group-sequential-design

Name: group-sequential-design
Author: RConsortium

88stars20forks

How to install group-sequential-design?

rconsortium/pharma-skills/group-sequential-design

$npx -y skills add RConsortium/pharma-skills --skill group-sequential-design

Installs into the current project.

Summary

Design group sequential clinical trials for survival endpoints (OS, PFS, DFS) with interim analyses, spending functions, boundaries, multiplicity, and event/enrollment prediction. Triggers on: Phase 3 trial design, sample size/events for survival endpoints, alpha spending, group

Files · 1

View on GitHub

SKILL.md

1	## Bundled Resources
2
3	\| File \| What it contains \| When to read \|
4	\|------\|-----------------\|--------------\|
5	\| `reference.md` \| Design guidance, parameter tables, spending functions, key rules, failure modes, analysis framework, protocol references \| After collecting user inputs, before writing R code \|
6	\| `examples.md` \| All R code examples organized by design pattern \| When you need code for a specific pattern — read only the relevant section \|
7	\| `post_design.md` \| IA timing checks (warning messages, user options), verification procedure (pass criteria, lrsim code, log template) \| After computing the design (step 6), before delivering \|
8	\| `scripts/gsd_report_template.py` \| Bundled Python report template — reads gsd_results.json, generates gsd_report.docx with zero hardcoded values \| Step 9: copy to output dir and run (or extend for design-specific features) \|
9
10	Do NOT read these files upfront. Read them only when you reach the corresponding step.
11
12	---
13
14	## Compact Instructions
15
16	When the conversation gets long and context is compressed, preserve information in this priority order:
17
18	1. Architecture decisions — never re-summarize these; they are settled
19	2. Key changes — what was computed, what parameters were chosen
20	3. Current verification status — pass/fail only, drop verbose logs
21	4. Tool outputs — can be deleted; keep only essential numbers (events, boundaries, power, sample size)
22
23	---
24
25	## Workflow
26
27	Task Progress Tracking: At the start of step 6 (after user confirms inputs), create a task checklist using `TaskCreate` for all remaining steps. Mark each task `in_progress` when starting and `completed` when done. Typical tasks:
28	1. Write and run R design script (gsd_design.R)
29	2. Check IA timing constraints
30	3. Run verification simulation (gsd_verification.R)
31	4. Generate Word report (gsd_report.py)
32
33	This gives the user a visual progress indicator (spinner → checkmark) throughout the computation.
34
35	Pacing — one artifact per turn. Do not bundle multiple major artifacts into a single response. Each of these is its own turn:
36	- Writing `gsd_design.R` (then stop — let the tool result come back)
37	- Running `gsd_design.R` (then stop)
38	- Writing `gsd_verification.R`
39	- Running the verification
40	- Writing the report
41
42	When you call `Write` for a script, do not also include long reasoning or a parallel `Bash` execution in the same turn. Save explanations for the final report. This keeps each assistant turn well under the output token cap and avoids truncation.
43
44	---
45
46	0. Scope check — do this before anything else.
47
48	This skill supports standard group sequential designs (GSD) for survival endpoints (OS, PFS, DFS). Before collecting any inputs, assess whether the user's request falls within this scope.
49
50	In scope — proceed to Step 1:
51	- Fixed set of hypotheses tested at pre-planned interim and final analyses
52	- Survival endpoints (time-to-event): OS, PFS, DFS, EFS
53	- Single or co-primary endpoints; single or multiple nested populations (subgroup + ITT)
54	- Pre-specified alpha spending, futility boundaries, multiplicity graphs
55	- Non-proportional hazard evaluation as a sensitivity analysis on a fixed GSD
56
57	Out of scope — stop and explain. If the user's request matches any of the patterns below, do NOT proceed with the GSD workflow. Instead, reply with the message template at the end of this step.
58
59	\| Pattern \| Key signals in the user's description \|
60	\|---------\|--------------------------------------\|
61	\| Adaptive enrichment \| "enrich to biomarker-positive patients based on interim data", "narrow the population if the subgroup is responding", "population selection at interim", "pick the responders mid-trial" \|
62	\| Sample size re-estimation (SSR) \| "increase the sample size based on interim results", "blinded or unblinded SSR", "adaptive sample size", "re-estimate events or N at interim" \|
63	\| Dose selection / seamless Phase 2/3 \| "select the best dose at interim and continue", "drop losing arms", "pick the winner dose", "seamless adaptive Phase 2/3" \|
64	\| Platform / master protocol \| "add or remove treatment arms dynamically", "master protocol", "basket or umbrella trial arm management" \|
65	\| Non-survival primary endpoint \| Binary response rate, continuous endpoint, count data as the primary — note: survival co-primaries alongside a binary endpoint are borderline; ask the user to clarify which endpoint drives the GSD \|
66	\| Adaptive randomization \| "response-adaptive randomization", "update allocation ratios based on

People who liked this also viewed

6 picks

Type

group-sequential-design

88stars20forks

View on GitHub

How to install group-sequential-design?

rconsortium/pharma-skills/group-sequential-design

$npx -y skills add RConsortium/pharma-skills --skill group-sequential-design

Installs into the current project.

Summary

Files · 1

View on GitHub

SKILL.md

1	## Bundled Resources
2
3	\| File \| What it contains \| When to read \|
4	\|------\|-----------------\|--------------\|
5	\| `reference.md` \| Design guidance, parameter tables, spending functions, key rules, failure modes, analysis framework, protocol references \| After collecting user inputs, before writing R code \|
6	\| `examples.md` \| All R code examples organized by design pattern \| When you need code for a specific pattern — read only the relevant section \|
7	\| `post_design.md` \| IA timing checks (warning messages, user options), verification procedure (pass criteria, lrsim code, log template) \| After computing the design (step 6), before delivering \|
8	\| `scripts/gsd_report_template.py` \| Bundled Python report template — reads gsd_results.json, generates gsd_report.docx with zero hardcoded values \| Step 9: copy to output dir and run (or extend for design-specific features) \|
9
10	Do NOT read these files upfront. Read them only when you reach the corresponding step.
11
12	---
13
14	## Compact Instructions
15
16	When the conversation gets long and context is compressed, preserve information in this priority order:
17
18	1. Architecture decisions — never re-summarize these; they are settled
19	2. Key changes — what was computed, what parameters were chosen
20	3. Current verification status — pass/fail only, drop verbose logs
21	4. Tool outputs — can be deleted; keep only essential numbers (events, boundaries, power, sample size)
22
23	---
24
25	## Workflow
26
27	Task Progress Tracking: At the start of step 6 (after user confirms inputs), create a task checklist using `TaskCreate` for all remaining steps. Mark each task `in_progress` when starting and `completed` when done. Typical tasks:
28	1. Write and run R design script (gsd_design.R)
29	2. Check IA timing constraints
30	3. Run verification simulation (gsd_verification.R)
31	4. Generate Word report (gsd_report.py)
32
33	This gives the user a visual progress indicator (spinner → checkmark) throughout the computation.
34
35	Pacing — one artifact per turn. Do not bundle multiple major artifacts into a single response. Each of these is its own turn:
36	- Writing `gsd_design.R` (then stop — let the tool result come back)
37	- Running `gsd_design.R` (then stop)
38	- Writing `gsd_verification.R`
39	- Running the verification
40	- Writing the report
41
42	When you call `Write` for a script, do not also include long reasoning or a parallel `Bash` execution in the same turn. Save explanations for the final report. This keeps each assistant turn well under the output token cap and avoids truncation.
43
44	---
45
46	0. Scope check — do this before anything else.
47
48	This skill supports standard group sequential designs (GSD) for survival endpoints (OS, PFS, DFS). Before collecting any inputs, assess whether the user's request falls within this scope.
49
50	In scope — proceed to Step 1:
51	- Fixed set of hypotheses tested at pre-planned interim and final analyses
52	- Survival endpoints (time-to-event): OS, PFS, DFS, EFS
53	- Single or co-primary endpoints; single or multiple nested populations (subgroup + ITT)
54	- Pre-specified alpha spending, futility boundaries, multiplicity graphs
55	- Non-proportional hazard evaluation as a sensitivity analysis on a fixed GSD
56
57	Out of scope — stop and explain. If the user's request matches any of the patterns below, do NOT proceed with the GSD workflow. Instead, reply with the message template at the end of this step.
58
59	\| Pattern \| Key signals in the user's description \|
60	\|---------\|--------------------------------------\|
61	\| Adaptive enrichment \| "enrich to biomarker-positive patients based on interim data", "narrow the population if the subgroup is responding", "population selection at interim", "pick the responders mid-trial" \|
62	\| Sample size re-estimation (SSR) \| "increase the sample size based on interim results", "blinded or unblinded SSR", "adaptive sample size", "re-estimate events or N at interim" \|
63	\| Dose selection / seamless Phase 2/3 \| "select the best dose at interim and continue", "drop losing arms", "pick the winner dose", "seamless adaptive Phase 2/3" \|
64	\| Platform / master protocol \| "add or remove treatment arms dynamically", "master protocol", "basket or umbrella trial arm management" \|
65	\| Non-survival primary endpoint \| Binary response rate, continuous endpoint, count data as the primary — note: survival co-primaries alongside a binary endpoint are borderline; ask the user to clarify which endpoint drives the GSD \|
66	\| Adaptive randomization \| "response-adaptive randomization", "update allocation ratios based on

People who liked this also viewed

6 picks

Type