Our Methodology

The 9-phase process

• Phase 0
  Intake

  Understand the agent, define training goals, scope boundaries. Who are you? What do you need to learn? What must not change?
• Phase 1
  Design Spec

  Architecture decisions: what to teach, how to structure it, which knowledge format fits the agent's framework.
• Phase 2
  Research

  Domain knowledge from books and web. We extract from primary sources — we don't generate from memory. Every finding is logged.
• Phase 2.5
  Synthesis

  Cross-reference all sources. Sources disagree — synthesis resolves contradictions explicitly. This step has caught errors in every project.
• Phase 3
  Playbooks

  Step-by-step decision procedures for each workflow. A role description alone adds ~0% over baseline — the measured +17% comes from playbooks, knowledge files and examples together (see the benchmark below).
• Phase 4
  Agent Core

  Compile identity, safety rules, and reasoning framework. The core captures who the agent is — so training doesn't accidentally overwrite it.
• Phase 5
  Reflection

  Self-learning loops, growth triggers, doubt tracking. We build in the mechanism for the agent to keep improving after delivery.
• Phase 6
  Compile

  Assemble all deliverables and deploy to the target platform: CLAUDE.md, .cursorrules, system prompt additions, or custom format.
• Phase 7
  Testing

  Smoke tests, adversarial tests, edge cases — minimum 12 per agent. Adversarial tests specifically probe for behavior we don't want.
• Phase 8
  Packaging

  Deliver to client with full documentation: what was done, what was found, what uncertainties remain, and how to use the pack.

Identity preservation

Evidence approach

We don't hide uncertainty. Every claim in our knowledge packs is tagged:

• [PROVEN]
  Verified against primary sources or direct measurement

  Tested, reproducible, backed by established literature or our own empirical data.
• [DIRECTIONAL]
  Likely correct, not yet fully verified

  Strong evidence, reasonable inference, not yet formally tested. Used where gaps exist but the claim is not misleading.
• [HYPOTHESIS]
  Uncertain, requires validation

  We include hypotheses when they are useful — but we label them clearly. An agent should know what it knows vs. what it suspects.

Training effectiveness

We measure training impact, not just completion. Benchmark results from the Warm Echoes dataset: 1,122 test runs across 17 agents, three conditions.

[DIRECTIONAL] — historical single-run screen (17 agents, 22 tasks each, one run per task): a directional result, not a replicated reliability benchmark. A decision-grade benchmark would need three matched runs per agent per task in each arm; this one has one. Methodology: backward design, comparative scoring across all three conditions, model-consistent evaluation (the same model in all three conditions).

See it in action

Read the full case study — 7 specialized game dev agents, built from scratch.