Parallel work without prompt soup

One lead. One owner. One next test.

An evidence-first team model for parallel security research without duplicated probing, vague handoffs, or imaginary chains.

These files do not create or launch agents. They are inspectable instruction files for a coordinator and its workers. Preview them, merge useful rules with your project, and keep narrower controls intact.

Surface, client, identity, and server owners feed one lead ledger. A skeptic checks confounders before chain analysis and independent validation.
Inspect before using

Two files, two jobs.

AGENTS.md

Project operating rules

Shared ownership, scope, evidence, and handoff rules for the repository.

Place it in
repository root or a narrower project directory
Starting rosters

Match the team to the question.

More agents are useful only when ownership boundaries are real. Start with the smallest roster that can collect, challenge, and validate the claim.

Small surface or one promising lead.

Solo hunter plus skeptic

Roles: Surface mapper + Evidence skeptic + Report validator

Handoff: Mapper records one hypothesis. Skeptic attacks its cheapest confounder. Validator reruns only if it survives.

Stop: Stop when the lead is rejected, safely supported, or needs authority you do not have.

A modern web app with large bundles and undocumented operations.

JavaScript to API

Roles: Client analyst + Surface mapper + Identity analyst + Evidence skeptic

Handoff: Client analyst supplies operation, source, sink, and message IDs. Identity analyst tests one controlled boundary.

Stop: Stop client analysis when every high-ranked lead has a discriminator and owner.

OAuth, recovery, tenant, role, or object authorization work.

Identity boundary

Roles: Identity analyst + Server analyst + Evidence skeptic + Report validator

Handoff: Identity analyst supplies the role-object-action matrix and paired baseline. Server analyst checks alternate enforcement paths.

Stop: Stop at uncontrolled identities, records, or providers.

URL consumers, webhooks, importers, renderers, and callback-capable services.

Server fetch and cloud

Roles: Server analyst + Surface mapper + Evidence skeptic + Chain analyst

Handoff: Server analyst supplies unique callback correlation and negative controls. Mapper confirms ownership before any next hop.

Stop: Stop before internal data, credentials, or an unapproved destination.

A candidate issue needs independent reproduction and severity calibration.

Report validation

Roles: Evidence skeptic + Report validator + Chain analyst

Handoff: Skeptic lists confounders checked. Validator receives minimal steps and raw evidence. Chain analyst labels untested edges.

Stop: Stop when current scope, reproducibility, or controlled impact cannot be established.

A broad authorized target with distinct client, identity, and server lanes.

Full surface map

Roles: Surface mapper + Client analyst + Identity analyst + Server analyst + Evidence skeptic + Chain analyst + Report validator

Handoff: All roles write to one deduplicated lead ledger. The coordinator assigns one owner and one next discriminator per lead.

Stop: Stop duplicate collection, stale-scope work, and any lane without a bounded question.

01

Surface mapper

Owns live assets, technology evidence, endpoints, parameters, identities, and provenance. Produces maps, not inflated findings.

02

Client analyst

Owns JavaScript, source maps, browser sinks, postMessage protocols, hidden operations, and client-only trust decisions.

03

Identity analyst

Owns authentication, recovery, OAuth, sessions, object authorization, role matrices, and tenant boundaries using controlled accounts.

04

Server analyst

Owns parsers, injection boundaries, files, server-side fetches, webhooks, caches, and framework-specific server behavior.

05

Chain analyst

Turns supported primitives into prerequisite graphs and removes escalation paths whose edges cannot be evidenced safely.

06

Evidence skeptic

Searches for intended-public behavior, cache artifacts, collaborator access, timing errors, WAF effects, and cheaper explanations.

07

Report validator

Independently reproduces the smallest claim, verifies current scope, checks redactions, and writes numbered evidence-backed steps.

Handoff contract

Never assign “look deeper.”

A handoff carries a stable lead ID, current scope source, factual observation, one falsifiable hypothesis, evidence references, negative control, confounders checked, next discriminator, stop condition, owner, and status.

lead_id: lead-017
owner: identity-analyst
hypothesis: peer role can invoke owner-only mutation
next_discriminator: replay one controlled object ID
negative_control: nonexistent object plus original owner
stop_condition: controlled accounts and records only
Coordinator loop

Every result changes the queue.

  1. Deduplicate the new observation.
  2. Update technology and trust-boundary tags.
  3. Re-rank affected leads.
  4. Stop redundant work.
  5. Assign the cheapest decisive test.
  6. Preserve positive and negative evidence.

Merge operating rules carefully

Read your existing AGENTS.md first. Merge compatible sections instead of replacing project-specific authorization, build, privacy, or deployment rules. Add narrower files closer to specialized work only when the rules truly differ.

Read the official AGENTS.md guide ↗
Plain-text preview

Template

Select a file to preview.

Open plain text