CONCEPTS · 32
Patterns& architectures.
Agent File Memory as Organisational Asset — When Markdown Notes Become a Corporate Record
An agent's persistent markdown memory, designed to solve the technical problem of context between sessions, turns into something else: a structured corporate record that can be read by the team and, in one case, used as the substrate for a commercial exit negotiation.
Context Before Action — Why an Agent Must Read the Chat Before It Speaks Into It
Before any message goes out, an agent reads the last 20–30 messages of the chat. The rule sounds trivial; the case that produced it is 18 templated 'thank-yous' sent to candidates who never replied. State files lie; the chat does not.
Incident-Driven Configuration — Growing an Agent's Rules by Precedent, Not by Design
An agent's behaviour file isn't designed upfront. Each incident gets logged, analysed, then promoted to a hard rule. The system grows by case law: each precedent becomes a new constraint. Over one month of operation, a config grew by 9+ rules — each tied to a specific mistake.
Persona Through Prohibitions — Engineering an AI Identity by What It May Never Do
A persona that holds across thousands of messages is built from named prohibitions, not friendly prescriptions. LLMs route around 'try to be concise' but comply with 'never use an exclamation mark'.
The Scope-Creep Ledger — How an AI Agent Recovers Money That a Human Producer Cannot See
Across one Swiss-watch project, an autonomous agent monitored 3000+ messages, classified each scope addition against an 8-stage approval pipeline and a pricing table, and assembled a 22-line ledger of significant additional costs. The number is real. The mechanism is generalisable. This is the single use case where the agent does work that a human producer is structurally incapable of doing.
Tiered Agent Autonomy — Routine Acts Auto, Decisions Through Approval
A working agent needs two tiers of authority, not one. Routine acts (acknowledgements, status, file requests, blocker probes) go autonomously; anything decisional (deliverables to client, scope, deadlines, costs) routes through the operator.
Two-Chat Architecture — Splitting a Project into an Autonomous Internal Chat and an Approval-Gated Client Chat
A single production project, two Telegram chats: one with the team running fully autonomous, one with the client where every reply routes through the operator. Information flows one way by rewriting (never forwarding); the reverse flow goes through a human.
Control-Archaeology Refactor
A long-lived AI tool accretes modes. After a few quarters the surface needs an excavation, not another button. The refactor: lift the primary axis to top tabs; drop the dead toggles; reorganize each tab around its natural flow.
The Four-Tier Prompt-Source Hierarchy
When an agent calls a tool to generate something, it usually has more context than the function signature can accept. Pass all four tiers — anchor, emotional, modifiers, fine-tuning — and rank them explicitly; otherwise outputs regress to a templated mean.
Generative UI in Agentic Chat
A chat agent can do more than emit text — it can emit typed UI components that render in-stream. The user's next message becomes a click on a button inside one of those components instead of free text.
Hyperfocus Skills — Six SKILL.md Files for On-Demand Executive Function
The full reference implementation of the Hyperfocus Engineering architecture: six composable Claude Code skills (focus-start, hyperfocus-sprint, dopamine-stack, cognitive-unload, anti-distraction, urgency-mode) built from a shared mechanics catalogue, with verbatim SKILL.md content for each.
Primitives-Library-First: When Brand Exploration Becomes a Design System
After the first exploratory batch ships and the user picks favorites, there's a fork: keep shipping one-offs, or extract a primitives library from the favorites and compose the next batch on top of it. The latter turns exploration into a system.
Programmatic Identity Generation
Treat wordmarks, monograms, and lockups as the output of code, not hand-craft. A primitives library plus palette/font tokens plus a style→technique map produces dozens of variants in minutes.
Happy Horse 1.0 — A Prompt-Engineer System Prompt
Production system prompt that converts a free-form brief into one Happy Horse 1.0 video prompt — slot-ordered, beat-structured, single-camera, language-pure, with a built-in checks line.
Happy Horse 1.0 — Prompt Rules
Production prompt-engineering rules for Happy Horse 1.0: 8-slot structure, front-loaded subjects, single camera move, language purity, mandatory beat plan.
I2V Storyboard Multi-Cut: a Multi-Shot Sequence from a Single I2V Input
An image-to-video trick: feed a 2×2 storyboard as the single input, then prompt the model to snap-cut into each panel in turn. Result — a 4-shot sequence from a single I2V call.
Brief → Scenario → Scenes: a Narrative-Ingestion Pipeline
A three-stage cascade that turns a free-form client brief into a structured, editable, shot-granular scene list — the upstream of a generative video pipeline.
Chaos-Frame Feed Architecture: One Hero, Eight Outtakes
A prompt architecture for 3×3 social-feed grids where only one of nine frames is the intentional hero — the other eight are deliberately chaotic. Polish is scarce by design.
Edit-Preservation: Apply Only the Requested Change
A prompt architecture for iterative LLM refinement: the baseline exists, the edit is surgical, everything else stays verbatim. Makes iteration cheaper than re-generation.
Mixed-Prompt Composition: One Image, Five Typed References
A single-LLM-call pattern that merges identity, pose, background, clothing, and product references into one coherent image-generation prompt — explicitly typed per reference so the generator never silently pulls the wrong dimension.
Seven-Element Shot Description
A prompt-engineering pattern between rigid JSON and free prose: mandatory coverage of seven named dimensions, expressed as one paragraph.
Why Storyboards Come Before Renders
Hand-drawn-style storyboards aren't decoration in a generative-video pipeline — they're a cheap visual lock on composition. Cheap to produce, fast to reject, and they double as vision references for the final-render stage.
Style-Prefix Architecture
Prefix every prompt in a class with the same style string and you turn the class's outputs into a neutral visual vocabulary downstream stages can read for composition, not style.
Template-Dispatch Prompts
Don't make one universal prompt handle every case. Use a typed flag to dispatch to a focused per-case prompt variant. Each variant is small, independent, and easy to debug; the LLM never sees the branching.
Two-Stage Architect Pattern
Split creative intent and technical generation across two LLM calls. A Creative Director emits prose concepts; a Technical Architect converts them into structured generator-ready prompts. Different temperatures, different system prompts, different failure modes — isolated.
Typed-Reference Composition: Every Reference Image Has One Job
A prompt architecture for multi-input image generation: every reference carries an explicit typed role, and the prompt names inline which reference contributes which dimension. Stops silent-reference hallucination at its source.
Seedance 2 — Censorship & Face-Filter Bypass Rules
A two-layer defense for Seedance 2 content moderation: a white grid overlay on the reference image, plus phrasing hygiene in the prompt itself.
Seedance 2 — A Cinematic Video Prompt Engineer
A production system prompt that turns a scenario plus reference stillshots into a single compact JSON object Seedance 2 consumes per shot.
A Workflow for Generating Seedance 2 JSON Prompts
End-to-end process: scenario plus reference stills in, production-ready Seedance 2 JSON out. Image pre-flight, scene mapping, character budget, censorship hygiene.
A.O.C. Application Skills — Three Composable Skills for Applied Anchor/Optics/Chemistry
Reference implementation for the A.O.C. framework: three Claude Code skills (aoc-auto, generation-prompts, typed-key-visual) that apply Anchor/Optics/Chemistry as an automatic backbone for any image or video generation prompt.
Creative Direction Skills — Brief → Moodboard and Brief → Scenario
Two Claude Code skills that handle the pre-prompt half of a generative pipeline: turning a brief into an 8-direction moodboard plan, then turning the chosen direction into a 6-shot scenario in the studio's 7-element shot structure.
MiniMax Voice Skill — Free Text to Production-Ready T2A JSON Request
A Claude Code skill that converts a free-form line (Russian or English) into ONE production-ready MiniMax T2A speech-synthesis JSON request, with inline pause markers, interjection tags, character-preset voice tuning, and a one-line tuning hint.