DONNA DELTA

DELTA-001 Source: Stop Prompting. Start Looping. — AI Edge / Miles Deutscher

4 Gaps Captured

Core Insight

Prompt engineering is dead. The new paradigm is loop engineering: AI tries → self-checks → fixes → retries → marks done. You move from being in the loop to being outside the loop. This is only possible because new models (Claude Fable 5, GPT 5.6) are smart enough to self-evaluate.

A loop has four components: Goal (defined completion criteria), Checklist (definition of done), Inspector (self-review pass), and Budget (stop after N attempts).

Two Loop Types

Inner loop: Single task — try, inspect, fix, retry. AI self-corrects before delivering.
Outer loop: Recurring scheduled job that runs inner loops each iteration. Self-improves over time.

Implementation Gap Analysis

Capability	Hermes Has	Loop Eng.	Gap
Scheduled recurring tasks	✅ 17 crons	Outer loop	None
Self-review within task	❌ Missing	Inner loop	BIG
Inspector/second pass	❌ Missing	Grader model	BIG
Hard definition-of-done	⚠️ Soft goals	Checklist	Medium
Skills auto-update from loops	❌ Missing	Loop→skill	Medium
Per-task attempt budget	⚠️ Repeat only	Stop after N	Small

Implementation Plan

P0 — Inner loop skill. Created loop-engineer skill. Pattern: GOAL + DONE WHEN checklist + STOP AFTER budget → delegate_task → inspect → retry up to N.
P0 — Inspector pass. Five-point checklist before any subagent output reaches Dwad: exists, meets checklist, no hallucinations, correct format, passes sniff test.
P1 — Upgrade crons. Add DONE WHEN checklists and STOP AFTER budgets to quality-critical cron jobs (audit-dojo, fabric-health-check first).
P2 — Skills auto-update. Add to every LLM-driven cron prompt: "If you discover anything that would improve the [skill], patch it before delivering."
P3 — Goal template library. Build reusable DONE WHEN checklists for recurring task types.

DELTA-002 Source: Stop prompting Claude. Use Andrej Karpathy's Loops Instead… — Dream Labs AI

6 Gaps Captured

Core Insight

Karpathy's auto-research pattern is now a native Claude Code feature — /goal and /loop. The new unlock is evaluator cartridges: custom scoring algorithms that grade AI output autonomously so the loop knows when to stop. No human in the loop. The combination — goal-driven agent + multi-cartridge evaluator + attempt budget — produces output that iteratively sharpens without you touching the keyboard.

This is DELTA-001's loop engineering with the inspector automated. The human moves from "checking every output" to "defining the scoring rubric once."

Evaluator Cartridge Concept

A cartridge is a self-contained scoring algorithm that takes output → returns a number (typically 1-10). Three cartridges shown:

1. Humanify Cartridge — scores voice, variety, predictability (AI slop detector). Only 9+ passes.
2. Hormozi AI Cartridge — scored against Alex Hormozi's public writing style for marketing punch.
3. Open Rate Analytics Engine — scores predicted open rate from past email data, audience behavior, subject line patterns.

The magic: composite scoring across multiple cartridges. The newsletter loop required 27/30 total, no single cartridge below 8. The agent ran 5 attempts autonomously, improving from 7.1 → 26.8 over ~40 minutes.

Claude Code Native Features (Not Us)

/goal — keeps Claude working until a clear condition is met. Autonomous, no human prompting between attempts.
/loop — wakes Claude up repeatedly to check on something. Condition-based, not schedule-based.

These are Claude Code native features — not available on Hermes. Our equivalents are manual (below).

Implementation Gap Analysis

Capability	Hermes Has	Source Says	Gap
Autonomous goal-driven loops	⚠️ Manual pattern	/goal native	BIG
Evaluator cartridges (scoring)	❌ None	Multi-cartridge	BIG
Condition-based wake-up	⚠️ Cron (schedule)	/loop (condition)	Medium
Subagent swarm → single goal	⚠️ Independent tasks	650 experiments	Medium
Attempt tracking + scoring history	⚠️ STOP AFTER only	Scoring table	Medium
Scoped file access for loops	❌ Full FS	Scoped access	Small

Implementation Plan

P1 — Simple cartridge: humanify scorer. Build a single evaluator cartridge as a standalone script. Takes text in → scores 1-10 on human-like voice, variety, predictability. Use our existing humanizer skill's criteria as the rubric. This is the beachhead — prove the pattern works on Hermes before building multi-cartridge.
P1 — Autonomous retry loop in loop-engineer. Upgrade loop-engineer from "orchestrator must manually inspect" to "subagent self-scores against a checklist, retries until pass or budget." The cartridge output becomes the pass/fail gate.
P2 — Multi-cartridge composite scoring. After P1 proves single-cartridge works, add composite: run N cartridges, sum scores, enforce per-cartridge minimums. This enables the "27/30 with no cartridge below 8" pattern.
P2 — Scoring table in cron output. When a cron job runs with loop-engineer, include the per-attempt scoring history in the delivery so Dwad can see the improvement trajectory.
P3 — Condition-based cron wake-up. Extend cronjob to support "run every N minutes UNTIL condition met" — merges schedule + goal into one primitive.
P4 — Swarm coordinator for single goal. Pattern where N subagents all optimize toward the same measurable goal with shared best-score state. The Karpathy 650-experiment pattern.