Skip to content

Worked examples

Five designs produced by the genesis skill, in recommended reading order. Each is the verbatim output of an architect session that loaded SKILL.md and applied the eight-step process. Numbers 03-05 are cold-load runs from a fresh agent context (no prior conversation, no human cleanup).

#ExampleWhat it shows
01README iteration with an alignment loopA8 ALIGNMENT LOOP applied to README iteration. Single skill, single goal-steward thread.
02Re-architecting a review panelRe-architecture lesson: a multi-lens panel anti-pattern (everything in one thread) and the corrected design.
03Release notes (minimum viable single skill)Minimal output: 1 skill + 2 assets + 3 scripts. A9 SUPERVISED EXECUTION + S7 + S4. A1 PANEL considered and rejected (lens-count gate did not fire).
04PR review (advisory panel)Multi-primitive panel: 6 personas + 4 assets + 3 scripts + trigger + entrypoint + rule + evals. A6 EVENT + A1 PANEL + DISSENT-WEIGHTED arbiter. R1 SPLIT considered, applied at lens content as R3 EXTRACT.
05PR review (verdict regime)Same prompt as 04 with one constraint removed (verdict required). Regime change: deterministic bridges, schema gate, post-emit verifier loop, graceful tool probes. A8 ALIGNMENT LOOP, B5 ESCALATION, R1 SPLIT considered and rejected with WHEN-clause grounding.

A general-purpose agent received only:

  1. Path to SKILL.md
  2. The operator prompt verbatim (no genesis vocabulary)
  3. Instructions to load assets per progressive disclosure
  4. Constraints: cite patterns with WHEN-clause quotes, render mermaid, apply W6 / W6.2 / W6.3, ASCII-only, stop at step 6 handoff packet

No prior genesis context was carried in. The output is what the skill, cold-loaded, produced.

The examples answer two questions the README cannot:

  • Does the skill produce the same output every time? No. It is prompt-sensitive in a disciplined way. Compare 03 (single skill) to 04 (17 primitives) to 05 (regime change with hardened pipeline).
  • Can I trust the design choices? Each example shows patterns CONSIDERED and REJECTED with WHEN-clause grounding. Most prompt engineering tools never show their rejection logic, so you cannot tell whether a design is justified or arbitrary.

Each example follows the same eight-step shape:

  1. Operator prompt (verbatim).
  2. Goal restated; success criteria pinned to B4 PLAN MEMENTO.
  3. Architectural pattern selection with WHEN-clause quote (the Tier-3 catalogue).
  4. Design pattern composition (the Tier-2 catalogue).
  5. Refactor passes considered (Tier-4) — applied or rejected with reasoning.
  6. Handoff packet: primitives, dependencies, deterministic gates, anti-patterns to watch.
  7. (Cut at step 6 for the cold-load runs.)

Read 01 first for the smallest viable composition, then 02 for the cautionary tale that anchors most architects’ intuition for why A1 PANEL needs fan-out across threads. After that, 03-04-05 in order shows the same architect facing escalating prompt complexity.