In the RCCOF Framework™ (Role, Context, Constraint, Output format, Few-shot), Context is the grounding layer. It supplies the reader, the situation and the facts the model is not permitted to invent. Role decides who is speaking; Context decides what is true and who is listening. A prompt with a strong Role and no Context produces confident prose about a customer who does not exist.
What Context in RCCOF is
Context in the RCCOF Framework is the block of the prompt that supplies empirical ground truth: who the reader is, what problem they arrived with, how much they already know, and which facts are fixed. It converts an open generation task into a bounded one by removing the model's need to guess at any of those four things.
The distinction from Role matters operationally. Role conditions the voice and the depth of authority. Context conditions the subject matter and the audience. The two layers fail in different directions: a missing Role produces competent writing with no point of view, while a missing Context produces confident writing about invented circumstances.
A language model asked to write without Context does not stop and ask. It fills the gap from its training distribution, which means the reader it addresses is the statistically average reader for that topic, and the facts it asserts are the statistically common facts. Both can be wrong in ways that read fluently.
Why missing Context is the largest source of invented detail
Missing Context is the condition under which a model has no choice but to invent, and that is what makes it the most expensive gap in a prompt. The model is not malfunctioning when it fabricates a customer pain point; it is completing a pattern, and the pattern it completes is whatever its training data suggests a paragraph of that shape usually contains.
Three kinds of invention follow predictably from three kinds of missing Context:
| What was left out | What the model supplies instead | How it reads |
|---|---|---|
| Who the reader is | A generic buyer with generic objections | Advice that fits nobody in particular |
| What is factually true about the subject | Plausible facts drawn from similar subjects | Specific, confident, and unverifiable |
| What the reader already knows | An assumed midpoint of awareness | Explaining the obvious, skipping the hard part |
The third row is the one teams notice last. An article that defines a term the reader already understands, then skips the step they were actually stuck on, has not hallucinated anything. It has simply been written for the wrong person, which is a Context failure with no visible error to point at.
The four inputs every Context block must carry
A Context block that carries these four inputs removes the four guesses a model would otherwise make. Fewer than four leaves a gap, and the gap is filled silently. More than four is usually Constraint or Few-shot material that has drifted into the wrong layer.
- The reader. Their role, their responsibility, and the decision they are trying to make. Not a demographic sketch. A person with a job and a pending decision.
- The situation. What prompted them to look for this now, and what they have already tried. This is what stops the output from starting at the beginning of the topic.
- The awareness level. What they already know and what they will find condescending. This sets the floor of the explanation.
- The ground truth. The facts, figures, product behaviours and constraints that are fixed and must not be reinvented, each one stated plainly enough to be checked.
Written out, a usable Context block runs longer than most teams expect, often longer than the instruction it supports. That ratio is correct. The instruction says what to do once; the Context governs every sentence produced.
Reader awareness stages, and how Context encodes them
Reader awareness is the single Context input that most changes the shape of the output, because it decides where the piece starts. Encoding it explicitly costs one line and replaces an assumption the model would otherwise make from the topic alone.
| Awareness stage | What the reader has | What Context must state | Where the output should open |
|---|---|---|---|
| Unaware | A symptom, no name for it | The symptom in their words | Naming the problem |
| Problem-aware | A named problem, no solution class | The name they use, which may differ from yours | The classes of solution that exist |
| Solution-aware | A solution class, no shortlist | The options they are weighing | Selection criteria and trade-offs |
| Product-aware | A shortlist, an objection | The specific objection holding them up | Evidence placed next to that objection |
The rows are not audience segments to be chosen between once. The same person occupies different stages on different questions, which is why the stage belongs in the Context of each individual prompt rather than in a brand-level brief written once a year.
Ground truth: supplying the facts the model may not invent
Ground truth is the part of Context that carries facts, and it earns its place by being checkable. Every statement placed here is one the model is forbidden to improve on, reword into something stronger, or supplement with a plausible neighbour drawn from its training data.
Three rules make a ground truth block hold:
- State the figure and its source together. A number with no provenance in the prompt will come back out with a confident framing the source never supported.
- State the boundaries of what you know. "We have no data on retention past six months" is Context. Leaving it out invites the model to fill the gap.
- Prefer the awkward specific to the tidy general. A model handed a tidy generalisation will elaborate on it. A model handed an awkward specific tends to reproduce it as given.
Ground truth and Constraint touch here but are not the same layer. Context says what is true. Constraint says what the model may not do, including inventing figures that were never supplied. Both are needed, because stating a fact does not by itself forbid the addition of others.
Comparative matrix: assumed context against supplied context
This comparative matrix sets an assumed context against a supplied context on the dimensions that change the output. It is intended as a diagnostic: if your prompts sit in the left column, the failures in the right-hand column are the ones to expect.
| Dimension | Context assumed | Context supplied |
|---|---|---|
| Reader | Statistical average for the topic | A named role with a pending decision |
| Opening line | Starts at the beginning of the subject | Starts where the reader is stuck |
| Figures | Plausible, unsourced, confidently stated | Only those supplied, with their provenance |
| Objections handled | The common ones for the category | The one actually holding this reader up |
| Failure mode | Fluent and irrelevant | Narrow, which is visible and fixable |
| Review effort | High, because errors read as well as facts | Lower, because claims can be checked against the block |
How much Context is too much
Context has an upper bound, and passing it degrades output rather than improving it. The useful test is not length in tokens but whether every element in the block changes something about what gets written. Material that changes nothing still competes for the model's attention.
Three signs a Context block has passed its useful size:
- Facts that never surface. If a supplied fact has not appeared in any output across several runs, it is not grounding anything.
- Repetition across layers. Brand voice instructions in Context are Few-shot material. Prohibitions in Context are Constraint material. Both belong where they can be maintained separately.
- Background nobody asked for. Company history, internal structure and founding narrative rarely change a sentence in the output unless the piece is about them.
Trimming is safer than it feels. Remove one element, run the prompt twice, and compare. A Context element that cannot be shown to change the output has been paying rent on attention for nothing.
Context against Few-shot: what each one actually transfers
Context and Few-shot are confused with each other often enough to be worth separating explicitly, because both of them supply the model with material rather than with instructions. The difference is what that material carries: Context transfers facts and audience, Few-shot transfers voice and shape.
| Property | Context | Few-shot |
|---|---|---|
| Transfers | What is true and who is reading | How an approved output sounds |
| Form | Statements and constraints of fact | Input and output pairs |
| Changes when | The subject or the audience changes | The brand voice or editorial standard changes |
| Failure when absent | Invented facts and invented readers | Correct content in the wrong register |
The practical consequence is maintenance. Facts change per piece, so Context is written per prompt. Voice changes rarely, so Few-shot exemplars are maintained centrally and reused. Merging them means rewriting the voice every time a fact changes.
Position of Context in the 5-layer RCCOF specification
Context occupies the second position in the RCCOF specification, directly after Role and before Constraint. The order is not decorative: each layer narrows what the layer before it left open, and Context narrows subject matter and audience once identity has been fixed.
- Role fixes who is speaking and with what authority.
- Context fixes what is true and who is being addressed.
- Constraint fixes what may not be done or claimed.
- Output Format fixes the structure the result must take.
- Few-shot fixes the register by demonstration rather than description.
Context placed after Constraint tends to be overridden in practice, because prohibitions written before the facts are known are written against an imagined subject. The full specification, including the anti-hallucination architecture the layers combine into, is published on the framework page.
Frequently asked questions about Context in RCCOF
These are the questions that come up most often when teams start writing Context blocks: how long they should be, whether reference documents can replace them, how they differ from a creative brief, and what to do when the facts are genuinely unknown.
Can I paste a reference document instead of writing a Context block?
You can, and it works for ground truth while doing nothing for the other three inputs. A reference document states what is true but rarely states who is reading, what they have already tried, or how much they know. Paste the document, then add those three in your own words.
How is Context different from a creative brief?
A brief is written for a person who will ask questions when something is unclear. A Context block is written for a system that will not ask, and will instead fill any gap from its training distribution. That single difference is why Context has to be explicit about things a human writer would infer.
What do I put in ground truth when the facts are genuinely unknown?
State the absence. Writing that no data exists for a given period is itself ground truth, and it is the instruction that stops the model producing a figure for that period. An empty space in the Context block is an invitation; a stated gap is a boundary.
Does Context need rewriting for every single prompt?
The reader and awareness inputs usually do, because they change per piece. Ground truth about stable product behaviour does not, and is better kept as a reusable block that individual prompts include. Splitting it that way keeps the per-prompt writing short enough that people actually do it.
Where must statistics and proper nouns be supplied?
Statistics, case study figures and proper nouns must be supplied in the Context layer of an RCCOF prompt. If a number is not there, the model is not allowed to produce one.
Which emotional triggers does the Context layer name?
The RCCOF master template names four emotional triggers for the Context layer: fear, want, hope and pride. One is chosen for the persona the prompt is written for.