@design-intelligence/ghost 0.28.0 → 0.29.0
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/dist/commands/skill-command.d.ts.map +1 -1
- package/dist/commands/skill-command.js +8 -1
- package/dist/commands/skill-command.js.map +1 -1
- package/dist/init-payloads/skeleton/glossary.md +3 -3
- package/dist/init-payloads/vessel-light/checks/shape-matches-job.md +2 -2
- package/dist/init-payloads/vessel-light/glossary.md +3 -3
- package/dist/init-payloads/vessel-light/grammar.conversation.md +1 -1
- package/dist/init-payloads/vessel-light/grammar.hierarchy.md +1 -1
- package/dist/init-payloads/vessel-light/grammar.job.md +13 -13
- package/dist/init-payloads/vessel-light/grammar.surfaces.md +1 -1
- package/dist/init-payloads/vessel-light/index.md +6 -11
- package/dist/init-payloads/vessel-light/materials/{ref → examples}/composition.conversation.html +3 -3
- package/dist/init-payloads/vessel-light/materials/{ref → examples}/composition.editorial.html +3 -3
- package/dist/init-payloads/vessel-light/materials/{ref → examples}/composition.form.html +3 -3
- package/dist/init-payloads/vessel-light/materials/{ref → examples}/composition.overlay.html +3 -3
- package/dist/init-payloads/vessel-light/materials/{ref → examples}/composition.table.html +3 -3
- package/dist/init-payloads/vessel-light/materials/{ref → examples}/email.html +3 -3
- package/dist/init-payloads/vessel-light/materials/tokens.css +1 -1
- package/dist/init-payloads/vessel-light/register.data-density.md +1 -1
- package/dist/init-payloads/vessel-light/register.editorial.md +1 -1
- package/dist/init-payloads/vessel-light/register.email.md +1 -1
- package/dist/scan/templates.d.ts +1 -1
- package/dist/scan/templates.js +1 -1
- package/dist/scan/templates.js.map +1 -1
- package/dist/skill-bundle/SKILL.md +6 -9
- package/dist/skill-bundle/references/authoring.md +145 -0
- package/dist/skill-bundle/references/concrete.md +131 -0
- package/dist/skill-bundle/references/making.md +4 -5
- package/dist/skill-bundle/references/nodes.md +138 -0
- package/dist/skill-bundle/references/schema.md +50 -57
- package/dist/skill-bundle/references/self-check.md +0 -1
- package/dist/skill-bundle/references/steering-audit.md +2 -11
- package/package.json +1 -1
- package/dist/skill-bundle/references/adapting-a-starter.md +0 -110
- package/dist/skill-bundle/references/authoring-scenarios.md +0 -187
- package/dist/skill-bundle/references/blocks.md +0 -179
- package/dist/skill-bundle/references/capture.md +0 -422
- package/dist/skill-bundle/references/concrete-tiers.md +0 -95
- package/dist/skill-bundle/references/distill.md +0 -205
|
@@ -0,0 +1,131 @@
|
|
|
1
|
+
---
|
|
2
|
+
name: concrete
|
|
3
|
+
description: Bind ghost guidance to concrete material in formats agents can use directly.
|
|
4
|
+
---
|
|
5
|
+
|
|
6
|
+
# Recipe: Bind Guidance To Concrete Material
|
|
7
|
+
|
|
8
|
+
**Goal:** make the intended continuation cheaper than the generic one without
|
|
9
|
+
turning the package into a second implementation or a fixed template library.
|
|
10
|
+
|
|
11
|
+
Concrete material commonly does three jobs:
|
|
12
|
+
|
|
13
|
+
| Form | What it constrains | What it buys |
|
|
14
|
+
| --- | --- | --- |
|
|
15
|
+
| Tokens | values and role names | stops plausible near-miss colors, sizes, and durations |
|
|
16
|
+
| Skeletons | opening structure | commits layout before a generic structure wins the first tokens |
|
|
17
|
+
| Components | reusable decisions | compresses variants, states, spacing, and behavior into names |
|
|
18
|
+
|
|
19
|
+
## Scope is not delivery
|
|
20
|
+
|
|
21
|
+
Scope says how much a material constrains. Delivery says whether the agent actually
|
|
22
|
+
receives it. Skeletons live in node bodies and `ghost pull` emits them last.
|
|
23
|
+
Files behind `materials` arrive only when pulled or inspected. Material the
|
|
24
|
+
agent never reads contributes nothing.
|
|
25
|
+
|
|
26
|
+
Anything that must never be missed belongs in prose or a Skeleton. Anything
|
|
27
|
+
behind `materials` must repay the inspection turn: explicit, skimmable, and
|
|
28
|
+
contract-first.
|
|
29
|
+
|
|
30
|
+
## Ship the syntax the agent emits
|
|
31
|
+
|
|
32
|
+
Every transform between what the agent reads and writes is an error
|
|
33
|
+
opportunity. If it reads `color.background.default: "#fafafa"` in nested JSON
|
|
34
|
+
but must write `var(--background)`, the model must reconstruct the mapping under
|
|
35
|
+
attention pressure. Near-miss names are priors filling that gap.
|
|
36
|
+
|
|
37
|
+
When a build pipeline owns the source of truth, point `materials` at the output
|
|
38
|
+
the agent writes against, not an upstream representation. Name each file;
|
|
39
|
+
glob patterns are invalid because a live repo can make them capture unintended
|
|
40
|
+
content.
|
|
41
|
+
|
|
42
|
+
## Choose only what helps
|
|
43
|
+
|
|
44
|
+
**Tokens are the visual floor.** Use one flat, contract-first file in the
|
|
45
|
+
emission syntax. For CSS, lead with a comment such as `/* style only with these
|
|
46
|
+
custom properties; never hardcode colors or timing. */`, then group flat
|
|
47
|
+
`--name: value` pairs by role. In a utility codebase, name the emitted classes
|
|
48
|
+
in prose and point to the built stylesheet. Avoid markdown token tables and
|
|
49
|
+
nested source JSON when the output uses neither.
|
|
50
|
+
|
|
51
|
+
**Skeletons are earned by composition-critical openings.** Put exactly one
|
|
52
|
+
fenced block under `## Skeleton`, in the target medium, with real token or class
|
|
53
|
+
names and `{placeholders}` for task facts. Include only enough structure to
|
|
54
|
+
commit the shape. A Skeleton in a materials file forfeits guaranteed delivery.
|
|
55
|
+
|
|
56
|
+
**Components are earned by repeated, diverse surfaces.** The source is the API
|
|
57
|
+
reference. Point to each component file explicitly and keep prose to usage
|
|
58
|
+
grammar: purpose, reach-when, neighboring alternative, and never. Do not copy
|
|
59
|
+
props into markdown. For a class vocabulary, lead each class block with a
|
|
60
|
+
one-line contract comment so a skimming agent can recover the grammar.
|
|
61
|
+
|
|
62
|
+
Examples are ordinary materials. When a complete runnable surface helps, point
|
|
63
|
+
to it and follow the example guidance in [nodes.md](nodes.md). Name the file for
|
|
64
|
+
the shape it shows, such as `composition.form.html`, not for arbitrary content.
|
|
65
|
+
|
|
66
|
+
Absence can be the correct stance. A package that asks agents to compose fresh
|
|
67
|
+
from tokens may reject a component kit. A package that values variation may use
|
|
68
|
+
few Skeletons. The reason for absence must be explicit.
|
|
69
|
+
|
|
70
|
+
## Bind a component library
|
|
71
|
+
|
|
72
|
+
An OSS component library should ship its own `.ghost/` packet beside the code,
|
|
73
|
+
then let adopters copy and own both. The packet is a taste floor, not the
|
|
74
|
+
consumer's brand.
|
|
75
|
+
|
|
76
|
+
1. Vendor or install the components and packet together.
|
|
77
|
+
2. Repoint every `materials` locator to the receiving repo's exact files.
|
|
78
|
+
3. State in the cover that local brand guidance overrides the vendored floor.
|
|
79
|
+
4. Give prose only to components whose purpose or divergence is not generic.
|
|
80
|
+
5. Put token and theming invariants in contract nodes.
|
|
81
|
+
6. Add checks only for countable, review-critical invariants.
|
|
82
|
+
7. Run `ghost validate`; dead-locator warnings identify paths that moved.
|
|
83
|
+
|
|
84
|
+
For a familiar library, the model may already know the API. Guidance still
|
|
85
|
+
matters for restraint and choice: which variant when, what not to combine, and
|
|
86
|
+
which neighboring component fits instead. For a novel library, the component
|
|
87
|
+
source supplies missing API knowledge; prose supplies the usage decision.
|
|
88
|
+
|
|
89
|
+
Do not add a component schema, registry mirror, or prop manifest to ghost. Those
|
|
90
|
+
copies drift. `materials` locates implementation; the node body explains why
|
|
91
|
+
and when.
|
|
92
|
+
|
|
93
|
+
## Bundle or reference
|
|
94
|
+
|
|
95
|
+
Put brand-owned artifacts that must travel through export or survive refactors
|
|
96
|
+
under `.ghost/materials/`: token output, logos, type files, motion data, and
|
|
97
|
+
portable examples. Reference living components, stories, tests, and styles at
|
|
98
|
+
their repository paths. Guidance stays in prose in both cases.
|
|
99
|
+
|
|
100
|
+
Use external locators when the authoritative material remains external. Add a
|
|
101
|
+
short `note` only when the locator itself does not tell the agent what it will
|
|
102
|
+
find.
|
|
103
|
+
|
|
104
|
+
## Drift bill
|
|
105
|
+
|
|
106
|
+
Concrete material repeats decisions from the prose. After changing guidance,
|
|
107
|
+
tokens, or component contracts, sweep Skeletons, components, examples, and
|
|
108
|
+
checks for stranded names or literals. Delete any copy whose maintenance cost
|
|
109
|
+
exceeds its steering value.
|
|
110
|
+
|
|
111
|
+
`ghost pull` inlines each distinct local material once per pull. Later nodes
|
|
112
|
+
keep a pointer to the first copy, so sharing a material across nodes is safe and
|
|
113
|
+
does not inflate its salience.
|
|
114
|
+
|
|
115
|
+
## Concrete self-check
|
|
116
|
+
|
|
117
|
+
1. Where does the emitted value vocabulary live?
|
|
118
|
+
2. What commits the most composition-critical opening?
|
|
119
|
+
3. Which reusable decisions are compressed into source-backed names?
|
|
120
|
+
4. What material shows decisions working together when prose is not enough?
|
|
121
|
+
5. Will the agent actually receive or inspect each material?
|
|
122
|
+
6. Is each material in the syntax the agent will emit?
|
|
123
|
+
7. Which copy goes stale when the underlying decision changes?
|
|
124
|
+
|
|
125
|
+
Any answer may be "absent, because." An unexplained absence is the gap.
|
|
126
|
+
|
|
127
|
+
## Never
|
|
128
|
+
|
|
129
|
+
- Never point materials at globs, directories, or generated junk.
|
|
130
|
+
- Never duplicate component APIs or pipeline source formats in prose.
|
|
131
|
+
- Never add concrete material only because a design-system convention says to.
|
|
@@ -38,7 +38,7 @@ selects, inspects, makes, renders, judges, and repairs in the same session.
|
|
|
38
38
|
- read inlined text materials;
|
|
39
39
|
- open referenced source, token, or component files;
|
|
40
40
|
- view image inspect-pointers instead of relying on filenames;
|
|
41
|
-
- inspect rendered
|
|
41
|
+
- inspect rendered examples, not just their contexts;
|
|
42
42
|
- use an available host connection for an external locator only when inspecting
|
|
43
43
|
it could materially affect the task;
|
|
44
44
|
- let the host run its normal authentication and permission flow;
|
|
@@ -51,10 +51,9 @@ selects, inspects, makes, renders, judges, and repairs in the same session.
|
|
|
51
51
|
- never modify an external resource unless the user explicitly asks;
|
|
52
52
|
- record remote, oversized, missing, or unreadable materials;
|
|
53
53
|
- never claim material grounding for something you did not inspect.
|
|
54
|
-
6. **
|
|
55
|
-
|
|
56
|
-
|
|
57
|
-
universal visual target.
|
|
54
|
+
6. **Follow the example instructions.** Keep what the node says to preserve,
|
|
55
|
+
change what belongs to the task, and use the example only in the situation it
|
|
56
|
+
covers. Do not turn one example into a universal target.
|
|
58
57
|
7. **Brief the work.** Follow [brief.md](brief.md). The brief is a steering step,
|
|
59
58
|
not a report. Keep the five-section shape, cap non-negotiables at five, and
|
|
60
59
|
keep ghost-backed claims separate from provisional local reasoning.
|
|
@@ -0,0 +1,138 @@
|
|
|
1
|
+
---
|
|
2
|
+
name: nodes
|
|
3
|
+
description: Write durable ghost guidance with discriminating context, purposeful prose, conditions, replacements, and useful examples.
|
|
4
|
+
---
|
|
5
|
+
|
|
6
|
+
# Recipe: Write Guidance Nodes
|
|
7
|
+
|
|
8
|
+
**Goal:** write the smallest node that forces a real brand decision. Package
|
|
9
|
+
shape belongs in [schema.md](schema.md); this recipe owns the guidance itself.
|
|
10
|
+
|
|
11
|
+
## One node, one purpose
|
|
12
|
+
|
|
13
|
+
A node is one coherent decision with one applicability. Split only when a body
|
|
14
|
+
contains another decision that should be gathered in a different situation.
|
|
15
|
+
Do not split by destination or component name.
|
|
16
|
+
|
|
17
|
+
Use `context` as retrieval payload, not summary. State the observable situation
|
|
18
|
+
in which the node applies. Read it alone: if it fits every brand or almost every
|
|
19
|
+
task, it will not help selection. Put what to do and why in the body.
|
|
20
|
+
|
|
21
|
+
Altitude lives in prose:
|
|
22
|
+
|
|
23
|
+
- State universal guidance plainly.
|
|
24
|
+
- Give narrower guidance a condition that names when it holds.
|
|
25
|
+
- Name the reversal condition when a decision can legitimately flip.
|
|
26
|
+
- Do not encode scope through folders, hierarchy, inheritance, or edges.
|
|
27
|
+
|
|
28
|
+
While drafting, ask three questions. They are prompts, not fields:
|
|
29
|
+
|
|
30
|
+
1. **Why:** what stance or tradeoff forces the choice?
|
|
31
|
+
2. **With what:** which concrete materials let the agent act on it?
|
|
32
|
+
3. **How assembled:** which relationships or opening structure must hold?
|
|
33
|
+
|
|
34
|
+
A node may answer one or all three. Put locators in `materials`; keep
|
|
35
|
+
interpretation in prose.
|
|
36
|
+
|
|
37
|
+
## Write decisions, not brand-deck language
|
|
38
|
+
|
|
39
|
+
A useful sentence rejects a plausible alternative. "We value clarity and
|
|
40
|
+
trust" does not steer. "State what remains safe before asking the person to try
|
|
41
|
+
again" does.
|
|
42
|
+
|
|
43
|
+
Keep real absolutes. `Never` and `always` are correct for human-ratified hard
|
|
44
|
+
lines. Elsewhere, remove filler and unchosen hedges: "elevate," "delight,"
|
|
45
|
+
"seamless," "best-in-class," "generally," "where possible," and "consider."
|
|
46
|
+
If the human has not picked a side, return to authoring rather than laundering
|
|
47
|
+
uncertainty into prose.
|
|
48
|
+
|
|
49
|
+
## Patterns bind and open
|
|
50
|
+
|
|
51
|
+
A pattern fixes part of a reusable structure and leaves the rest available:
|
|
52
|
+
|
|
53
|
+
- **Bound:** what must not be redecided, such as what appears first, exactly
|
|
54
|
+
once, or never competes for attention.
|
|
55
|
+
- **Open:** where the agent may choose within limits, such as evidence form,
|
|
56
|
+
secondary content, or tone within a range.
|
|
57
|
+
|
|
58
|
+
Everything bound is a template. Nothing bound is vibes. Use a Skeleton only
|
|
59
|
+
when the opening structure itself must arrive verbatim; see
|
|
60
|
+
[concrete.md](concrete.md).
|
|
61
|
+
|
|
62
|
+
## Anti-goals replace
|
|
63
|
+
|
|
64
|
+
A strong anti-goal says **not X; instead Y; recognize the switch by Z**.
|
|
65
|
+
Negation alone makes the rejected default more salient without committing the
|
|
66
|
+
replacement.
|
|
67
|
+
|
|
68
|
+
```markdown
|
|
69
|
+
---
|
|
70
|
+
context: Building or reviewing a performance dashboard.
|
|
71
|
+
---
|
|
72
|
+
|
|
73
|
+
Not rounded cards, celebratory gradients, and a wall of equal metrics.
|
|
74
|
+
Instead use one accountable fact, its evidence, and one priced next action on a
|
|
75
|
+
flat surface. Recognize the switch: removing the logo should not leave a generic
|
|
76
|
+
SaaS dashboard behind.
|
|
77
|
+
```
|
|
78
|
+
|
|
79
|
+
Purge the rejected pattern from examples and starter structures. Use checks to
|
|
80
|
+
catch the hard regression; do not make anti-goal prose carry review alone.
|
|
81
|
+
|
|
82
|
+
## Explain examples
|
|
83
|
+
|
|
84
|
+
If a node includes or points to an example, say what the agent should keep and
|
|
85
|
+
what should change with the task. Name the situation it fits when needed. If
|
|
86
|
+
you cannot give those instructions, leave the example out.
|
|
87
|
+
|
|
88
|
+
## Reusable blocks earn prose by purpose
|
|
89
|
+
|
|
90
|
+
Generic primitives need no node merely because they exist. A primitive earns
|
|
91
|
+
prose only for its brand-specific divergence. A reusable block earns a node when
|
|
92
|
+
it encodes a distinguishable user moment or arrangement.
|
|
93
|
+
|
|
94
|
+
For a reusable block, write a short body in this rhythm:
|
|
95
|
+
|
|
96
|
+
- **For:** the user need, not the widget.
|
|
97
|
+
- **Reach when:** the user's first question.
|
|
98
|
+
- **Not when:** the neighboring purpose to choose instead.
|
|
99
|
+
- **Never:** the misuse that stretches it beyond its job.
|
|
100
|
+
|
|
101
|
+
Two blocks that answer the same first question are one purpose, not two nodes.
|
|
102
|
+
Keep props and API reference out of the body. The implementation is swappable;
|
|
103
|
+
the purpose is durable.
|
|
104
|
+
|
|
105
|
+
## Drafting gate
|
|
106
|
+
|
|
107
|
+
Before curation, ask:
|
|
108
|
+
|
|
109
|
+
| Dimension | Question |
|
|
110
|
+
| --- | --- |
|
|
111
|
+
| Testimony | Can you name the human words or evidence behind this? |
|
|
112
|
+
| Discrimination | Does the context select a real situation rather than a topic? |
|
|
113
|
+
| Force | Does the body decide something and reject a plausible alternative? |
|
|
114
|
+
| Altitude | Is it universal on purpose, or conditioned? |
|
|
115
|
+
| Residue | Is it free of starter prose, API mirroring, and brand-deck filler? |
|
|
116
|
+
|
|
117
|
+
A weak answer returns to the human or evidence. Scores do not make guidance
|
|
118
|
+
canonical; human curation does.
|
|
119
|
+
|
|
120
|
+
## Match the form to the failure
|
|
121
|
+
|
|
122
|
+
| If the agent keeps... | Author... |
|
|
123
|
+
| --- | --- |
|
|
124
|
+
| missing guidance | sharper `context`; universal guidance may belong on the cover |
|
|
125
|
+
| inventing values | a material-backed node with exact vocabulary |
|
|
126
|
+
| producing generic output | replacement anti-goal plus a well-explained example |
|
|
127
|
+
| choosing the wrong structure | bound/open pattern and, when needed, a Skeleton |
|
|
128
|
+
| crossing a hard line | invariant prose plus a review check |
|
|
129
|
+
| applying guidance too broadly | a condition or reversal condition |
|
|
130
|
+
| making the wrong tradeoff | a decision trace with the losing alternative |
|
|
131
|
+
|
|
132
|
+
## Never
|
|
133
|
+
|
|
134
|
+
- Never write a node the human neither said, showed, nor accepted.
|
|
135
|
+
- Never make a node a container for observations or implementation inventory.
|
|
136
|
+
- Never duplicate API documentation unless the API itself is the guidance.
|
|
137
|
+
- Never use a broad context to compensate for unrelated decisions in one body.
|
|
138
|
+
- Never ship a blacklist-only anti-goal.
|
|
@@ -1,12 +1,10 @@
|
|
|
1
1
|
---
|
|
2
2
|
name: schema
|
|
3
|
-
description: The ghost package
|
|
3
|
+
description: The deterministic ghost package contract: layout, manifest, nodes, materials, Skeletons, checks, and command behavior.
|
|
4
4
|
---
|
|
5
5
|
|
|
6
6
|
# ghost Package Reference
|
|
7
7
|
|
|
8
|
-
Canonical package:
|
|
9
|
-
|
|
10
8
|
```text
|
|
11
9
|
.ghost/
|
|
12
10
|
manifest.yml ghost.package/v1: schema + id + optional cover
|
|
@@ -18,28 +16,31 @@ Canonical package:
|
|
|
18
16
|
```
|
|
19
17
|
|
|
20
18
|
Reserved at the root: `manifest.yml`, `glossary.md`, `materials/`, and
|
|
21
|
-
`checks/`. Every other `*.md` is a node.
|
|
19
|
+
`checks/`. Every other `*.md` is a node. The corpus is flat: no hierarchy,
|
|
20
|
+
inheritance, or edges.
|
|
22
21
|
|
|
23
22
|
## Manifest
|
|
24
23
|
|
|
25
24
|
`manifest.yml` declares `schema`, `id`, and optionally `cover`. `cover` is a
|
|
26
25
|
node id. When present and resolved, `ghost gather` inlines that node above the
|
|
27
|
-
menu on every invocation.
|
|
28
|
-
essence, temperature, and brand-only refusals.
|
|
26
|
+
menu on every invocation.
|
|
29
27
|
|
|
30
|
-
`ghost validate`
|
|
31
|
-
|
|
32
|
-
budget (1500 bytes) is a warning.
|
|
28
|
+
`ghost validate` reports a missing referenced cover as an error, an undeclared
|
|
29
|
+
cover as a warning, and a cover body over 1500 bytes as a warning.
|
|
33
30
|
|
|
34
|
-
##
|
|
31
|
+
## Glossary and identity
|
|
32
|
+
|
|
33
|
+
`glossary.md` declares the package's kind vocabulary and defines each kind. A
|
|
34
|
+
node's id is its filename minus `.md`; its kind is the first dotted filename
|
|
35
|
+
segment. A bare filename has no kind. Undeclared kind prefixes warn.
|
|
35
36
|
|
|
36
|
-
|
|
37
|
+
## Nodes
|
|
37
38
|
|
|
38
39
|
```markdown
|
|
39
40
|
---
|
|
40
41
|
context: Placing, sizing, or choosing a logo lockup or glyph.
|
|
41
42
|
materials:
|
|
42
|
-
- brand/logo
|
|
43
|
+
- brand/logo-primary.svg
|
|
43
44
|
- https://figma.com/file/example?node-id=logo-lockups
|
|
44
45
|
- locator: mcp://brand-assets/logo-lockups
|
|
45
46
|
note: Source lockups and glyph exports
|
|
@@ -48,36 +49,26 @@ materials:
|
|
|
48
49
|
Use the full lockup when recognition matters.
|
|
49
50
|
```
|
|
50
51
|
|
|
51
|
-
-
|
|
52
|
-
-
|
|
53
|
-
|
|
54
|
-
|
|
55
|
-
|
|
56
|
-
|
|
57
|
-
|
|
58
|
-
|
|
59
|
-
|
|
60
|
-
- `materials` accepts repo-relative file paths plus supported external locators
|
|
61
|
-
using `https:`, `mcp:`, `figma:`, or `github:`. Name each file explicitly:
|
|
62
|
-
glob patterns are not supported and fail validation, because in a live
|
|
63
|
-
repo a glob can capture unintended files into pulls. Items may be bare
|
|
64
|
-
locator strings or `{ locator, note }` objects. Use a short `note` only when
|
|
65
|
-
an opaque locator needs retrieval context. The external locator tells the
|
|
66
|
-
host how to connect; ghost does not fetch or authenticate. The list locates
|
|
67
|
-
material, while guidance stays in the node body.
|
|
68
|
-
|
|
69
|
-
ghost derives whether a node carries concrete material from structure:
|
|
70
|
-
non-empty `materials`, a fenced code block of at least 3 lines, or a
|
|
71
|
-
`## Skeleton` section. `gather` reports these payload labels for clarity; they
|
|
72
|
-
are not ranking signals.
|
|
73
|
-
|
|
74
|
-
## Skeleton convention
|
|
75
|
-
|
|
76
|
-
A `## Skeleton` section contains the literal opening structure for a surface.
|
|
77
|
-
It should contain exactly one fenced block; `ghost validate` warns, never fails,
|
|
78
|
-
when a Skeleton section has zero or multiple fences.
|
|
52
|
+
- `context` is the retrieval payload shown by `ghost gather`.
|
|
53
|
+
- `description` remains a deprecated read alias for one release; validation
|
|
54
|
+
warns until it is renamed.
|
|
55
|
+
- `materials` accepts explicit repo-relative file paths and external locators
|
|
56
|
+
using `https:`, `mcp:`, `figma:`, or `github:`.
|
|
57
|
+
- Glob patterns are invalid. Each local file must be named explicitly.
|
|
58
|
+
- A material may be a bare locator or `{ locator, note }`.
|
|
59
|
+
- External locators describe access; ghost does not fetch or authenticate.
|
|
60
|
+
- Frontmatter may contain additional descriptive keys. Guidance stays in prose.
|
|
79
61
|
|
|
80
|
-
|
|
62
|
+
A node is concrete when it has non-empty `materials`, a fenced code block of at
|
|
63
|
+
least three lines, or a `## Skeleton` section. `gather` reports those payload
|
|
64
|
+
labels; they are not ranking signals.
|
|
65
|
+
|
|
66
|
+
## Skeletons
|
|
67
|
+
|
|
68
|
+
A `## Skeleton` section contains literal opening structure. It must contain
|
|
69
|
+
exactly one fenced block; zero or multiple fences warn.
|
|
70
|
+
|
|
71
|
+
````markdown
|
|
81
72
|
## Skeleton
|
|
82
73
|
|
|
83
74
|
```tsx
|
|
@@ -86,10 +77,10 @@ when a Skeleton section has zero or multiple fences.
|
|
|
86
77
|
<button>{nextStep}</button>
|
|
87
78
|
</section>
|
|
88
79
|
```
|
|
89
|
-
|
|
80
|
+
````
|
|
90
81
|
|
|
91
|
-
`ghost pull`
|
|
92
|
-
|
|
82
|
+
`ghost pull` removes Skeleton sections from node bodies and emits their fences
|
|
83
|
+
last under the begin-from-this-structure banner.
|
|
93
84
|
|
|
94
85
|
## Checks
|
|
95
86
|
|
|
@@ -107,16 +98,18 @@ references:
|
|
|
107
98
|
Grade whether the change preserves the logo guidance in `asset.logo`.
|
|
108
99
|
```
|
|
109
100
|
|
|
110
|
-
`references`
|
|
111
|
-
review instructions for the host agent
|
|
112
|
-
|
|
113
|
-
|
|
114
|
-
##
|
|
115
|
-
|
|
116
|
-
- `ghost gather` emits the cover
|
|
117
|
-
|
|
118
|
-
- `ghost pull` emits selected nodes in steering order
|
|
119
|
-
materials
|
|
120
|
-
|
|
121
|
-
- `ghost review` matches
|
|
122
|
-
checks, and emits
|
|
101
|
+
`references` contains node ids with optional heading anchors. Check bodies are
|
|
102
|
+
review instructions for the host agent. ghost validates and transports checks;
|
|
103
|
+
it does not grade them.
|
|
104
|
+
|
|
105
|
+
## Command behavior
|
|
106
|
+
|
|
107
|
+
- `ghost gather` emits the cover, coverage counts, then a complete, unfiltered,
|
|
108
|
+
unranked node menu. Checks are absent.
|
|
109
|
+
- `ghost pull` emits selected nodes in steering order, inlines eligible local
|
|
110
|
+
text materials once, leaves later duplicate pointers, turns binary materials
|
|
111
|
+
into inspect-pointers, and leaves external materials as locators.
|
|
112
|
+
- `ghost review` matches touched files to exact local material paths, offers
|
|
113
|
+
relevant checks, and emits an advisory packet for the host agent.
|
|
114
|
+
- `ghost export` bundles the package and audits which locators travel.
|
|
115
|
+
- `ghost pulse` summarizes local gather and pull events.
|
|
@@ -35,7 +35,6 @@ Before generating, can you cite:
|
|
|
35
35
|
- the governing principle, stance, or tradeoff?
|
|
36
36
|
- the concrete materials, if exactness matters?
|
|
37
37
|
- the applicable pattern, if structure matters?
|
|
38
|
-
- the relevant exemplar and what it is normative for?
|
|
39
38
|
- the anti-goal that blocks the generic version?
|
|
40
39
|
- the hard invariants?
|
|
41
40
|
- the conditions that apply or do not apply?
|
|
@@ -37,24 +37,15 @@ Report first:
|
|
|
37
37
|
| Row | Status | Evidence | Next move |
|
|
38
38
|
| --- | --- | --- | --- |
|
|
39
39
|
| Retrieval | strong / weak | contexts, ids, cover | sharpen contexts or move universal guidance to the cover |
|
|
40
|
-
| Concreteness | strong / thin | materials, fenced examples, Skeletons | add concrete locators,
|
|
40
|
+
| Concreteness | strong / thin | materials, fenced examples, Skeletons | add concrete locators, examples, or opening structures |
|
|
41
41
|
| Anti-goals | present / missing / vague | `anti-goal.*`, review packet | write not-X-instead-Y replacements and material locators |
|
|
42
|
-
| Consistency | clean / conflicting |
|
|
42
|
+
| Consistency | clean / conflicting | guidance vs concrete material | update or remove stale material |
|
|
43
43
|
| Stance | present / missing | cover, `principle.*` | write forced-choice principles |
|
|
44
44
|
| Materials | present / missing | `materials`, inspect-pointers | point at real assets/components/tokens |
|
|
45
|
-
| Exemplars | annotated / unannotated / converging / missing | fenced samples, screenshots, repeated runs | say what to copy and what is incidental; tighten applicability when unrelated tasks converge |
|
|
46
45
|
| Patterns | bound-open / loose / missing | `pattern.*`, Skeletons | state applies / bound / open and add a Skeleton when opening structure matters |
|
|
47
46
|
| Checks | covered / partial / missing | checks/, review packet | add checks for high-risk invariants |
|
|
48
47
|
| Silence posture | defined / missing | cover | say when to proceed provisionally or ask |
|
|
49
48
|
|
|
50
|
-
## Exemplar convergence
|
|
51
|
-
|
|
52
|
-
When auditing repeated runs, compare unrelated asks that pulled the same
|
|
53
|
-
exemplar. If they reproduce the same composition, section order, or decorative
|
|
54
|
-
treatment without the exemplar's stated condition requiring it, the exemplar is
|
|
55
|
-
steering beyond its intent. The next move is to tighten applicability, add a
|
|
56
|
-
shape-appropriate exemplar, or remove the exemplar.
|
|
57
|
-
|
|
58
49
|
## Task-level readiness
|
|
59
50
|
|
|
60
51
|
For a task, gather, pull, and report:
|
package/package.json
CHANGED
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
{
|
|
2
2
|
"name": "@design-intelligence/ghost",
|
|
3
|
-
"version": "0.
|
|
3
|
+
"version": "0.29.0",
|
|
4
4
|
"description": "ghost CLI for portable .ghost/ packages of brand guidance your agent reads before it builds, with advisory review after",
|
|
5
5
|
"license": "Apache-2.0",
|
|
6
6
|
"author": "Block, Inc.",
|
|
@@ -1,110 +0,0 @@
|
|
|
1
|
-
---
|
|
2
|
-
name: adapting-a-starter
|
|
3
|
-
description: Transplant a starter ghost package — a body like vessel-light or the naked skeleton — into your own brand, one procedure from manifest id to regenerated refs.
|
|
4
|
-
handoffs:
|
|
5
|
-
- label: Validate the adapted package
|
|
6
|
-
command: ghost validate --format json
|
|
7
|
-
prompt: Does the adapted package validate, and did any pruned rule orphan a check reference?
|
|
8
|
-
---
|
|
9
|
-
|
|
10
|
-
# Recipe: Adapt A Starter Package
|
|
11
|
-
|
|
12
|
-
**Goal:** turn an installed starter — a full body (`ghost init --body
|
|
13
|
-
vessel-light`) or the naked skeleton (`ghost init`) — into *your* brand's
|
|
14
|
-
brand guidance without shipping a self-contradicting package.
|
|
15
|
-
|
|
16
|
-
The naked skeleton is factored by what must happen during adaptation:
|
|
17
|
-
|
|
18
|
-
| Stratum | Files | On adaptation |
|
|
19
|
-
| --- | --- | --- |
|
|
20
|
-
| Cover | `brand.md` | Write it — essence in one paragraph, temperature, and brand-only refusals. Apply the admission test to every line: a violation visible in one element belongs in that element's chapter; visible in one view, in the composition chapter; visible only across the whole body of work, here. Keep a one-screen budget and delete every scaffolding sentence when real. |
|
|
21
|
-
| Foundation | `foundation.*` | Answer each chapter's open section (`Palette`, `Typeface`, `Radius`, `Character`, `Warmth`), restate it as the brand's current answer, and edit misuse lists to the brand's real failure modes. |
|
|
22
|
-
| Context | `context.*` | Add or re-tune after the open questions are answered; context nodes state only what inverts in a matching situation. |
|
|
23
|
-
| Cliche floor | `cliche.*` | Prune — delete entries a foundation misuse list absorbs, delete the paired check flag with it, and replace generic entries with refusals only this brand makes. If a refusal passes the cover admission test, graduate it to `brand.md`. |
|
|
24
|
-
| Materials and refs | `materials/*`, referenced implementations, refs | Add or regenerate so the prose has concrete material to inspect. |
|
|
25
|
-
| Checks | `checks/*` | Keep paired with the nodes they enforce; delete or rewrite flags when their source rule changes. |
|
|
26
|
-
|
|
27
|
-
Do the steps **in order and in one sitting** where possible. A half-adapted
|
|
28
|
-
package is worse than an unadapted one: stale refs steer harder than any prose
|
|
29
|
-
you rewrote, so stopping before refs and checks are current ships a package
|
|
30
|
-
that contradicts itself.
|
|
31
|
-
|
|
32
|
-
## The procedure
|
|
33
|
-
|
|
34
|
-
1. **Change the manifest id.** Edit `id:` in `manifest.yml` to your brand's
|
|
35
|
-
name. This is deliberately first: it is the explicit act that marks the
|
|
36
|
-
adaptation as begun. Until it changes, the package honestly claims to be the
|
|
37
|
-
starter, and every consuming agent cites it as a starter default.
|
|
38
|
-
2. **Write the cover.** Replace `brand.md` with the brand's essence in one
|
|
39
|
-
paragraph, the shared temperature of words and motion, and the refusals only
|
|
40
|
-
this brand makes. Test every line: a violation visible in one element moves
|
|
41
|
-
to that element's chapter; visible in one view, to the composition chapter;
|
|
42
|
-
only what shows solely across the whole body of work stays. Keep the cover
|
|
43
|
-
to one screen and delete every scaffolding sentence when it is real.
|
|
44
|
-
3. **Answer the foundation chapters.** Walk each `foundation.*` node. Keep the
|
|
45
|
-
usage rules that describe the closed vocabulary; answer the open section as
|
|
46
|
-
the brand's current answer. Ask the human for undecided values; never
|
|
47
|
-
freehand a value and present it as brand-backed. Edit each misuse list to the
|
|
48
|
-
failures this brand actually needs to avoid.
|
|
49
|
-
4. **Add or edit materials.** Record literal values where the implementation or
|
|
50
|
-
asset materials live, then add `materials` locators to the nodes that explain
|
|
51
|
-
them. Name each file explicitly; do not reach for glob patterns, which in a
|
|
52
|
-
live repo can capture unintended files into pulls. Role names can stay
|
|
53
|
-
stable while values change. The point is that a realizing agent can inspect
|
|
54
|
-
the same concrete source the prose governs. Starter and vendored packages
|
|
55
|
-
ship locators written against the source repo's layout, so after vendoring,
|
|
56
|
-
verify every locator resolves against this repo's actual paths, repoint the
|
|
57
|
-
ones that moved, and run `ghost validate` — `material-locator-dead` warnings
|
|
58
|
-
name the stale ones.
|
|
59
|
-
5. **Re-tune contexts.** Revisit each `context.*` node after the foundation
|
|
60
|
-
answers exist. Keep only inversions that apply in that situation; delete
|
|
61
|
-
generic foundation rules repeated there, and add new contexts only when a
|
|
62
|
-
task situation truly bends the defaults.
|
|
63
|
-
6. **Prune `cliche.median`.** Each rule is a `###` heading section; delete the
|
|
64
|
-
whole section for every rule a foundation misuse list now absorbs or your
|
|
65
|
-
brand legitimately violates. Do not rewrite surviving measured defaults into
|
|
66
|
-
taste. Then run `ghost validate`: every check reference orphaned by a pruned
|
|
67
|
-
heading surfaces as its own warning — delete the paired flag and its
|
|
68
|
-
reference from the check.
|
|
69
|
-
7. **Regenerate refs.** Exemplars dominate prose: a prose rule contradicted by a
|
|
70
|
-
stale ref loses. Rebuild each ref against the new values and chapter answers,
|
|
71
|
-
keep any annotation headers current, and make each ref demonstrate its closed
|
|
72
|
-
sets completely. If the starter shipped no refs, generate them now — a
|
|
73
|
-
package with no exemplars steers at half strength.
|
|
74
|
-
8. **Rewrite checks and near-miss nodes.** For each answered foundation chapter,
|
|
75
|
-
name the failure mode one step away from the answer and its replacement. Put
|
|
76
|
-
pre-generation guidance in nodes and review assertions in `checks/`; never
|
|
77
|
-
leave a check that asserts an obligation no node states.
|
|
78
|
-
9. **Re-run the checks, including against refs.** `ghost validate` for package
|
|
79
|
-
shape, then review the regenerated refs against the median and value checks
|
|
80
|
-
(stage the ref changes and run `ghost review`). The floor only holds if the
|
|
81
|
-
refs hold it too: a ref that trips a median flag will teach every future
|
|
82
|
-
generation the violation.
|
|
83
|
-
|
|
84
|
-
## Consuming an unadapted starter
|
|
85
|
-
|
|
86
|
-
Work does not block on adaptation. Before the procedure runs (or midway through
|
|
87
|
-
it), cite starter content honestly:
|
|
88
|
-
|
|
89
|
-
- Cover scaffolding and unanswered foundation sections: **provisional** — never
|
|
90
|
-
plain brand guidance.
|
|
91
|
-
- Foundation usage rules: **ghost-backed starter structure** until the human
|
|
92
|
-
answers the open sections and edits the misuse lists.
|
|
93
|
-
- Surviving median rules: **owner-backed after init** — ghost stamps this
|
|
94
|
-
measured model behavior into every initialized package; you own the pruning and
|
|
95
|
-
any adaptation thereafter.
|
|
96
|
-
- Context nodes: **conditional** — read only when their situation matches, and
|
|
97
|
-
revisit them after the foundation answers change.
|
|
98
|
-
- The manifest id tells you which state you are in: a starter id means
|
|
99
|
-
unadapted.
|
|
100
|
-
|
|
101
|
-
## Never
|
|
102
|
-
|
|
103
|
-
- Never leave cover scaffolding in a real ghost package.
|
|
104
|
-
- Never present your provisional value as the brand's answer.
|
|
105
|
-
- Never leave stale refs alongside new values — regenerate or delete; a stale
|
|
106
|
-
exemplar outweighs rewritten prose.
|
|
107
|
-
- Never prune a median rule without deleting its paired check flag, or keep a
|
|
108
|
-
check that asserts an obligation no node states.
|
|
109
|
-
- Never pre-write the manifest id change into automation — it is the human's
|
|
110
|
-
act of ownership, the one step an agent should not take alone.
|