@mrciphersmith/keryx 0.2.162 → 0.2.163
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/cli.js
CHANGED
|
@@ -53468,7 +53468,7 @@ import path47 from "path";
|
|
|
53468
53468
|
// package.json
|
|
53469
53469
|
var package_default = {
|
|
53470
53470
|
name: "@mrciphersmith/keryx",
|
|
53471
|
-
version: "0.2.
|
|
53471
|
+
version: "0.2.163",
|
|
53472
53472
|
description: "Version-controlled project context for AI coding agents: code graph, architecture wiki, project memory, relevant tests, quality signals, and task flows.",
|
|
53473
53473
|
private: false,
|
|
53474
53474
|
publishConfig: {
|
package/package.json
CHANGED
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
{
|
|
2
2
|
"name": "@mrciphersmith/keryx",
|
|
3
|
-
"version": "0.2.
|
|
3
|
+
"version": "0.2.163",
|
|
4
4
|
"description": "Version-controlled project context for AI coding agents: code graph, architecture wiki, project memory, relevant tests, quality signals, and task flows.",
|
|
5
5
|
"private": false,
|
|
6
6
|
"publishConfig": {
|
|
@@ -0,0 +1,153 @@
|
|
|
1
|
+
---
|
|
2
|
+
description: "Claude-only. Prompting and workflow standards for Claude Opus 5.5 and newer reasoning-by-default Claude models: no reasoning instructions, finish-line task framing, stopping rules, long-run state, blockers-first acceptance."
|
|
3
|
+
alwaysApply: false
|
|
4
|
+
agent_requires: ["claude"]
|
|
5
|
+
---
|
|
6
|
+
|
|
7
|
+
# Opus 5.5 Prompting Standards
|
|
8
|
+
|
|
9
|
+
## Scope — Claude only
|
|
10
|
+
|
|
11
|
+
This rule describes behavior specific to Claude Opus 5.5 and newer reasoning-by-default
|
|
12
|
+
Claude models. It is **not** a general prompting rule.
|
|
13
|
+
|
|
14
|
+
- It is cited from the repository's `CLAUDE.md`, not from `AGENTS.md`, so agents that read
|
|
15
|
+
only `AGENTS.md` never load it.
|
|
16
|
+
- `agent_requires: ["claude"]` records the same restriction in metadata, alongside the
|
|
17
|
+
`stack_requires` convention used by stack-specific rules.
|
|
18
|
+
- Do not apply it to prompts aimed at non-Anthropic models, or at Claude models older than
|
|
19
|
+
4.6 — those may still need explicit reasoning instructions.
|
|
20
|
+
|
|
21
|
+
Source: Anthropic, *Getting the most out of Claude Opus 5.5* —
|
|
22
|
+
https://claude.dev/blog/getting-the-most-out-of-opus-5-5/
|
|
23
|
+
|
|
24
|
+
## Purpose
|
|
25
|
+
|
|
26
|
+
Claude Opus 5.5 reasons before every reply. Prompt patterns written for models that had to be
|
|
27
|
+
*told* to reason are now dead weight: they consume context, add no capability, and can degrade
|
|
28
|
+
output by over-constraining the model's own decomposition.
|
|
29
|
+
|
|
30
|
+
## When To Apply
|
|
31
|
+
|
|
32
|
+
- Writing or reviewing a prompt, system prompt, or tool description aimed at Claude Opus 5.5+.
|
|
33
|
+
- Authoring or editing `CLAUDE.md`, rule files, or skill instructions read by Claude.
|
|
34
|
+
- Composing a subagent dispatch, per `subagent-context-construction.md`.
|
|
35
|
+
- Designing a long-running or agentic workflow (migration, audit, multi-step implementation).
|
|
36
|
+
- Auditing existing instructions for patterns written against older models.
|
|
37
|
+
|
|
38
|
+
## Mandatory Behavior
|
|
39
|
+
|
|
40
|
+
### 1. Delete reasoning instructions
|
|
41
|
+
|
|
42
|
+
The model always thinks before replying. Remove, do not soften:
|
|
43
|
+
|
|
44
|
+
- `think carefully`, `think step by step`, `let's think about this`
|
|
45
|
+
- `рассуждай пошагово`, `подумай внимательно`
|
|
46
|
+
- requests to show, explain, or narrate reasoning as a way of improving quality
|
|
47
|
+
- instructions referencing `<thinking>` tags or a fixed thinking budget
|
|
48
|
+
|
|
49
|
+
Replace each with an explicit completion criterion.
|
|
50
|
+
|
|
51
|
+
### 2. Frame tasks by finish line, not by steps
|
|
52
|
+
|
|
53
|
+
Give the whole task in one message with a stated definition of done. Do not hand-decompose
|
|
54
|
+
work the model decomposes better with full context.
|
|
55
|
+
|
|
56
|
+
Acceptable decomposition: when steps encode a **constraint** — ordering that matters, or a gate
|
|
57
|
+
that must pass before the next action — not when they merely restate the task.
|
|
58
|
+
|
|
59
|
+
### 3. Be specific about what to avoid
|
|
60
|
+
|
|
61
|
+
Name the concrete patterns to avoid. Generic negative direction ("don't make it generic",
|
|
62
|
+
"avoid boilerplate") does not constrain output.
|
|
63
|
+
|
|
64
|
+
### 4. Encode stopping rules in `CLAUDE.md`
|
|
65
|
+
|
|
66
|
+
State where the agent proceeds autonomously and where it stops and asks. Destructive operations
|
|
67
|
+
keep their permission prompts — never disable confirmations to reduce friction.
|
|
68
|
+
|
|
69
|
+
### 5. Externalize state on long runs
|
|
70
|
+
|
|
71
|
+
A long run fills the context window and the early portion is summarized away. Keep progress in a
|
|
72
|
+
file so state survives compaction. In a keryx project that file already exists: flow state is the
|
|
73
|
+
record, and it is written through the `keryx flow` CLI, never by hand.
|
|
74
|
+
|
|
75
|
+
### 6. Delegate breadth, verify evidence
|
|
76
|
+
|
|
77
|
+
For repository-wide audits and migrations, dispatch parallel subagents and verify the evidence
|
|
78
|
+
each returns before accepting it. A subagent's claim is not a result — this is the same
|
|
79
|
+
requirement `subagent-status-protocol.md` encodes in its status contract.
|
|
80
|
+
|
|
81
|
+
### 7. Accept results blockers-first
|
|
82
|
+
|
|
83
|
+
1. Look first for anything the agent is waiting on — blockers outrank the summary.
|
|
84
|
+
2. Request a review pass before reading the diff.
|
|
85
|
+
3. For research and investigation tasks, require unconfirmed findings to be marked as such.
|
|
86
|
+
|
|
87
|
+
### 8. Feed finished artifacts
|
|
88
|
+
|
|
89
|
+
Attach images, spreadsheets, and documents directly rather than describing them. Opus 5.5 reads
|
|
90
|
+
charts, diagrams, and screenshots more accurately than its predecessor; a paraphrase discards
|
|
91
|
+
information the model could have used.
|
|
92
|
+
|
|
93
|
+
### 9. Settle earlier answers in long projects
|
|
94
|
+
|
|
95
|
+
State that previously agreed answers are settled, or the model re-litigates decisions during
|
|
96
|
+
follow-up questions.
|
|
97
|
+
|
|
98
|
+
## Output Contract
|
|
99
|
+
|
|
100
|
+
When this rule drives an edit to prompts or instruction files:
|
|
101
|
+
|
|
102
|
+
- Report each change as `file:line`, the pattern removed, and what replaced it.
|
|
103
|
+
- Instructions controlling **output format** are out of scope — never removed under this rule.
|
|
104
|
+
- Safety, permission, and confirmation rules are out of scope — never removed under this rule.
|
|
105
|
+
- Instructions inside files aimed at non-Claude agents (`AGENTS.md` and its imports) are out of
|
|
106
|
+
scope.
|
|
107
|
+
|
|
108
|
+
## Examples
|
|
109
|
+
|
|
110
|
+
### Reasoning instruction → completion criterion
|
|
111
|
+
|
|
112
|
+
```
|
|
113
|
+
# Before
|
|
114
|
+
Think carefully step by step, then migrate the auth endpoints.
|
|
115
|
+
|
|
116
|
+
# After
|
|
117
|
+
Migrate every auth endpoint to the v2 contract. Done when `bun test auth` passes
|
|
118
|
+
and no route still imports `legacy-auth`.
|
|
119
|
+
```
|
|
120
|
+
|
|
121
|
+
### Hand-decomposition → finish line
|
|
122
|
+
|
|
123
|
+
```
|
|
124
|
+
# Before
|
|
125
|
+
1. Read the schema. 2. List the affected models. 3. Write the migration.
|
|
126
|
+
4. Update the types. 5. Run the tests.
|
|
127
|
+
|
|
128
|
+
# After
|
|
129
|
+
Migrate the schema to add `tenant_id` across all affected models.
|
|
130
|
+
Done when migrations apply cleanly on a fresh database and the type check passes.
|
|
131
|
+
```
|
|
132
|
+
|
|
133
|
+
### Generic direction → named patterns
|
|
134
|
+
|
|
135
|
+
```
|
|
136
|
+
# Before
|
|
137
|
+
Design the settings page, avoid a generic look.
|
|
138
|
+
|
|
139
|
+
# After
|
|
140
|
+
Design the settings page. Avoid: centered hero headings, three-column feature
|
|
141
|
+
grids, gradient-on-card backgrounds, and full-width primary buttons.
|
|
142
|
+
```
|
|
143
|
+
|
|
144
|
+
### Stopping rule in `CLAUDE.md`
|
|
145
|
+
|
|
146
|
+
```markdown
|
|
147
|
+
## Stopping rules
|
|
148
|
+
|
|
149
|
+
- Proceed without asking: reads, searches, test runs, local commits on a feature branch.
|
|
150
|
+
- Stop and ask: force-push, branch deletion, schema migration against a shared database,
|
|
151
|
+
publishing a package, any change under `infra/`.
|
|
152
|
+
- Keep permission prompts enabled for destructive commands.
|
|
153
|
+
```
|