Editing Openai/6969614e-00e4-8001-ac42-dda7410c3c4a (section)

=== Assistant: Below is a practitioner-level answer—assuming the person already understands prompt basics and wants durable, high-sign… ===
Below is a practitioner-level answer—assuming the person already understands prompt basics and wants durable, high-signal behavior for analysis, studying, reasoning, and critique.

No emojis, no fluff.

==== Prompting is not control; it is conditioning within a session and calibration across sessions. ====

ChatGPT does not truly “learn” a user over time unless:
* preferences are stored via custom instructions / memory, or
* behavior is re-established reliably each session.

So the correct model is:

: 

Not “one perfect prompt.”

==== ### ====

Use both, but for different layers.

===== #### =====

Use these for:
* Tone (formal, skeptical, precise)
* Depth expectation (graduate-level, professional)
* Epistemic stance (cite uncertainty, flag assumptions)
* Interaction style (challenge weak premises, ask clarifying questions only when necessary)

Good for:

: 

Bad for:
* Task-specific logic
* Methodology
* Domain-specific frameworks (they bloat memory and dilute signal)

Think of custom instructions as personality + epistemology, not workflow.

====== This is a short document you paste at the start of serious work. ======

A good charter defines:
* Role
* Standards
* Methods
* Failure modes to avoid
* Output format norms

Example (compressed):

<syntaxhighlight>Charter:
You are an analytical assistant operating at graduate / professional level.
Priorities:
# Accuracy > speed
# Explicit reasoning chains
# Critique weak premises rather than accept them
# Distinguish facts, inferences, and speculation
# Offer counterarguments when relevant

Defaults:
* Ask clarifying questions only if ambiguity materially affects conclusions
* Use structured reasoning (headings, steps, tables where useful)
* Cite uncertainty explicitly

If tradeoffs exist, explain them.

</syntaxhighlight>

This is far more reliable than long custom instructions alone.

==== ### ====

Most people write prompts like requests.

===== Write prompts like specifications. =====

===== <syntaxhighlight>Context: =====
(What the model must assume as true)

Task:
(What to do)

Constraints:
(What to avoid, limits, standards)

Method:
(How to think, not just what to output)

Output format:
(So structure doesn’t drift)

</syntaxhighlight>

Example for critique:

<syntaxhighlight>Context:
Assume the audience is expert but skeptical.

Task:
Critique the argument below for logical validity and evidentiary sufficiency.

Constraints:
* Do not restate the argument unless needed for clarity
* Flag unsupported assumptions explicitly
* Separate logical flaws from empirical gaps

Method:
Evaluate premises → inference → conclusion.
Consider counterexamples.

Output format:
* Summary judgment (2–3 sentences)
* Major issues (bulleted)
* Minor issues (bulleted)

</syntaxhighlight>

This produces repeatable behavior, even across sessions.

==== ### ====

Meta-constraints like:
* “Be concise”
* “Think step by step”
* “Act like a professor”

Are weak unless operationalized.

Better:
* “Limit summaries to 3 sentences”
* “Show reasoning as numbered steps”
* “Assume the reader already knows X”

===== - You reuse the same skeleton =====
* You only change content, not structure

This reduces drift and hallucination.

==== These are the real killers: ====

===== Long, multi-role, contradictory instructions cause: =====
* Priority confusion
* Shallow compliance
* Style drift

Fix: keep them short and principled.

===== Example: =====
* “Be maximally critical”
* then later: “Be supportive and validating”

The model doesn’t resolve value conflicts well.

Fix: explicitly declare mode shifts:

: 

===== If each task is phrased differently, behavior won’t stabilize. =====

Fix: reuse scaffolds.

===== Saying “that’s not what I meant” is weak. =====

Better:

: 

The model responds better to pattern correction than dissatisfaction.

===== Memory is best for preferences, not procedures. =====

Procedures should live in:
* charters
* templates
* saved prompts

==== Think in terms of components: ====

===== - Custom instructions =====
* Short, value-level guidance

===== - Analysis template =====
* Critique template
* Study / synthesis template

===== Explicit commands like: =====
* “Switch to examiner mode”
* “Assume adversarial peer review”

===== Occasionally ask: =====

: 

This surfaces drift early.

==== : ====

So:
* Put charters at the top
* Restate critical constraints when stakes are high
* Don’t assume past chats transfer behavior

Consistency is engineered, not remembered.

==== If they want long-term, consistent analytical behavior, the best method is: ====
* Minimal, principled custom instructions
* A reusable charter for serious work
* Standardized prompt scaffolds
* Explicit mode control
* Active correction of drift

That’s how you turn ChatGPT from a clever assistant into a reliable analytical instrument.