Prompt Lifecycle Design Rationale
Status: Practical Guide / Design Rationale — not a finalized protocol, benchmark, or experiment template.
1. Purpose
This document explains why each prompt form in the CIP/PAL workflow exists, what problem it was designed to solve, and where in the Identity Lifecycle it belongs.
The prompts described here are not a linear improvement series. Each emerged from a different operational problem encountered at a different stage of image-generation and sequence-generation practice. They should be understood as distinct artifacts with distinct roles, preconditions, and constraints.
Prompts are not ranked by quality or power. Each exists because a different problem required a different solution.
2. CIP/PAL Framing
Within the CIP/PAL reconstruction model:
A → (A + C) → A′ → B′ ≠ B
Where:
- A = human original intent, constraints, or approved state
- C = model-side or execution-structure mediation that transforms A into A′
- A′ = task understanding or execution form after C-mediated transformation
- B′ = generated output or candidate output subject to adoption
- B = intended or human-expected result
This is a risk-condition expression. It does not claim that every output necessarily diverges from B. It keeps the observation structure visible: C may mediate A into A′, and B′ may differ from the intended B.
CIP/PAL does not directly control C or A′. It supports governance of the workflow conditions around the transformation from A to A′, including validation, rejection, purge, re-binding, re-convergence, and adoption.
3. Prompt Lifecycle
The Identity Lifecycle proceeds through the following prompt sequence:
Source A
↓
Minimum Requirements
↓
YAML Prompt (Source of Truth)
↓
Minimal Prompt (Character Exploration)
↓
Long Prompt (Identity Establishment)
↓
Optimization Prompt (Convergence Adjustment)
↓
Ultra-Compressed Prompt (Identity Recall)
After convergence is established, series generation uses a repeating state transition:
Converged State
↓
Transition Prompt
↓
Next Approved State
Each stage is described below with its design rationale, problem context, phase, and constraints.
4. Minimum Requirements
Problem it solves
Before any prompt is written, there must be a clear definition of what is required. Without this, prompts may reflect the writer's assumptions rather than the human operator's approved intent.
Design Rationale
Minimum Requirements establishes the human-approved source state A. It is the authoritative baseline against which all derived prompts and generated outputs are compared. Without it, there is no stable A to drift from.
Phase
Pre-prompt. The starting point of the entire lifecycle.
What it does not replace
Nothing. It is the source, not a derived artifact.
Prerequisites
None.
5. YAML Prompt (Source of Truth)
Problem it solves
Natural-language prompts are lossy. When a character definition is translated into English prose, useful redundancy is removed, prohibitions are weakened, priority relationships collapse, and structural separation disappears. Each editing cycle degrades the original intent further.
Design Rationale
The YAML Prompt preserves the structured source state in a human-reviewable, diffable, and comment-supporting format. It separates immutable identity elements, controlled variables, scene variables, reference image policies, and workflow governance into named sections. Natural-language prompts are derived projections of the YAML, not equivalents.
The YAML Prompt is not a generation prompt by default. It is a reviewable master specification.
YAML = human-approved source state. Natural-language prompts = derived artifacts.
Phase
Source state management. Parallel to all other prompt forms.
What it does not replace
The YAML Prompt does not replace the Long Prompt, Minimal Prompt, or any other derived form. It is the source from which those forms are derived.
Prerequisites
Human-approved Minimum Requirements.
6. Minimal Prompt (Character Exploration)
Problem it solves
Long prompts activate more C-mediated transformation, increasing the risk of unintended A′. Observing how much identity can be recovered from minimal input reveals which attributes are load-bearing.
Design Rationale
The Minimal Prompt reduces prompt surface area to reduce optimization pressure from the model's training priors. It is used to explore reconstruction behavior, not to establish identity for production use.
A minimal prompt is not merely a short prompt. It is a controlled projection from a structured source state.
Minimality without source-of-truth becomes accidental compression.
Phase
Character exploration. Pre-convergence.
What it does not replace
The Minimal Prompt is not a substitute for the Long Prompt in identity-establishment workflows. It is not sufficient for first-generation use when identity continuity matters.
It is not appropriate as the sole prompt for:
- identity fixation (establishing a repeatable, human-approved character state);
- series generation (producing a sequence of images with consistent identity);
- continuous small-change generation (pose, expression, or scene variation within an established identity).
In these cases, additional anchoring, stronger constraints, and human approval checkpoints are required. For post-convergence series generation or continuous small-change workflows, approved anchor material is normally required.
Relationship to short-prompt and tag-list workflows
Some generative AI workflows — including those common in Stable Diffusion and similar systems — rely on short prompts or tag lists as a primary generation input. These approaches can be effective for exploration and variation within a familiar style or training distribution.
However, using minimal prompts or short tag lists without sufficient anchoring and constraints for identity fixation purposes may increase drift risk. The model may fill in unspecified attributes through C-mediated reconstruction, producing outputs that appear plausible but diverge from the intended identity.
A short tag list may specify broad attributes such as hair color, age range, clothing, or visual style, while leaving identity-relevant details such as facial structure, eye shape, hair flow, body proportion, expression, and atmosphere underspecified. In such cases, checkpoint behavior, sampler behavior, seed, prompt weighting, learned distribution, LoRA, DreamBooth, reference images, ControlNet, IP-Adapter, img2img, and other execution conditions may fill those unspecified regions. These execution-side conditions are forms of C in the CIP/PAL framing. They are not inherently harmful, but they become governance-relevant when identity continuity across multiple generations is required and the unspecified regions are identity-critical.
This is not a claim that short-prompt workflows are inherently less stable. It is an observation that minimality without source-of-truth and without human approval checkpoints may allow C to reconstruct unspecified attributes in unintended directions, particularly when identity continuity across multiple generations is required.
Prerequisites
Human-approved Minimum Requirements. Preferably a YAML source state. Convergence is not required.
7. Long Prompt (Identity Establishment)
Problem it solves
Without a sufficiently detailed initial prompt, C-mediated transformation fills missing information with distributional defaults, producing a B′ that diverges from the intended B. Identity attributes that are not explicitly stated may be reconstructed in an unintended direction.
Design Rationale
The Long Prompt makes A dense and explicit before the model forms A′. It re-injects context, identity conditions, constraints, priorities, prohibited transformations, and decision boundaries. It is the primary tool for initial convergence and for re-convergence after drift.
The Long Prompt does not eliminate C. The model may still compress, summarize, reprioritize, or reinterpret the instruction before producing B′. However, it reduces the missing information that C might otherwise fill in freely.
In practical use, the Long Prompt functions more like a convergence-stage prompt for re-establishing the intended context, constraints, identity conditions, and decision boundaries before compressed prompts can safely be used for re-binding or recall.
Phase
Identity establishment. First generation and re-convergence after drift.
What it does not replace
The Long Prompt is a derived artifact — a controlled projection of the YAML source state into natural language. It does not replace the YAML master.
Prerequisites
Human-approved YAML master or equivalent source state.
8. Optimization Prompt (Convergence Adjustment)
Problem it solves
After observing the first generation result, specific drift points become visible. A general-purpose correction may improve image quality while introducing new drift. What is needed is a targeted correction that addresses only the observed deviation from A.
Design Rationale
The Optimization Prompt targets specific observed drift. Correction must be aimed at returning to A, not simply improving image quality. It is not a general optimization of the prompt. It is a drift-correction instrument.
Used before convergence exists, the Optimization Prompt may fill missing information through C and produce a coherent B′ that is not sufficiently bound to A.
Phase
Convergence adjustment. After first generation review, not before.
What it does not replace
The Optimization Prompt must not be used for first generation. It must not replace the Long Prompt or the YAML master.
Prerequisites
Observed B′ and specific drift points must exist. The human-approved source state remains authoritative. Not for first-generation identity establishment.
9. Ultra-Compressed Prompt (Identity Recall)
Problem it solves
After convergence is established, re-generating from the full Long Prompt for every subsequent generation is unnecessary and risks introducing new C at each step. A lighter recall mechanism is needed that leverages the existing convergence state.
Design Rationale
The Ultra-Compressed Prompt works because a human-approved state, effective constraints, and rejected-drift history already exist. Without those preconditions, it is insufficient for identity establishment.
Compressed prompts are not initialization substitutes. Compress only after converging.
The Ultra-Compressed Prompt preserves only the identity core, essential fixed features, allowed variation range, critical drift warnings, and major rejection boundaries. It is a lightweight key for returning toward a human-approved anchor, not a complete definition of A.
Phase
Identity recall after stabilization. Post-convergence only.
What it does not replace
The Ultra-Compressed Prompt must not be used for first generation, identity recovery, new character creation, or unapproved scene change. It must not replace the YAML master or the Long Prompt.
Prerequisites
Human-approved convergence state. Anchor image. Established convergence history. Compression safety review passed.
10. Transition Prompt (Sequence Continuation)
Problem it solves
In sequence-generation workflows, a character must move through a series of scenes without losing identity continuity. The Long Prompt is designed for identity establishment, not for controlled scene-to-scene transition within an established sequence.
Design Rationale
The Transition Prompt assumes convergence already exists and focuses on the state change from one approved scene to the next. It preserves the established identity while allowing controlled variation in scene, pose, or context.
The Transition Prompt is the tool used when the problem is not "who is this character" but "how does this character move to the next state."
Phase
Sequence continuation. After convergence is established.
What it does not replace
The Transition Prompt must not be used before convergence exists. It does not replace the Long Prompt for initial identity establishment. It does not replace the YAML master.
Prerequisites
Established and human-approved convergence state. Approved anchor image. Defined target scene or sequence state.
11. Key Distinctions
Prompt compression is not initialization
The compressed prompt works partly because a human-approved state, effective constraints, and rejected-drift history already exist. A short prompt is not a replacement for that convergence history.
| Prompt type | Operational role |
|---|---|
| Long Prompt | Initial convergence / task-frame re-establishment |
| Optimization Prompt | Drift correction after first generation |
| Ultra-Compressed Prompt | Lightweight recall after stabilization |
| Transition Prompt | Scene or state change within established sequence |
Identity Establishment vs Identity Recall
Identity Establishment is the governed process of creating a first human-approved convergence state. It requires the Long Prompt and human approval at each step.
Identity Recall is the process of re-invoking an already-established identity state using a lighter mechanism. It requires the convergence state, anchor image, and Ultra-Compressed Prompt — none of which can substitute for each other.
Identity Recall is not identity recovery. If the convergence state has been lost or contaminated, recovery requires re-establishment, not recall.
YAML is the source, not the prompt
The YAML Prompt is the human-approved master specification. All natural- language prompts are derived projections. When generation quality degrades, return to the YAML source — not to the English prompt.
12. Non-Claims
This document does not claim:
- any prompt form eliminates C;
- any prompt form guarantees identity preservation;
- any prompt form directly controls A′ or model behavior;
- prompts are ranked by quality, power, or version number;
- CIP/PAL directly controls C, A′, or the generative model;
- this document defines a finalized protocol, benchmark, or standard.
13. Summary
Each prompt form in the CIP/PAL lifecycle exists because a different problem required a different solution at a different stage of identity-governed generation work. They are not a linear improvement series.
Understanding why each prompt exists — what problem it solves, what it does not replace, and what preconditions it requires — is more important than knowing which is "shorter" or "stronger."
Human approval remains the control point at every stage.
This document is a Practical Guide and Design Rationale record. It is not a finalized protocol, benchmark, or experiment template.
See also: YAML Prompt Source-of-Truth Pattern — Prompt Compression After Human-Approved Convergence — Translation Loss and the YAML-First Principle — Glossary