are where the suite optimizes its own tooling against an eval set; they are often the highest-fit autoresearch targets and the easiest to overlook while hunting only runtime numbers.

Add a family only if a lane can name its receipt. Stop rule: if a lane needs another lane's finding to score a candidate, return the dependency — do not braid the lanes. A target whose verifier depends on a second target is two targets or none. Done when: every family lane has returned and every nomination carries a receipt.

1. Synthesize. Dedupe across lanes onto one flat list. Delete every receiptless nomination without ceremony.

Done when: one list, each entry with a distinct receipt.

1. Gate each candidate. Apply the seven gates and the Goodhart pre-mortem from the gates skill. Assign green / amber / red and, for anything not green, name the first gate it failed.

Done when: every candidate has a tier; every non-green names its failed gate and (if amber) its promotion.

1. Rank and return the ledger. Sort by tier, then by payoff × tractability. For each amber, name the one promotion that would green it — usually a calibrated judge-panel or a held-out harness. Flag the top green candidates for frame-loop.

Done when: the ledger is returned in the contract below and the next action for each candidate is explicit.

Output contract

Return one block per candidate, highest tier first. Omit no field; a missing receipt is a missing candidate.

Handle:      
Family:      
Receipt:     
Current:     
True goal:   
Tier:        green | amber | red
Failed gate: 
Guard:       
Promotion:   
Disposition: frame now | build harness | judge-panel | keep human | discard

Close with the handoff: name the green candidates to send to frame-loop, and the amber candidates whose promotion is a judge-panel.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: TylerJNewman
• Source: TylerJNewman/skills
• License: MIT

Install and usage instructions live in the source repository linked above.