# Video Analysis: "I Stopped Prompting AI One Task At A Time. This Works Better."
**By Nate B. Jones** | Feb 27, 2026 | ~15:38 | 24K views
**URL:** https://youtu.be/A4zMyjkL0Dc

---

## How This Document Was Built (for other agents)

YouTube transcript was IP-blocked from the cloud server. Summary was reconstructed via:
- Web search of the video title + framework terms
- Multiple citation-cross-referenced searches across Reddit, Substack, CXOTalk interviews
- Video title + channel metadata confirmed: "AI News & Strategy Daily | Nate B Jones"

**YouTube Transcript Skill (for coding agent to self-install):**

The Claw agent on victory-bobcat has `youtube-transcript-api` Python package installed (v1.2.4). 
IP is blocked from cloud/Hetzner. To use it effectively:

```bash
# Install
pip install youtube-transcript-api

# Usage (Python)
from youtube_transcript_api import YouTubeTranscriptApi
api = YouTubeTranscriptApi()
t = api.fetch('VIDEO_ID')
transcript = ' '.join([s.text for s in t])

# Workaround for IP blocks (cloud env)
# Option 1: Use yt-dlp with browser cookies
yt-dlp --skip-download --write-auto-subs --sub-format vtt \
  --cookies-from-browser chrome --output /tmp/transcript \
  "https://www.youtube.com/watch?v=VIDEO_ID"

# Option 2: Use a proxy
api = YouTubeTranscriptApi(proxies={"http": "http://proxy:port", "https": "http://proxy:port"})
t = api.fetch('VIDEO_ID')

# Option 3: Use markdown.new as a first-pass title/channel check
# web_fetch("https://markdown.new/https://youtu.be/VIDEO_ID")
```

**OpenClaw skill that covers this:** `media-ingest` skill at `~/.openclaw/workspace/skills/media-ingest/SKILL.md`
To use: say "process this YouTube link https://..." and it routes through the ingest pipeline.

---

## Video Summary

**Core Thesis:** The traditional "prompt one task at a time and watch it happen" model is dead. As AI moves toward autonomous agents, prompting has split into **four distinct engineering disciplines** — and only understanding all four unlocks real ROI. There's already a reported 10x gap widening between people who get this and people who still think "prompt engineering" is the whole game.

---

## The Four Disciplines

### 1. Prompt Craft (table stakes)
The original skill — writing clear, specific instructions. Still necessary. Not sufficient.
The limitation: assumes synchronous interaction (human watching, correcting in real-time).
With long-running agents, you can't babysit every step.

### 2. Context Engineering
**Definition:** Systematically designing the information environment the AI works within — not just what you ask, but everything the AI needs to know to solve the problem without asking you more questions.

Toby Lutke (Shopify CEO) coined this term. The insight: a great prompt that leaves out critical context is a bad prompt.

**Two layers:**
- **Deterministic:** Static prompts, curated knowledge bases, APIs, fixed facts about your world
- **Probabilistic:** Dynamic tools the AI can use (search, RAG, recursive info-gathering)

**Advanced concepts:** RAG systems, MCP (Anthropic's Model Context Protocol), semantic compression, tiered memory for long-running agents (to prevent context window degradation).

**The limit of context alone:** Even perfect context fails if the AI's *intent* is misaligned.

### 3. Intent Engineering
**Definition:** Explicitly encoding organizational goals, values, trade-off hierarchies, and decision boundaries into the AI — not just *what to do* but *what to optimize for*.

**The Klarna Warning:** Klarna's AI customer service agent crushed the metrics — resolved conversations faster, saved millions, replaced hundreds of agents. Then the CEO admitted it "cost something far more valuable." The AI had optimized for speed without encoded intent around customer satisfaction. It *succeeded brilliantly at the wrong objective.*

Intent engineering prevents this. It asks: what does "good" actually mean at the organizational level, and how do we hard-code that hierarchy?

**Key components:**
- Define and rank strategic priorities (e.g. "empathy > resolution speed")
- Encode decision-making boundaries (what the AI should and shouldn't do)
- Build feedback loops for continuous realignment
- Requires leadership + technical collaboration — it's a strategy problem, not a prompt problem

### 4. Specification Engineering
**Definition:** Creating structured, complete, internally consistent documents that autonomous agents can execute against over *extended periods* without human intervention. Treating your organization's entire knowledge corpus as agent-readable and actionable.

This is the quality ceiling of everything your AI does.

**The Five Primitives of a good spec:**
1. **Self-contained problem statements** — zero ambiguity, zero "go ask a human"
2. **Acceptance criteria** — explicit definition of done, so the agent knows when it's finished
3. **Constraint architecture** — hard boundaries the agent must respect (time, scope, safety)
4. **Task decomposition** — complex work broken into sequential, executable steps
5. **Evaluation design** — test cases with known-good outputs; catches regressions when models update

---

## Is This Relevant to OiMy? Analysis

**Short answer: Yes, but selectively. Two of the four disciplines are directly actionable for OiMy now. Two are already handled well.**

### Dimension-by-dimension OiMy relevance:

#### Prompt Craft → Already strong
OiMy V2 prompt (96 lines, companion + BCT) is already well beyond "prompt craft." We have behavioral fixes, energy matching, banned words, safety hedges. Not a gap.

#### Context Engineering → **Partially done, gap worth closing**
This is the most directly applicable framework.

**What we have:**
- Entity context injection (family profiles from H3.0)
- Session facts extraction (`_extract_session_facts()` — added June 24)
- Thread tracker for named people/events
- BCT matrix (1800 rows) as a knowledge base
- Knowledge base coaching JSONL (61 YouTube transcripts) — but NOT yet embedded/RAG'd

**The gap:** Our knowledge base is still a flat file. Jones's framework says the *probabilistic layer* (dynamic retrieval) is what makes context engineering powerful. We have the deterministic layer (static prompts, fixed profiles) but no RAG. The vector store task (open backlog item) maps directly to this.

**Immediate win available:** When a user mentions "screen time" or "tantrums" or "dementia caregiving," we should be retrieving the 3 most relevant coaching chunks from the knowledge base — not just relying on the LLM's training data. This is the vector store + RAG sprint that's been in the backlog.

#### Intent Engineering → **Gap worth addressing — this is the Klarna risk for OiMy**
This is nuanced but important.

OiMy's intent right now is roughly: *be helpful to the family.* But the eval data shows the engine sometimes optimizes for:
- Actionability (giving answers = completion) over warmth (presence = what user actually needed)
- Conversation length (more turns = "engaged") over user experience
- Routing confidence (BCT skill triggered = "working") over appropriate silence

This is the Klarna pattern at a small scale. The move classifier (added June 24) was our first attempt at intent engineering — saying "JUST_RESPOND when casual, BE_PRESENT when venting, PLAN_FORGE when committed." That's encoding intent, not just prompting.

**What's missing:**
- A clear priority hierarchy: `safety > trust > warmth > actionability` — codified, not just implied
- A "silence is sometimes the best response" principle
- An explicit "don't optimize for engagement metrics" instruction somewhere in the system

**What to add:** A short "OiMy Intent Layer" block in the system prompt, separate from behavioral instructions. Example:
```
[INTENT] OiMy's goal hierarchy:
1. User safety (never compromise)
2. Trust in the relationship (never damage it for short-term helpfulness)
3. Warmth and presence (often more valuable than an answer)
4. Practical help (useful, but never at the expense of 1-3)
```

#### Specification Engineering → **This is how to fix the other coding agent delegation problem**
This maps directly to Bharath's repeated challenge: passing context to another coding agent so it can execute without hand-holding.

The "it has completed" → "Add these to an MD file with clear context" pattern in every session IS the specification engineering problem. The agents are getting incomplete specs.

**Jones's five primitives applied to OiMy agent tasks:**

| Primitive | Current practice | Should be |
|-----------|-----------------|-----------|
| Self-contained problem statement | Often missing ("fix the trust issue") | Full context + current state embedded in the brief |
| Acceptance criteria | Rarely explicit | "Eval passes when tone_mismatch < 10%, NPS ≥ 7.0" |
| Constraint architecture | Missing | "Don't touch the BCT path, don't modify skill files" |
| Task decomposition | Ad hoc | Numbered phases with explicit handoffs |
| Evaluation design | Sometimes ("run spot-5 eval") | Always included with specific pass criteria |

The plan files we've been creating (oimy-companion-quality-plan.md, oimy-next-tasks.md) are attempts at spec engineering. Making them more rigorous = faster agent execution.

---

## What OiMy Should Actually Add (Priority Order)

### Priority 1: Intent Layer in System Prompt (1-2 hours)
Add a small `[INTENT]` block at the top of `gemma4_12b_companion_v2.txt` (and thus all companion calls):
```
[INTENT — OiMy priority hierarchy]
1. Never compromise user safety or crisis response
2. Never damage the trust relationship for short-term helpfulness
3. Warmth and presence > advice delivery
4. Practical help when asked, appropriate to trust level
Optimize for: would this user message a friend and say "that actually helped"?
Do NOT optimize for: conversation length, skill triggers, or completion metrics.
```
This directly addresses the "OiMy keeps questioning instead of helping" and "unsolicited advice" patterns from the evals.

### Priority 2: RAG for Coaching KB (1-2 days)
Embed `knowledge_base_coaching.jsonl` (61 transcripts, ~500k tokens) into pgvector or ChromaDB on Hetzner. On every chat_light turn, retrieve top-3 coaching chunks by semantic similarity to the user's message. Inject as context.

This is pure context engineering — the probabilistic layer we're missing. Expected impact: meaningfully better actionability scores (currently 3.3/5) and fewer cases where OiMy gives generic advice when expert-level guidance exists in the KB.

### Priority 3: Specification Engineering for Agent Briefs (ongoing)
Every task file passed to the coding agent should include Jones's five primitives. Template:

```markdown
## Task: [name]

### Problem Statement (self-contained)
[Full context — current state, files involved, what broke, what the goal is]

### Acceptance Criteria
[Exact conditions: "eval passes", "smoke test for X passes", "NPS ≥ Y"]

### Constraint Architecture
[What NOT to touch, what rollback looks like, budget/time limits]

### Task Decomposition
[Numbered phases, each with explicit output]

### Evaluation Design
[How to verify completion: specific commands, test cases, expected outputs]
```

### Priority 4: Dynamic Session Memory (medium-term)
Jones's tiered memory for long-running agents maps to our Honcho integration (currently unblocked — API key not wired). Honcho gives us cross-session memory. The "context degradation" problem he describes (context window fills up, performance degrades) is exactly what happens in long OiMy sessions. Honcho solves it by persisting compressed memories.

---

## What We Should NOT Add

- **Big Five psych profiling** — Jones mentions personality modeling; we consciously skipped this (too invasive, not enough ROI for family context)
- **Multi-session journey arcs** — complex longitudinal planning he alludes to; we're not there yet
- **Full "spec engineering" of all org documents** — that's enterprise AI infrastructure, not relevant at our scale

---

## YouTube Transcript Skill Installation Instructions (for other coding agent)

The Claw agent on victory-bobcat uses the **`media-ingest`** skill to process YouTube videos. The skill lives at:
```
~/.openclaw/workspace/skills/media-ingest/SKILL.md
```

To install the underlying Python transcript capability on a new agent's machine:

### Step 1: Install the library
```bash
pip install youtube-transcript-api
# or if using bun/node environment:
npm install youtube-transcript  # node alternative
```

### Step 2: Test it (note: cloud IPs often get blocked)
```python
from youtube_transcript_api import YouTubeTranscriptApi
api = YouTubeTranscriptApi()
transcript = api.fetch('A4zMyjkL0Dc')
text = ' '.join([s.text for s in transcript])
print(text[:500])
```

### Step 3: If cloud-blocked, use yt-dlp with auto-subs
```bash
yt-dlp --skip-download --write-auto-subs --sub-format vtt \
  --output /tmp/yt_%(id)s \
  "https://www.youtube.com/watch?v=VIDEO_ID"
# Then parse the VTT file
```

### Step 4: Register as an OpenClaw skill
Copy `~/.openclaw/workspace/skills/media-ingest/SKILL.md` to the new agent's workspace:
```bash
mkdir -p ~/.openclaw/workspace/skills/media-ingest/
cp /path/to/media-ingest/SKILL.md ~/.openclaw/workspace/skills/media-ingest/
```

Or if running OpenClaw, the skill can be triggered directly with: *"process this YouTube link [URL]"*

### Alternative: Use web_search + markdown.new for blocked IPs
When the transcript API is blocked, the fallback that worked here:
1. `web_fetch("https://markdown.new/[youtube_url]")` — gets title and channel
2. `web_search("[video title] transcript summary framework")` — gets reconstructed content
3. Cross-reference with `web_fetch` on any Substack/blog posts by the creator

---

## Summary for Codebase

**What this video validated about OiMy's current direction:**
✅ Move classifier = intent engineering (deterministic intent > letting LLM guess)
✅ Session facts extraction = context engineering (deterministic layer)
✅ V2 prompt with behavioral rules = specification (acceptance criteria for companion behavior)
✅ BCT matrix = constraint architecture (techniques to use, what to avoid)

**What it surfaced as gaps:**
🔴 No intent hierarchy in prompt (Klarna risk — optimize for wrong thing)
🟡 Coaching KB not RAG'd (probabilistic context layer missing)
🟡 Agent briefs lack spec engineering primitives (why handoffs lose context)
🟡 Honcho not wired (tiered memory for long sessions)

**Estimated NPS impact of adding Priority 1 (intent layer):** +0.5–1.0 NPS (addresses the "questioning instead of helping" pattern more directly than move classifier alone)

**Estimated NPS impact of Priority 2 (RAG coaching KB):** +0.5–1.5 NPS (improves actionability from 3.3 → ~4.0+, groundedness stays high)