The Holistic Prompt & Context Canvas (HPC-10)

Three Layers. Ten Elements.

Before the first module, meet the framework the whole course is built around. You'll use it diagnostically from Module 3 onward, so it's worth reading once now and returning to as a reference — many learners print it.

HPC-10 is the deliberate, smaller sibling of the Holistic Agent course's own HAC-20 framework (Five Planes, Twenty Elements), which governs full agent systems. HPC-10 covers the ten elements that determine whether a single AI interaction — a prompt, a request, a task given to an assistant — succeeds. Master this, and HAC-20 will feel like an extension of something you already know, not a new framework to learn from zero.

Layer 1 — INTENT: what you want

#	Element	The question it answers	Failure if unanswered
1	Purpose	What outcome are we actually trying to achieve — stated as an outcome, not an activity?	The AI optimises for a plausible-sounding answer instead of your actual goal
2	Instructions	What, specifically, should the AI do — and in what order, if sequence matters?	Vague requests get vague or arbitrarily-shaped output
3	Constraints	What rules, boundaries, or exclusions must the output respect?	Technically-responsive output that's unusable — wrong tone, wrong length, non-compliant

Layer 2 — INFORMATION: what the AI needs to know

#	Element	The question it answers	Failure if unanswered
4	Context	What must be in view right now — background, prior turns, your actual situation?	A generic answer to a generic version of your question
5	Knowledge	What facts, documents, or data should this be grounded in?	Hallucination — fluent, confident, ungrounded output
6	Resources	What examples, references, tools, or systems can the AI draw on or use?	The AI guesses at a format or capability instead of using what's actually available

Layer 3 — ASSURANCE: how you know it worked, and how it gets better

#	Element	The question it answers	Failure if unanswered
7	Output	What should the result look like — format, structure, length, audience?	Correct content in a shape nobody can actually use
8	Evaluation	How do you determine whether the result is actually good?	You can't tell improvement from noise — every judgement is a vibe
9	Verification	How are specific claims or actions checked before you rely on them?	A confident wrong answer reaches a decision unchecked
10	Iteration	What did you learn, and what changes next time?	The same failure recurs indefinitely; nothing compounds

Using HPC-10 diagnostically

Almost every disappointing AI response maps to one missing element. Once you know the mapping, fixing a bad result stops being guesswork.

Symptom	Missing element	Fix
Confident, plausible, wrong answer	5 Knowledge or 9 Verification	Ground it in sources; check the claim before trusting it
The AI did a different task well	1 Purpose or 2 Instructions	Restate the outcome in one sentence; make the task explicit
Right content, wrong shape	7 Output	Specify format, audience, and length before generating
Worked once, fails on the next similar case	8 Evaluation	You had a demo, not a test
Same mistake keeps happening	10 Iteration	Nothing is capturing what you learned last time
Generic, textbook-sounding answer	4 Context	The AI never received your actual situation
The AI invents a tool, fact, or format that doesn't exist	6 Resources	It was never told what's actually available
Response ignores an obvious rule	3 Constraints	The rule was assumed, never stated

The Prompt–Context Cycle

The companion to the canvas — what you actually do, in order, each time:

DESIGN → TEST → OBSERVE → DIAGNOSE → MODIFY → RETEST → EVALUATE → STANDARDISE

Design against the ten elements. Test on real inputs, not one lucky example. Observe what actually came back — not what you expected. Diagnose which element was missing, using the table above. Modify only that element. Retest before declaring victory. Evaluate against a standard, not a feeling. And once it holds, standardise it into a reusable template — Module 4's exercises and the Prompt Library later in this course exist because re-deriving a good prompt from scratch every time is wasted effort.

Two rules the rest of this course keeps coming back to:

More context is not automatically better context. The smallest set of high-signal information beats the largest set of possibly-relevant information — a finding with real architectural grounding, covered in Module 5.

A prompt that works once is a demonstration. A prompt that works reliably is engineered. Covered fully in Module 11.