Convert analysis findings into steelmanned, actionable plans with conditional recommendations. Every recommendation includes a derivation chain explaining WHY. Requires a source analysis from /ar, /aw, /u, /araw, or /uaua.
Convert analysis findings into steelmanned, actionable plans with conditional recommendations. Every recommendation includes a derivation chain explaining WHY. Requires a source analysis from /ar, /aw, /u, /araw, or /uaua.
output
{"format":"prose"}
PROPOSE - Evidence-Derived Plan Synthesis
Input: $ARGUMENTS
Interpretations
Before executing, identify which interpretation matches the user's input:
Interpretation 1 — Convert prior analysis to plans: The user has already run /ar, /aw, /u, /araw, or /uaua and wants to synthesize those findings into actionable, evidence-derived plans with conditional recommendations.
Interpretation 2 — Generate plans without prior analysis: The user wants actionable plans but has not run a source analysis first — redirect them to /pss or suggest running a source analysis first.
Interpretation 3 — Evaluate existing plans against evidence: The user already has candidate plans and wants to steelman, evaluate, and rank them using evidence from a prior analysis.
If ambiguous, ask: "I can help with converting analysis findings into plans, generating plans from scratch (via /pss), or evaluating existing plans against evidence — which fits?"
If clear from context, proceed with the matching interpretation.
Core Principles
Plans derive from evidence only. Every plan must trace back to specific numbered items (F-numbers, R-numbers, W-numbers, U-numbers) from a source analysis. If you cannot cite a source item, you don't have a plan -- you have a guess. This skill REFUSES to operate without a source analysis.
Steelman before evaluating. Every plan is presented in its strongest possible form BEFORE being evaluated. A plan that was never steelmanned was never fairly considered.
Conditional recommendations as default. "Use this if X / don't use this if Y" is the standard format. Binary RECOMMEND / NOT RECOMMEND is a failure mode -- reality has conditions.
Every recommendation has a derivation chain. Source items -> what they show -> why that leads to this verdict. No recommendation without the chain.
NOT RECOMMENDED plans get full treatment. Rejected plans include: steelman, why the steelman still fails, and what would change the verdict. Anything less is strawman rejection.
What you give up is part of the plan. Every plan forecloses alternatives. If you can't state what a plan makes impossible, you haven't analyzed it.
Three phases, strict separation. Phase 1 extracts (no evaluation). Phase 2 compiles (no new plans). Phase 3 recommends (only from the registry). Never mix phases.
Prerequisite: Source Analysis
This skill REQUIRES a prior analysis from one of:
/ar (Assume Right) -- source items are R-numbered
/aw (Assume Wrong) -- source items are W-numbered
/u (Universalize) -- source items are U-numbered
/araw (Assume Right / Assume Wrong) -- source items are C-numbered (claims) and F-numbered (findings)
/uaua (Universalize-ARAW-Universalize-ARAW) -- source items are U-numbered and F-numbered
If no source analysis exists in the conversation or is provided as input:
PROPOSE REFUSED
===============
No source analysis found.
/p converts analysis findings into plans. It does not generate plans from scratch.
To use /p:
1. Run /ar, /aw, /u, /araw, or /uaua on your topic first
2. Then run /p to convert the findings into actionable plans
For generating plans without prior analysis, use /pss instead.
Stop here. Do not proceed.
Phase 1: PLAN EXTRACTION
Step 1: Identify the Source
SOURCE ANALYSIS: [which skill produced it -- ar/aw/u/araw/uaua]
SOURCE TOPIC: [what was analyzed]
SOURCE DEPTH: [depth level if specified]
ITEM PREFIX: [R/W/U/C+F -- the numbering used in the source]
TOTAL SOURCE ITEMS: [count]
Step 2: Extract Plan Seeds
Scan the source registry for "plan seeds" -- items that point toward action. Every seed must cite a source item number.
Seed types to extract:
Seed Type
Where to Find
Example
DO_FIRST actions
Synthesis section of ARAW/UAUA
"Test assumption X" from F23
Derived alternatives
AW findings, multi-valued AW
Alternative Y from W8 or F15
Foreclosures
AR findings, foreclosure nodes
"Can no longer do Z" from R6 or F6
CRUX points
CRUX registry in ARAW
"If X is true, then plan A; if false, plan B" from CRUX-3
Tensions
Tension registry
"X contradicts Y" -- resolve via plan
Convergent implications
AR/ARAW where multiple branches require same thing
Group related seeds into 3-7 concrete plans. Each plan must:
Have a clear action statement (what you DO)
Cite at least 2 source items
Be distinct from other plans (not a variant of the same thing)
[PLAN-1] [clear action statement]
Seeds: [SEED-numbers]
Source items: [item numbers from original analysis]
[PLAN-2] [clear action statement]
Seeds: [SEED-numbers]
Source items: [item numbers from original analysis]
...
Cluster check: If you have fewer than 3 plans, the source analysis may be too narrow -- note this. If you have more than 7, look for plans that can merge.
Step 4: Steelman Each Plan
For EVERY plan (including ones you suspect will be rejected), present the strongest possible case:
[PLAN-1] STEELMAN:
STRONGEST CASE: [2-4 sentences -- the best argument for this plan]
BEST CONDITIONS: [when/where this plan would work best]
IDEAL OUTCOME: [what success looks like if everything goes right]
SOURCE SUPPORT: [which source items support this plan, with brief explanation]
Steelman quality check: If your steelman is 1 sentence or dismissive, you failed. A steelman should make someone genuinely consider the plan.
Step 5: Evaluate Against Source Evidence
For EVERY plan, evaluate honestly using source evidence:
[PLAN-1] EVALUATION:
SUPPORTING EVIDENCE: [source items that support -- item numbers + what they show]
OPPOSING EVIDENCE: [source items that oppose -- item numbers + what they show]
EVIDENCE BALANCE: [which side has stronger/more bedrock evidence]
CRITICAL DEPENDENCY: [what must be true for this plan to work -- cite items]
FAILURE MODE: [how this plan most likely fails -- cite items]
Phase 2: PLAN REGISTRY
After ALL extraction and evaluation is complete, compile into a structured registry. No new plans introduced here.
PLAN REGISTRY
=============
SOURCE: [skill] on [topic] at [depth]
PLANS EXTRACTED: [N]
[PLAN-1] [action statement]
STEELMAN: [strongest case -- 1-2 sentences]
STRENGTHS: [from source items -- cite numbers]
WEAKNESSES: [from source items -- cite numbers]
RISKS: [what could go wrong -- cite items]
DEPENDENCIES: [what must be true -- cite items]
REVERSIBILITY: [easy to undo / hard to undo / one-way door]
COST: [what this requires -- time, money, effort, opportunity]
GIVES UP: [what this plan forecloses -- what you can no longer do]
FIRST ACTIONS: [concrete next steps if this plan is chosen]
[PLAN-2] ...
...
COMPARISON TABLE:
| Plan | Strengths | Weaknesses | Reversible? | Cost | Gives Up |
|------|-----------|------------|-------------|------|----------|
| PLAN-1 | [brief] | [brief] | [Y/N/Partial] | [brief] | [brief] |
| PLAN-2 | [brief] | [brief] | [Y/N/Partial] | [brief] | [brief] |
...
PLAN INTERACTIONS:
- [PLAN-X] and [PLAN-Y]: [compatible / incompatible / sequential]
- [PLAN-X] enables [PLAN-Z]: [how]
- [PLAN-X] blocks [PLAN-Z]: [how]
Phase 3: RECOMMENDATION SYNTHESIS
Derived entirely from the plan registry. No new plans or evidence introduced here.
Categorize Each Plan
Every plan gets ONE of four categories:
RECOMMENDED -- Evidence strongly supports. Use this.
[PLAN-N] RECOMMENDED
DERIVATION CHAIN:
Source items [numbers] show [what they show]
-> This means [intermediate conclusion]
-> Therefore [why this plan follows]
USE IF: [conditions where this is the right choice]
DON'T USE IF: [conditions where this is wrong]
FIRST ACTIONS:
1. [specific action] -- resolves [what]
2. [specific action] -- resolves [what]
CONDITIONALLY RECOMMENDED -- Evidence supports under specific conditions.
[PLAN-N] CONDITIONALLY RECOMMENDED
DERIVATION CHAIN:
Source items [numbers] show [what they show]
-> This means [intermediate conclusion]
-> Therefore [why this plan follows WHEN conditions hold]
CONDITION: [precise condition that must be true]
HOW TO CHECK: [how to verify the condition]
USE IF: [when the condition holds + what else must be true]
DON'T USE IF: [when the condition fails + what to do instead]
IF CONDITION UNKNOWN: [what to do to find out]
FIRST ACTIONS:
1. [specific action] -- resolves [what]
NOT RECOMMENDED -- Evidence opposes. Don't use this.
[PLAN-N] NOT RECOMMENDED
STEELMAN REMINDER: [restate the strongest case -- 1-2 sentences]
DERIVATION CHAIN:
Source items [numbers] show [what they show]
-> This means [intermediate conclusion]
-> Despite the steelman, this fails because [why]
WHY IT FAILS: [specific source items that defeat the steelman]
WHAT WOULD CHANGE THIS VERDICT: [what new evidence would flip to RECOMMENDED]
USE INSTEAD: [which other plan addresses the same need]
UNRESOLVED -- Evidence insufficient to recommend or reject.
[PLAN-N] UNRESOLVED
DERIVATION CHAIN:
Source items [numbers] show [what they show]
-> But [what's missing or contradictory]
-> Cannot determine recommendation because [why]
WHAT'S MISSING: [specific information needed]
HOW TO RESOLVE: [specific actions to get that information]
LEANS TOWARD: [which direction evidence slightly favors, if any]
Final Summary
RECOMMENDATION SUMMARY
======================
SOURCE: [skill] on [topic]
RECOMMENDED:
- [PLAN-N]: [one-line summary] -- USE IF: [brief condition]
CONDITIONALLY RECOMMENDED:
- [PLAN-N]: [one-line summary] -- CONDITION: [brief]
NOT RECOMMENDED:
- [PLAN-N]: [one-line summary] -- BECAUSE: [brief]
WOULD CHANGE IF: [brief]
UNRESOLVED:
- [PLAN-N]: [one-line summary] -- NEEDS: [brief]
SUGGESTED SEQUENCE:
[If multiple plans are compatible, suggest an order]
1. [First action from highest-priority RECOMMENDED plan]
2. [Check condition for CONDITIONALLY RECOMMENDED plan]
3. [Resolve UNRESOLVED plans by gathering missing info]
WHAT THIS ANALYSIS DOES NOT COVER:
[Explicitly state limitations -- what /p can't tell you]
Anti-Failure Checks
Failure Mode
Signal
Fix
Plan from thin air
Plan doesn't cite source item numbers
Delete it. Every plan traces to source items.
Soft steelman
1-sentence dismissive steelman
Rewrite. Steelman should make someone genuinely consider the plan.
Recommendation without derivation
"RECOMMENDED" with no chain
Add the chain: source items -> what they show -> why this verdict.
Binary-only recommendations
All RECOMMENDED or NOT RECOMMENDED, no conditions
Add USE IF / DON'T USE IF to every recommendation. Reality has conditions.
Missing "gives up"
Plan entry has no GIVES UP field
Every plan forecloses something. Find what.
Strawman rejection
NOT RECOMMENDED without steelman reminder
Always restate the steelman before explaining why it fails.
Cherry-picked evidence
Derivation chain cites 2 items but source has 30
Check ALL relevant source items, not just the ones that support your verdict.
Missing failure mode
Plan has no failure mode identified
Every plan can fail. How does this one fail? Cite source items.
No interactions
Plans listed independently with no interaction analysis
Check: are plans compatible? Sequential? Does one block another?
Pre-Completion Check
Source analysis identified and referenced (skill, topic, item prefix)
All plan seeds cite source item numbers
3-7 plans extracted (if fewer, noted why)
EVERY plan steelmanned (including rejected ones)
Steelmans are genuine (not dismissive 1-liners)
EVERY plan evaluated against source evidence (both supporting and opposing)
EVERY plan has: strengths, weaknesses, risks, dependencies, reversibility, cost, gives up
Comparison table complete
Plan interactions analyzed
EVERY recommendation has a derivation chain (source items -> intermediate conclusion -> verdict)
EVERY recommendation has USE IF / DON'T USE IF conditions
NOT RECOMMENDED plans include steelman reminder + what would change verdict
UNRESOLVED plans include what's missing + how to resolve
No plans introduced without source item citations
Cherry-pick check: Are there source items you ignored? If yes, check if they affect any plan.
Strawman check: Re-read each NOT RECOMMENDED. Would a proponent feel their plan was fairly considered?
Condition check: Are there RECOMMENDED plans that should be CONDITIONALLY RECOMMENDED? Binary confidence is suspicious.
