TheStreamCode/agentic-rd-skill

Agentic R&D Workflow

Use this self-contained skill package to automate a Markdown-first, multi-agent laboratory workflow inside the current workspace.

This is a general-purpose adaptation of the Agent Laboratory pattern: independent evidence review, collaborative plan formulation, task execution or investigation, results analysis, quality review, and final report writing. It is not limited to scientific research; apply the same lab workflow to product, business, technical, strategy, legal, compliance, security, UX, market, and operational use cases.

Designed for skill-compatible coding agents with filesystem access. Prefer real parallel subagents whenever the environment supports them; use isolated simulated passes only when real subagents are unavailable.

Package Contents

• SKILL.md: activation rules and core operating instructions.
• references/generalized-lab-model.md: universal Agent Laboratory model and use-case mapping.
• references/workflow.md: full phase sequence and approval gate.
• references/agent-roles.md: specialist role selection guidance.
• references/quality-rules.md: evidence, uncertainty, review, and safety standards.
• references/implementation-notes.md: install, scaffold, subagent, and file-handling notes.
• references/example-run.md: a compact end-to-end example showing the expected file trail and output quality.
• assets/templates/: Markdown templates for all workflow files.
• scripts/init-rd-workflow.mjs: optional scaffold script.
• scripts/validate-skill.mjs: repository validation for the public skill package.

First Action

1. Look for project-brief.md in the current workspace.
2. If it exists, read it completely before any other workflow step.
3. If it does not exist and the user gave enough context, create it with assets/templates/project-brief.md.
4. If essential context is missing, ask only the minimum blocking questions.
5. After the brief exists, continue without asking for permission between normal workflow phases.

Run scripts/init-rd-workflow.mjs to scaffold project-brief.md and the required work/ files. The script must not create work/06-final-output.md.

Required Outputs

Create and use this structure in the active workspace:

work/
├── 00-lab-notes.md
├── 01-orchestration-plan.md
├── 02-specialist-outputs/
├── 03-team-collaboration.md
├── 04-cross-review-notes.md
├── 05-stage-gate-review.md
└── 06-final-output.md

Each workflow phase maps to a specific output path. The numbered laboratory phases (evidence review, plan formulation, execution, results analysis) are not separate files: they are materialized as one file per specialist inside work/02-specialist-outputs/.

Phase	Output path
Lab setup	work/00-lab-notes.md
Orchestration	work/01-orchestration-plan.md
Evidence review, plan, execution, results (one file per specialist)	work/02-specialist-outputs/NN-<role>.md
Team collaboration (consolidated by the orchestrator)	work/03-team-collaboration.md
Cross-review	work/04-cross-review-notes.md
Stage-gate review	work/05-stage-gate-review.md
Final synthesis	work/06-final-output.md

Do not create or write work/06-final-output.md until work/05-stage-gate-review.md says Approved.

Workflow

1. Read references/workflow.md for the canonical phase sequence.
2. Read references/generalized-lab-model.md to map the user's use case onto the laboratory phases.
3. Read references/agent-roles.md before selecting specialists.
4. Read references/quality-rules.md before cross-review, stage-gate review, and final synthesis.
5. Use references/implementation-notes.md for script use, subagent behavior, and file-handling details.
6. Use templates from assets/templates/ for generated workflow files.
7. Consult references/example-run.md for a worked end-to-end example when unsure about expected output quality.

Scale To The Brief

Match workflow weight to the brief. Do not run a full multi-wave laboratory for a small question.

• Trivial brief: the orchestrator may do a single inline pass with no subagents. Record evidence, assumptions, and risks in work/00-lab-notes.md, write a short work/05-stage-gate-review.md self-check, then work/06-final-output.md. The intermediate specialist, collaboration, and cross-review files are optional at this tier. The stage-gate rule still holds: do not write 06-final-output.md until 05-stage-gate-review.md says Approved.
• Small or narrow brief: use one or two specialists, run a single wave, and fold collaboration into the orchestrator's notes instead of a separate exchange. Keep cross-review, stage-gate, and final synthesis as a single quick pass each.
• Medium brief: use a few specialists across one or two waves with a light team-collaboration step.
• Large or regulated brief: use the full phase sequence, multiple waves, explicit team collaboration, and the relevant review roles.

Always keep the mandatory minimum regardless of size: evidence/context review, plan formulation, execution or investigation, results analysis, risk/assumptions coverage, and the stage-gate quality gate before final synthesis. Everything else is optional and added only when it materially changes the deliverable. See references/generalized-lab-model.md for the scaling detail.

Execution Rules

• Select the fewest specialist roles that cover the brief well.
• Always include evidence/context review, plan formulation, execution or investigation, results analysis, and risk/assumptions coverage.
• Prefer real parallel subagents. They are the primary execution model for this skill, not an optional enhancement.
• In each phase, spawn all independent specialists before waiting for results. Do not run independent specialists one at a time when parallel subagents are available.
• Otherwise, simulate specialists as isolated passes. Each pass must read only the brief, orchestration plan, and its assigned scope before writing its own file.
• Never collapse specialist analysis, cross-review, stage-gate review, and final synthesis into one response.
• Parallelize independent specialist work, but keep dependency gates sequential: review before plan, plan before execution, execution before results analysis, review before final report.
• Require citations or source notes for factual claims when sources are available.
• Mark unsupported claims as assumptions or inferences.
• Preserve uncertainty and regulated-domain review boundaries.

Collaborative Team Protocol

• Treat subagents as a collaborating team, not disconnected report writers.
• The Master Orchestrator defines team roles, ownership, dependencies, and collaboration rules in work/01-orchestration-plan.md.
• Each subagent writes only its own specialist output file, and returns its questions, dependencies, disagreements, and handoff notes to the orchestrator (in its result, not by editing a shared file).
• The orchestrator consolidates those returned notes into work/03-team-collaboration.md. Never have multiple subagents write the same file concurrently.
• The team must reconcile shared assumptions before cross-review.
• Cross-review evaluates both specialist outputs and the team collaboration record.
• Final synthesis must reflect resolved agreements, unresolved disagreements, and important minority views.

Stop Conditions

Stop and ask the user only when:

• The subject or goal is not identifiable.
• The desired output cannot be inferred.
• Key constraints conflict.
• Required external data, credentials, private systems, paid tools, or unavailable network access block the work.
• Legal, medical, financial, compliance, security, or safety-critical advice cannot be kept within a safe informational boundary.

Completion

When the workflow completes, tell the user which files were created and whether the stage gate approved final synthesis. Keep the completion message concise.