Proposal Synthesis

Proposal Synthesis

Turn an open problem into a small design space of complete proposal briefs. This skill sits between loose ideation and critique: it should create critique-ready options, not just brainstorm fragments and not yet run the bounded critique loop itself.

Read these references only when needed:

• output-contract.md for the proposal_packet shape and handoff fields
• domain-lens-packs.md for default divergence criteria across research, development, and general problem-solving
• subagent-patterns.md for bounded multi-subagent role sets and isolation rules
• prompt-shaping.md for how to tighten the input with thinking-lenses

Use This Skill For

• vague goals that need 2-5 complete proposal options rather than one recommended direction
• requests that say give me several plans, compare multiple approaches, or generate proposals before critique
• cross-domain work where the user wants research, engineering, or decision options in a common proposal shape
• cases where proposal-critique-refine would be premature because no stable proposal_brief exists yet
• bounded multi-subagent divergence where each subagent should generate a different candidate instead of critiquing one shared draft

Do Not Use

• the user only wants loose exploration, sketches, or clarifying conversation
  • use brainstorming
• a coherent proposal already exists and mainly needs critique, repair, or stress testing
  • use proposal-critique-refine
• the task is mostly product discovery, stakeholder research, or market analysis
  • use business-analyst
• the user wants implementation planning or coding right now
  • use delivery-conductor or a narrower execution lane
• the main missing ingredient is external fact-finding rather than proposal generation
  • do the research first instead of inventing around missing evidence

Input Contract

Require only:

• the problem, opportunity, or desired outcome

Infer when possible:

• target user or stakeholder
• constraints
• decision criteria
• time horizon
• what makes the options meaningfully different

Ask follow-up questions only when a missing fact would make the candidate proposals non-comparable, unsafe, or obviously mis-scoped.

Output Contract

Always produce a proposal_packet containing:

• problem_frame
• decision_criteria
• candidate_proposals
• recommended_candidate
• open_questions
• handoff_hint

Each candidate proposal must be complete enough for a downstream critic to attack it without inventing the missing core.

Workflow

1. Frame the Problem

Normalize the ask into:

• objective
• target user or stakeholder
• constraints
• success criteria
• biggest unknown

Keep the frame compact and faithful. Do not quietly upgrade a small request into a bigger strategy exercise.

2. Tighten the Prompt When Needed

If the request is underspecified, high-cost, or likely to branch too widely, first tighten the generation prompt with thinking-lenses. Prefer the lightest shaping that still protects quality:

• inline-structured for most medium tasks
• pre-optimizer only when the solution space is wide, ambiguous, or expensive

Do not spend more effort polishing the prompt than producing the proposals.

3. Choose the Divergence Shape

Default to 3 candidate proposals.

Use:

• 2 when the scope is small or urgency is high
• 4-5 when the task is broad and the additional options remain meaningfully different

Pick one domain lens pack from domain-lens-packs.md:

• research
• development
• general problem-solving

The candidate set should differ in mechanism, risk posture, or sequencing, not just wording.

4. Generate Candidate Proposal Briefs

Each candidate should include:

• title
• core_thesis
• approach
• why_this_could_work
• main_tradeoff
• key_assumptions
• main_risks
• first_validation_step

Favor complete, attackable proposals over a long list of thin ideas.

5. Use Subagents Only When They Improve Coverage

If subagents are available and the design space is genuinely wide, use up to 4 bounded proposal generators. Give every subagent:

• the same problem_frame
• the same hard constraints
• the same decision criteria
• one distinct role or lens

Do not show one subagent another subagent's output before synthesis. If subagents are not available, run serial passes and explicitly label the result as reduced-independence perspective-taking.

6. Synthesize the Design Space

Merge duplicates. Keep only the 2-5 options that remain materially different and decision-relevant.

Then produce:

• a comparison grid
• the best-fit recommendation
• one sentence on why each rejected option was not selected

If a hybrid is recommended, explain what is borrowed from each source option and why the hybrid is cleaner than either parent.

7. Prepare the Critique Handoff

Select one proposal_brief as the critique candidate. Package it with:

• alternative options that were considered
• decision criteria
• unresolved questions
• recommended critique lenses

Use the schema in output-contract.md.

8. Stop at the Proposal Layer

This skill stops once the proposal set is ready.

Next handoffs:

• use proposal-critique-refine when the best candidate now needs bounded critique and repair
• use delivery-conductor when the direction is good enough to push into planning, implementation, testing, or verified delivery

Do not silently continue into implementation planning unless the user explicitly changes the task.

Multi-Agent Mode

Use multi-agent divergence only when it materially improves coverage.

Default role sets live in subagent-patterns.md. Each subagent should return one candidate proposal with:

• thesis
• main_tradeoff
• largest_assumption
• first_validation_step

Good uses:

• cross-disciplinary research directions
• product or engineering strategy with real architecture tradeoffs
• ambiguous problem-solving where reversibility and risk posture differ

Poor uses:

• tiny scoped tasks
• cases where all roles would generate near-identical proposals
• requests where the real blocker is missing facts, not missing options

Guardrails

• do not confuse this skill with free-form brainstorming
• do not run a critique loop here and pretend it replaces proposal-critique-refine
• do not let more candidates crowd out meaningful differences
• do not invent evidence to make weak proposals look grounded
• do not produce one polished option plus cosmetic variants
• do not present same-model serial passes as independent validation
• do not jump into implementation plans, task lists, or code changes unless the user explicitly asks to continue

Example Triggers

• Use $proposal-synthesis to turn this vague direction into 3 critique-ready proposal briefs.
• Use $proposal-synthesis to generate multiple research plans, recommend one, and package it for critique.
• Use $proposal-synthesis to explore several engineering directions before we send one to proposal-critique-refine.
• Use $proposal-synthesis with bounded subagent divergence so I can compare multiple complete solutions instead of one draft.