brand guidance is fractured across tools. the agent is now the maker. ghost gathers what applies and uses it to steer the work.
ghost
a model generates the likeliest thing. ghost makes your brand the likeliest thing.
0 thesis
1 fracture
Teams keep different parts of the brand in different tools.
Product teams maintain components and tokens. Marketing keeps campaign systems and asset libraries. By the time a specialist applies the shared brand guidelines, each discipline has adapted them to its own work.
product
components, tokens, interaction patterns, UI copy
marketing
campaign systems, messaging frameworks, asset libraries
content
voice and editorial guidance
support
macros and service conventions
motion
behavior and timing
legal
approved claims and language
Specialists bridge the gaps through interpretation and review, then repeat the work on the next surface.
2 maker
Agents work across the boundaries that split brand guidance into pieces.
Agents can build a payment screen, write its confirmation email, produce a launch campaign, and revise the support response. The agent still receives a component library for one task, a voice guide for another, and a campaign deck for the next. Different people provide each source at different times.
3 guidance
ghost records brand decisions in prose that agents can find and use.
A `.ghost/` package holds the brand's broad stances that should survive every medium, the narrower decisions that belong to a particular situation, and pointers to the materials where those decisions become concrete.
One decision can guide a screen, email, campaign, and support response. Each medium still keeps its own form.
.ghost/pattern.crop-with-intent.md ---
description: Photography. Gather when a composition uses a photo.
materials:
- brand/photography/**
---
# Crop with intent
Crop around one clear subject. Let the subject meet at least one edge.
Reject cautious full-object framing and collages that avoid choosing a focal point..ghost/condition.blocked-progress.md ---
description: Blocked progress. Gather when someone cannot continue a task.
materials:
- src/components/error-state/**
---
# Keep a path forward
Keep the explanation and one next action together.
If the person cannot resolve the problem, state what happens next. Do not end on a disabled control..ghost/asset.color-roles.md ---
description: Exact color roles and values. Gather before assigning color.
materials:
- src/styles/brand-tokens.css
---
# Color roles
Ink carries content. Signal marks selection. Correction marks review.
```css
--ink: #171714;
--signal: #e8df55;
--correction: #c83e36;
```ink signal correction 4 gather
A useful ghost package may hold hundreds of decisions. Loading all of them into every task would turn shared guidance into one giant prompt.
ghost gather gives the agent a compact view of the
complete ghost package. The agent chooses, then pulls the guidance
that applies to the work.
brand
any task that should express this brand · always in context
voice
anything with words
motion
anything that moves on screen
email.transactional
money moved and the reader is checking
layout.spacing
laying out a page
never.ai-defaults
any first-draft visual surface
logo.usage
placing or sizing the mark
pick a task above
the agent reads the full menu before deciding what applies
observable
Gather shows the complete menu before the agent chooses. If useful guidance does not reach the work, you can see whether the node was missing, its description was unclear, or the agent skipped it.
efficiency
Gather represents every node with a compact ID and description. Pull loads the full prose and materials only for the selected nodes. The agent sees the shape of the whole brand while keeping generation context small and free of unrelated instructions.
One ghost package can guide product, marketing, support, and motion. The agent loads only the guidance that applies to its current task.
5 steer
Steering quality comes from the whole delivery and correction loop, not from a larger prompt.
ghost's optimization target is the first useful draft: work that starts with the right brand constraints, uses the right materials, and leaves a clear trail when it misses. Prompts and prose influence work rather than enforce it. Requirements that must always hold belong in code, schemas, linters, tests, or other executable checks. Human review remains part of the correction loop, not the enforcement boundary.
what reaches the agent
The first bet is relevant over complete. A useful package can be large, but the task should not inherit every rule in it. Gather exposes the whole menu, then pull delivers the bodies selected for the work. The agent sees enough of the brand to choose, and only the applicable guidance becomes generation context.
A small shared stance can be present early without becoming a second handbook. The menu lets the agent choose what applies. Selected prose and real materials arrive for the task. A matching opening structure sits close to generation, where it can shape the first commitment without burying the work under unrelated instruction.
what carries influence
Instructions steer more when they name the replacement move. Do not only forbid a weak pattern. Show the stronger alternative in prose, attach the material the agent can inspect, and include the example that makes the choice concrete. A crop rule, a density pattern, or a support phrase carries more weight when it points at the thing it governs.
Misses are easier to repair when they are diagnosed narrowly.
knowledge gap
the agent invented a token, component, asset, or value because the real material was absent
supply the real files, names, values, or material pointers
steering gap
the agent used the right materials but composed them generically
strengthen the examples, conditions, pattern, or opening structure
delivery gap
the relevant guidance existed but did not reach the task
add the node, sharpen its description, or inspect why it was skipped
correction gap
the output follows the words but still feels wrong in context
render the work, review it separately, and tune the package from the miss
when it arrives
Timing changes the work. Guidance that arrives before generation can shape the plan, the search, and the first draft. Guidance that arrives after the draft mostly becomes critique. ghost commits to the opening structure because the earliest frame is the one the agent builds around.
Volume changes the work too. More instructions can flatten each other until none of them wins. ghost favors compact descriptions, selective pulls, and material-backed detail so the important constraints are visible at the moment of making.
correction
The second half of steering happens after the artifact exists. Render visual work. Read the result separately from the generation prompt. When checks exist, use review as a later pass, not as more instruction stuffed into the first request.
Then tune the package from the evidence. Pulse can show whether a node was seen and pulled. Review can show which guidance applied to the diff. Human feedback can show where the prose was vague, the material was missing, or the example taught the wrong move.
how the bets compound
Each bet depends on the others. Guidance cannot help if it is not selected and delivered. Selection cannot prevent invention unless the package points to real materials. Materials alone do not determine composition. Generation cannot verify every rendered property. Review only improves later work when feedback becomes a package edit.
Together, they make brand guidance more likely to survive the path from request to draft to review. The advantage comes from the complete delivery and correction loop, not any one technique. The exact amount, order, and phrasing still need tuning against real work.