reviewing-instructions

Name: reviewing-instructions
Author: alexei-led

3installs5forks

How to install reviewing-instructions?

alexei-led/cc-thingz/reviewing-instructions

$npx -y skills add alexei-led/cc-thingz --skill reviewing-instructions

Installs into the current project.

Summary

The host SKILL.md is canonical. This overlay only adds Claude-specific argument, task, and aggregation behavior.

Files · 1

View on GitHub

SKILL.md

1	# Instruction Review
2
3	## Claude platform additions
4
5	The host SKILL.md is canonical. This overlay only adds Claude-specific argument,
6	task, and aggregation behavior.
7
8	## Argument parsing
9
10	From `$ARGUMENTS`:
11
12	- first non-flag token: file path, directory path, plugin name, or omitted scope
13	- `--model <name>`: override model resolution
14	- `--team`: use parallel review agents for large scopes
15	- `--rerank`: use calibration anchors and pairwise comparison when comparing versions
16
17	## Task use
18
19	Use direct review for one file or one small skill folder.
20
21	Use parallel tasks only when scope contains multiple independent files or plugins.
22	Launch at most 3 review tasks at once. Batch deterministically by sorted path, not
23	by estimated difficulty.
24
25	Each task prompt must include:
26
27	- exact file list
28	- resolved model context or instruction to resolve it
29	- path to `references/scoring-rubric.md`
30	- path to `references/model-resolution.md`
31	- path to `references/skill-architecture.md` when the scope includes `SKILL.md`, `AGENT.md`, `body.md`, or agent-facing references
32	- requirement to cite evidence for every score
33	- requirement to apply gates and caps before final score
34
35	## Aggregation
36
37	When task results return:
38
39	1. Verify each reviewed file was in scope.
40	2. Recompute caps when a task forgot them.
41	3. Deduplicate findings by file plus rule or dimension.
42	4. If two scores differ by more than 1 point, compare gates and caps first.
43	5. Use the lower-confidence score only as a signal; do not average incompatible scopes.
44	6. Put unresolved disagreements in the report as low-confidence notes.
45
46	## Rerank mode
47
48	When `$ARGUMENTS` contains `--rerank`, read `references/calibration.md`.
49	For two versions of a file, compare gates first, then dimensions, then final score.
50	If the difference is less than 0.5, report a tie.
51
52	## Web and model docs
53
54	Use web lookup for model docs only when local references are missing or the user asks
55	for current vendor guidance. If web access fails, use local references or generic
56	context and report the gap.

People who liked this also viewed

6 picks

Type

reviewing-instructions

3installs5forks

View on GitHub

How to install reviewing-instructions?

alexei-led/cc-thingz/reviewing-instructions

$npx -y skills add alexei-led/cc-thingz --skill reviewing-instructions

Installs into the current project.

Summary

The host SKILL.md is canonical. This overlay only adds Claude-specific argument, task, and aggregation behavior.

Files · 1

View on GitHub

SKILL.md

1	# Instruction Review
2
3	## Claude platform additions
4
5	The host SKILL.md is canonical. This overlay only adds Claude-specific argument,
6	task, and aggregation behavior.
7
8	## Argument parsing
9
10	From `$ARGUMENTS`:
11
12	- first non-flag token: file path, directory path, plugin name, or omitted scope
13	- `--model <name>`: override model resolution
14	- `--team`: use parallel review agents for large scopes
15	- `--rerank`: use calibration anchors and pairwise comparison when comparing versions
16
17	## Task use
18
19	Use direct review for one file or one small skill folder.
20
21	Use parallel tasks only when scope contains multiple independent files or plugins.
22	Launch at most 3 review tasks at once. Batch deterministically by sorted path, not
23	by estimated difficulty.
24
25	Each task prompt must include:
26
27	- exact file list
28	- resolved model context or instruction to resolve it
29	- path to `references/scoring-rubric.md`
30	- path to `references/model-resolution.md`
31	- path to `references/skill-architecture.md` when the scope includes `SKILL.md`, `AGENT.md`, `body.md`, or agent-facing references
32	- requirement to cite evidence for every score
33	- requirement to apply gates and caps before final score
34
35	## Aggregation
36
37	When task results return:
38
39	1. Verify each reviewed file was in scope.
40	2. Recompute caps when a task forgot them.
41	3. Deduplicate findings by file plus rule or dimension.
42	4. If two scores differ by more than 1 point, compare gates and caps first.
43	5. Use the lower-confidence score only as a signal; do not average incompatible scopes.
44	6. Put unresolved disagreements in the report as low-confidence notes.
45
46	## Rerank mode
47
48	When `$ARGUMENTS` contains `--rerank`, read `references/calibration.md`.
49	For two versions of a file, compare gates first, then dimensions, then final score.
50	If the difference is less than 0.5, report a tie.
51
52	## Web and model docs
53
54	Use web lookup for model docs only when local references are missing or the user asks
55	for current vendor guidance. If web access fails, use local references or generic
56	context and report the gap.

People who liked this also viewed

6 picks

Type