ai-research-reproduction

Name: ai-research-reproduction
Author: lllllllama

161kinstalls10forks

How to install ai-research-reproduction?

lllllllama/rigorpilot-skills/ai-research-reproduction

$npx -y skills add lllllllama/RigorPilot-Skills --skill ai-research-reproduction

Installs into the current project.

Summary

Rigor Reproduce compatible skill slug for README-first deep learning repository reproduction. Use when the user wants an end-to-end, minimal-trustworthy flow that reads the repository first, selects the smallest documented inference or evaluation target, coordinates intake, setup

Files · 1

View on GitHub

SKILL.md

1	# ai-research-reproduction
2
3	## Purpose
4
5	Use this as the Rigor Reproduce compatible skill slug for README-first deep
6	learning repository reproduction. The installed slug remains
7	`ai-research-reproduction` for compatibility. The skill guides the agent toward
8	a minimal trustworthy run with auditable evidence; it should not micromanage
9	implementation details that the model can infer from the repository.
10	Reproduction is not "make it run by changing anything"; it means faithfully
11	reading the README, environment, weights, datasets, and documented commands,
12	then recording results and deviations.
13
14	Start from the shared operating principles in
15	`../../references/agent-operating-principles.md`, then load
16	`../../references/research-rigor-principles.md` and
17	`../../references/deep-learning-experiment-principles.md` when scientific
18	meaning, comparability, or experiment details are at stake.
19
20	## Fit
21
22	Use this skill when all are true:
23
24	- The target is an AI code repository with a README, scripts, configs, or
25	documented commands.
26	- The request spans multiple trusted phases such as intake, setup, execution,
27	training verification, analysis, paper-gap resolution, and reporting.
28	- The desired result is a small reproducible target, not broad experimentation.
29
30	Do not use this skill for paper summaries, generic environment setup, isolated
31	repo scanning, standalone command execution, open-ended research design, or
32	explicit candidate-only exploration.
33
34	## Trusted Target Selection
35
36	Choose the smallest target that can honestly demonstrate repository-grounded
37	reproduction:
38
39	1. documented inference
40	2. documented evaluation
41	3. documented training startup or partial verification
42	4. full training only after explicit user confirmation
43
44	Treat README guidance as the primary reproduction intent. Use repository files
45	to clarify the README, not to silently replace it. When the README and paper
46	conflict, record the conflict and use `paper-context-resolver` only for the
47	narrow reproduction-critical gap.
48
49	## Workflow
50
51	1. Read the README and nearby repo signals.
52	2. Use `repo-intake-and-plan` to extract documented commands and candidate
53	targets.
54	3. Select and justify the minimum trustworthy target.
55	4. Use `env-and-assets-bootstrap` only for target-specific environment,
56	checkpoint, dataset, and cache assumptions.
57	5. Use `analyze-project` only when structure, insertion points, or suspicious
58	implementation patterns need read-only clarification.
59	6. Use `minimal-run-and-audit` for documented inference, evaluation, smoke, or
60	sanity execution.
61	7. Use `run-train` instead when the selected trusted target is training startup,
62	short-run verification, full kickoff, or resume.
63	8. Pause for human review before fuller training claims or any change that could
64	alter dataset, split, checkpoint, preprocessing, metric, loss, model
65	semantics, or result interpretation.
66	9. Write the standardized outputs and give a concise final note in the user's
67	language when practical.
68
69	## Patch Boundary
70
71	Prefer no repository edits. If edits are needed, keep them conservative and
72	auditable:
73
74	- Try command-line arguments, environment variables, path fixes, dependency
75	version fixes, or dependency-file fixes before code changes.
76	- Reproduction fixes are allowed when needed, but they must not be hidden. State
77	what changed, why it was necessary, whether it changes scientific meaning,
78	and whether it affects comparability with the paper, README, or baseline.
79	- Avoid changing model architecture, core inference semantics, training logic,
80	loss functions, or experiment meaning.
81	- If repository files must change, create a branch named
82	`repro/YYYY-MM-DD-short-task`, keep verified patch commits sparse, and record
83	README-fidelity impact in `PATCHES.md`.
84
85	See `references/patch-policy.md`.
86
87	## Outputs
88
89	Always target `repro_outputs/`:
90
91	```text
92	SUMMARY.md
93	COMMANDS.md
94	LOG.md
95	SCIENTIFIC_CHANGELOG.md
96	COMPARABILITY_REPORT.md
97	status.json
98	PATCHES.md # only if patches were applied
99	```
100
101	Use the templates under `assets/` and the field rules in
102	`references/output-spec.md`.
103
104	- Put the shortest high-value summary in `SUMMARY.md`.
105	- Put copyable commands in `COMMANDS.md`.

People who liked this also viewed

6 picks

Type

ai-research-reproduction

161kinstalls10forks

View on GitHub

How to install ai-research-reproduction?

lllllllama/rigorpilot-skills/ai-research-reproduction

$npx -y skills add lllllllama/RigorPilot-Skills --skill ai-research-reproduction

Installs into the current project.

Summary

Files · 1

View on GitHub

SKILL.md

1	# ai-research-reproduction
2
3	## Purpose
4
5	Use this as the Rigor Reproduce compatible skill slug for README-first deep
6	learning repository reproduction. The installed slug remains
7	`ai-research-reproduction` for compatibility. The skill guides the agent toward
8	a minimal trustworthy run with auditable evidence; it should not micromanage
9	implementation details that the model can infer from the repository.
10	Reproduction is not "make it run by changing anything"; it means faithfully
11	reading the README, environment, weights, datasets, and documented commands,
12	then recording results and deviations.
13
14	Start from the shared operating principles in
15	`../../references/agent-operating-principles.md`, then load
16	`../../references/research-rigor-principles.md` and
17	`../../references/deep-learning-experiment-principles.md` when scientific
18	meaning, comparability, or experiment details are at stake.
19
20	## Fit
21
22	Use this skill when all are true:
23
24	- The target is an AI code repository with a README, scripts, configs, or
25	documented commands.
26	- The request spans multiple trusted phases such as intake, setup, execution,
27	training verification, analysis, paper-gap resolution, and reporting.
28	- The desired result is a small reproducible target, not broad experimentation.
29
30	Do not use this skill for paper summaries, generic environment setup, isolated
31	repo scanning, standalone command execution, open-ended research design, or
32	explicit candidate-only exploration.
33
34	## Trusted Target Selection
35
36	Choose the smallest target that can honestly demonstrate repository-grounded
37	reproduction:
38
39	1. documented inference
40	2. documented evaluation
41	3. documented training startup or partial verification
42	4. full training only after explicit user confirmation
43
44	Treat README guidance as the primary reproduction intent. Use repository files
45	to clarify the README, not to silently replace it. When the README and paper
46	conflict, record the conflict and use `paper-context-resolver` only for the
47	narrow reproduction-critical gap.
48
49	## Workflow
50
51	1. Read the README and nearby repo signals.
52	2. Use `repo-intake-and-plan` to extract documented commands and candidate
53	targets.
54	3. Select and justify the minimum trustworthy target.
55	4. Use `env-and-assets-bootstrap` only for target-specific environment,
56	checkpoint, dataset, and cache assumptions.
57	5. Use `analyze-project` only when structure, insertion points, or suspicious
58	implementation patterns need read-only clarification.
59	6. Use `minimal-run-and-audit` for documented inference, evaluation, smoke, or
60	sanity execution.
61	7. Use `run-train` instead when the selected trusted target is training startup,
62	short-run verification, full kickoff, or resume.
63	8. Pause for human review before fuller training claims or any change that could
64	alter dataset, split, checkpoint, preprocessing, metric, loss, model
65	semantics, or result interpretation.
66	9. Write the standardized outputs and give a concise final note in the user's
67	language when practical.
68
69	## Patch Boundary
70
71	Prefer no repository edits. If edits are needed, keep them conservative and
72	auditable:
73
74	- Try command-line arguments, environment variables, path fixes, dependency
75	version fixes, or dependency-file fixes before code changes.
76	- Reproduction fixes are allowed when needed, but they must not be hidden. State
77	what changed, why it was necessary, whether it changes scientific meaning,
78	and whether it affects comparability with the paper, README, or baseline.
79	- Avoid changing model architecture, core inference semantics, training logic,
80	loss functions, or experiment meaning.
81	- If repository files must change, create a branch named
82	`repro/YYYY-MM-DD-short-task`, keep verified patch commits sparse, and record
83	README-fidelity impact in `PATCHES.md`.
84
85	See `references/patch-policy.md`.
86
87	## Outputs
88
89	Always target `repro_outputs/`:
90
91	```text
92	SUMMARY.md
93	COMMANDS.md
94	LOG.md
95	SCIENTIFIC_CHANGELOG.md
96	COMPARABILITY_REPORT.md
97	status.json
98	PATCHES.md # only if patches were applied
99	```
100
101	Use the templates under `assets/` and the field rules in
102	`references/output-spec.md`.
103
104	- Put the shortest high-value summary in `SUMMARY.md`.
105	- Put copyable commands in `COMMANDS.md`.

People who liked this also viewed

6 picks

Type