@esoteric-logic/praxis-harness 2.16.0 → 3.0.0

package/base/skills/px-prompt/SKILL.md

@@ -11,35 +11,91 @@ Single entry point for the prompt engine. Detects what needs to happen based on
 
 ## Invocation
 - `/px-prompt <project-name>` — create, generate, or regenerate a project's prompts
+- `/px-prompt --deal <deal-name> [--type recompete|new-start|task-order|idiq]` — Maximus capture deal (0 questions)
+- `/px-prompt --advance <deal-name>` — advance deal to next capture phase, update scoring expectations
+- `/px-prompt --edit <project-name> "<change>"` — targeted edit, auto-regenerate platform outputs
 - `/px-prompt --sync` — recompile all projects, report diffs and budgets
+- `/px-prompt --dashboard` — project index with status, budgets, and staleness
+- `/px-prompt --refresh <project-name>` — re-run Perplexity research, diff and update
+- `/px-prompt --deploy <project-name>` — copy outputs to clipboard with deployment URLs
+- `/px-prompt --scan <project-name>` — full audit of project quality and budgets
 - `/px-prompt --list` — list all projects and their status
 
 ---
 
+## Initial Menu Behavior
+
+**NEVER hardcode individual project names in the initial options menu.**
+When the user runs `/px-prompt` without arguments, show ONLY these generic options:
+1. Create new project
+2. Sync all projects
+3. Work on existing project → THEN dynamically list projects from `node bin/prompt-compile.js --list`
+
+Do NOT show "Work on praxis" or "Work on maximus" or any project name — the engine is project-agnostic. Discover projects dynamically at runtime.
+
+---
+
+## Output File Naming Convention
+
+Platform outputs use suffixed filenames so users can distinguish them at a glance:
+
+| Platform | Output Filename | Budget |
+|----------|----------------|--------|
+| Source (Claude Projects) | `system-prompt.md` | 5,000 chars |
+| Claude Desktop / Projects | `project-instructions-claude-desktop.md` | 2,500 chars |
+| Perplexity Spaces | `space-instructions-perplexity.md` | 4,000 chars |
+| Claude Code | `CLAUDE.md` | 250 lines |
+
+All output files live in `prompts/projects/<project-name>/`.
+Reference/knowledge files live in `prompts/projects/<project-name>/references/`.
+
+---
+
 ## Routing Logic
 
 When invoked with a project name, detect the right action:
 
 ```
-/px-prompt <project-name>
+/px-prompt <args>
+  │
+  ├─ --deal <deal-name> [--type ...]?
+  │   → ACTION: DEAL (Step 7) — 0 questions, Perplexity researches, auto-scaffold
+  │
+  ├─ --advance <deal-name>?
+  │   → ACTION: ADVANCE PHASE (Step 12) — move deal to next capture phase
+  │
+  ├─ --edit <project-name> "<change>"?
+  │   → ACTION: TARGETED EDIT (Step 8) — edit one section, regenerate outputs
+  │
+  ├─ --dashboard?
+  │   → ACTION: DASHBOARD (Step 9) — project index with staleness
+  │
+  ├─ --refresh <project-name>?
+  │   → ACTION: REFRESH (Step 10) — re-run Perplexity research, diff, update
   │
-  ├─ Project doesn't exist?
+  ├─ --deploy <project-name>?
+  │   → ACTION: DEPLOY (Step 11) — clipboard + URLs
+  │
+  ├─ --scan <project-name>?
+  │   → ACTION: SCAN & AUDIT (Step 6)
+  │
+  ├─ --sync?
+  │   → ACTION: SYNC ALL (Step 5)
+  │
+  ├─ <project-name> — project doesn't exist?
   │   → ACTION: CREATE (Step 1)
   │
-  ├─ Project exists, mode: standalone, system-prompt.md missing?
-  │   → ACTION: GENERATE FROM SCRATCH (Step 2)
+  ├─ <project-name> — exists, standalone, system-prompt.md missing?
+  │   → ACTION: GENERATE (Step 2)
   │
-  ├─ Project exists, mode: standalone, system-prompt.md exists, platform outputs missing?
+  ├─ <project-name> — exists, standalone, platform outputs missing?
   │   → ACTION: CONDENSE (Step 3)
   │
-  ├─ Project exists, mode: standalone, all files present?
-  │   → ACTION: VALIDATE + offer to regenerate (Step 4)
-  │
-  ├─ Project exists, mode: compiled?
-  │   → ACTION: COMPILE (Step 4)
+  ├─ <project-name> — exists, all files present?
+  │   → ACTION: VALIDATE (Step 4)
   │
-  └─ --sync flag?
-      → ACTION: SYNC ALL (Step 5)
+  └─ <project-name> — exists, compiled?
+      → ACTION: COMPILE (Step 4)
 ```
 
 ---
@@ -48,73 +104,86 @@ When invoked with a project name, detect the right action:
 
 **Triggered when:** project folder doesn't exist.
 
-### 1a. Core intake (always ask)
+### 1a. Single-question intake
 
-1. **Description** — "Describe this project in one sentence."
+Ask ONE question: **"Describe this project in 1-2 sentences."**
 
-2. **Role** — "Who is the AI in this project?"
-   - Show available identity blocks: `node bin/prompt-blocks.js --category identity`
-   - User can pick one OR describe a custom role
+That's it. Infer everything else using the inference engine below.
 
-3. **Target platforms** — "Which platforms will you deploy to?"
-   - Multi-select: Claude Projects, Perplexity Spaces, Claude Code
-   - Default: Claude Projects + Perplexity Spaces
+### 1b. Inference engine — derive config from description
 
-4. **Complexity** — Based on the description, recommend a mode:
-   - Simple/standard project → **compiled** (block-based, continue to 1b)
-   - Complex multi-role agent → **standalone** with AI generation (continue to 1c)
-   - User already has a prompt → **standalone** paste-in (scaffold folder, skip to Step 3)
+Run these rules against the description to auto-populate the project config:
 
-### 1b. Compiled project setup
+**Role inference** (keywords → identity block):
+- "architect", "design", "infrastructure", "cloud", "azure", "aws" → `solutions-architect`
+- "engineer", "developer", "code", "build", "implement" → `senior-engineer`
+- "research", "analysis", "investigate", "study" → `research-partner`
+- "capture", "proposal", "federal", "deal", "RFP", "maximus" → `federal-deal-sa` (use `maximus-sa` profile)
+- No match → `solutions-architect` (default)
 
-5. **Domain expertise** — Show available domain blocks:
-   ```bash
-   node bin/prompt-blocks.js --category domains
-   ```
-   User picks from list or describes custom domains.
-   For custom domains: create a new block file at `prompts/blocks/domains/<id>.md`.
+**Domain inference** (keywords → domain blocks):
+- "cloud", "azure", "aws", "terraform", "infrastructure" → `cloud-infrastructure`
+- "federal", "govcon", "government", "compliance", "FedRAMP" → `govcon`
+- "web", "react", "frontend", "UI", "design" → `web-development`
+- "capture", "proposal", "RFP", "deal" → `govcon-capture`, `govcon-proposal`
+- No existing block match → standalone mode (AI generates custom domain content from research)
+
+**Platform inference:**
+- Default: Claude Projects + Perplexity Spaces
+- Add Claude Code if description contains: "code", "repo", "implement", "build", "develop", "engineering", "CLI"
+- Perplexity-only if description contains: "research only", "analysis only", "investigation"
+
+**Mode inference:**
+- If description triggers `maximus-sa` profile → compiled with `maximus-sa`
+- If ≥2 existing domain blocks match → compiled with matched blocks
+- If description suggests multi-role, custom workflow, or complex agent → standalone
+- If no domain blocks match → standalone with AI generation (Perplexity fills the gap)
+- Default for unknown domains → standalone
 
-6. **Research domains** (if Perplexity selected) — "What topics should Perplexity prioritize?"
-   Suggest based on selected domains. User accepts or customizes.
+**Research domain inference:**
+- Derive from description keywords + matched domain names
+- For standalone: use the full description as research seed
+- Always generate at least 2 research queries
+
+### 1c. Show confirmation card
+
+Instead of asking more questions, show what was inferred:
+
+```
+Project: <project-name>
+
+  Role:       <inferred identity block or "custom via research">
+  Domains:    <matched blocks or "custom — AI-generated from research">
+  Platforms:  <inferred list>
+  Mode:       <compiled | standalone>
+  Profile:    <matched profile or "none — custom blocks">
+  Research:   <inferred research topics>
+  Knowledge:  auto-generate from research
+
+  [Proceed] [Edit]
+```
 
-7. **Knowledge files** — "Reference documents to upload alongside? (compliance matrices, standards, playbooks)"
+- **Proceed** → run full pipeline (scaffold → research → generate → condense → results)
+- **Edit** → ask ONE targeted follow-up on the specific field to change
 
-8. **Claude Code extras** (only if Claude Code selected) — tech stack, commands, git identity
+### 1d. Build project and run pipeline
 
-9. **Build project folder:**
+After confirmation:
+
+1. **Scaffold folder:**
    ```bash
    mkdir -p prompts/projects/<project-name>/references
    ```
 
-   Write `prompt-config.yaml`:
-   ```yaml
-   project: <project-name>
-   description: <from intake>
-   mode: compiled
-   version: "1.0"
-   platforms: [claude-project, perplexity-space]
-   profile: null
-   blocks:
-     identity: [<matched-block>]
-     domains: [<selected-blocks>]
-     behaviors: []
-     formats: []
-     context: [<auto-selected by platform>]
-   research_domains: [<from intake>]
-   knowledge_files: [<from intake>]
-   ```
-
-   **Auto-add context blocks by platform:**
-   - Perplexity → `official-docs-first`, `flag-confidence`
-   - Claude Code → `vault-integration`, `mcp-servers`, `praxis-workflow`
-
-10. **Compile:** `node bin/prompt-compile.js <project-name>`
+2. **Write `prompt-config.yaml`** with inferred values
 
-11. → Go to **Step 4** (results + deployment)
+3. **If compiled mode:** `node bin/prompt-compile.js <project-name>` → Step 4
 
-### 1c. Standalone project with AI generation
+4. **If standalone mode:** → Step 2 (Perplexity research + generation)
 
-→ Go to **Step 2** (generate from scratch)
+**Auto-add context blocks by platform (compiled mode):**
+- Perplexity → `official-docs-first`, `flag-confidence`
+- Claude Code → `vault-integration`, `mcp-servers`, `praxis-workflow`
 
 ---
 
@@ -122,15 +191,13 @@ When invoked with a project name, detect the right action:
 
 **Triggered when:** standalone project exists but `system-prompt.md` is empty/missing, OR user explicitly requests generation.
 
-### 2a. Intake (if not already gathered in Step 1)
+### 2a. Use inference from Step 1
 
-1. **Description** — 2-3 sentences about the project
-2. **Role** — primary AI responsibility
-3. **Domains** — expertise areas (free-form, not limited to existing blocks)
-4. **Key behaviors** — rules beyond defaults
-5. **Target audience** — who reads the output
-6. **Knowledge files** — reference documents
-7. **Target platforms** — deployment targets
+All intake was gathered in Step 1 (description → inference engine → confirmation card).
+Do NOT ask additional questions here. Use the confirmed config from Step 1c.
+
+The description provides: role, domains, audience, platforms.
+Behaviors default from `_base`. Knowledge files auto-generate from research.
 
 ### 2b. Domain research via Perplexity
 
@@ -158,7 +225,7 @@ If Perplexity unavailable: state "Domain research could not be completed — pro
 
 ### 2c. Generate system-prompt.md
 
-Using intake + Perplexity research, generate following the 5-layer skeleton:
+Using intake + Perplexity research, generate following the 7-layer skeleton:
 
 ```markdown
 ---
@@ -189,15 +256,41 @@ generated_by: px-prompt
 ## Knowledge Interaction Rules
 [How to use reference files, when to cite, quote-before-answer]
 
-## Accuracy Standards
-- Flag confidence levels when synthesizing across sources
-- Distinguish verified facts from analytical inferences
-- If sources disagree, cite both and explain the discrepancy
-- Never fabricate version numbers, citations, or references
-- When information may be outdated, note this explicitly
+## Reasoning Approach
+Think step-by-step through complex problems. Break work into clear phases:
+1. **Understand** — Restate the question to confirm what's being asked
+2. **Research** — Check knowledge files and sources before forming an answer
+3. **Analyze** — Evaluate options, weigh tradeoffs, identify risks
+4. **Recommend** — Present a clear recommendation with rationale
+5. **Verify** — Self-check: does the answer address the actual question? Are claims sourced?
+
+For multi-step tasks, complete each step fully before moving to the next. Do not skip ahead or assume intermediate results. Show your reasoning when the logic matters — hide it when the answer is straightforward.
+
+## Quality Controls
+### Source Verification
+- Cross-reference claims against uploaded knowledge files before presenting as fact
+- When synthesizing from multiple sources, flag any contradictions
+- Distinguish between: verified (from knowledge files), corroborated (multiple sources agree), inferred (logical deduction), and speculative (educated guess)
+
+### Anti-Hallucination Protocol
+- Never fabricate version numbers, dates, statistics, citations, URLs, or API signatures
+- If you cannot find a specific fact in knowledge files or your training data, say so — do not approximate
+- When quoting standards, frameworks, or regulations: cite the specific document name and section
+- If a user asks about something not covered in knowledge files, explicitly state "this is not covered in the provided references" before offering general knowledge
+- For numerical claims (costs, timelines, metrics): only state numbers you can trace to a source
+- When information may be outdated (>12 months), flag it: "As of [date], [claim] — verify for current status"
+
+### Output Quality
+- Every recommendation includes rationale and tradeoffs
+- Tables for comparisons, not paragraphs
+- Structured outputs: use headers, bullets, numbered steps — not walls of text
+- Lead with the answer, then the reasoning — not the other way around
+- Every deliverable has a clear structure: BLUF (bottom line), body (evidence/analysis), next steps (actions)
+- Self-check before delivering: Does this answer the specific question? Are all claims sourced? Is the structure scannable?
 
 ## When Uncertain
 State uncertainty explicitly. Ask one clarifying question rather than guessing.
+Flag confidence level: HIGH (verified from sources), MEDIUM (corroborated), LOW (inferred or speculative).
 ```
 
 **Generation rules:**
@@ -205,7 +298,8 @@ State uncertainty explicitly. Ask one clarifying question rather than guessing.
 - Positive framing: "Do X" over "Don't do Y"
 - No few-shot examples (breaks Perplexity)
 - Under 5,000 characters
-- Include Accuracy Standards + When Uncertain (mandatory)
+- Quality Controls + When Uncertain sections are **mandatory** — never omit
+- Anti-Hallucination Protocol is **mandatory** in all generated prompts
 
 Write to `prompts/projects/<project-name>/system-prompt.md`.
 
@@ -213,19 +307,69 @@ Write to `prompts/projects/<project-name>/system-prompt.md`.
 
 → Go to **Step 3** (condense to platform outputs)
 
+### 2e. Generate knowledge files from Perplexity research
+
+**Always runs by default** after prompt generation. Do not ask — just generate.
+If the user declined in the confirmation card, skip this step.
+
+For each research domain:
+
+1. Run deep Perplexity queries per domain:
+   ```
+   perplexity_research: "Comprehensive overview of [domain]:
+   key concepts, current standards, best practices, common terminology,
+   frameworks, tools, and workflows. Include specific version numbers,
+   dates, and authoritative sources."
+   ```
+
+2. Structure findings into a knowledge file:
+   ```markdown
+   ---
+   domain: [domain-name]
+   generated: [today]
+   source: perplexity-research
+   ---
+
+   # [Domain] — Reference Guide
+
+   ## Key Concepts & Terminology
+   [Terms with definitions from research]
+
+   ## Current Standards & Frameworks
+   [Standards with version numbers and dates]
+
+   ## Best Practices
+   [Actionable practices from research]
+
+   ## Common Workflows
+   [Step-by-step workflows from research]
+
+   ## Tools & Technologies
+   [Relevant tools with current versions]
+
+   ## Sources
+   [Citations from Perplexity research]
+   ```
+
+3. Write to `prompts/projects/<project-name>/references/<domain-slug>.md`
+
+4. Update `prompt-config.yaml` knowledge_files list
+
+**Budget awareness:** Each knowledge file should be under 10,000 chars for Claude Projects upload limits. Split large domains into multiple files if needed.
+
 ---
 
 ## Step 3 — CONDENSE: Generate platform outputs from system-prompt.md
 
-**Triggered when:** standalone project has `system-prompt.md` but missing `space-instructions.md` or `CLAUDE.md`.
+**Triggered when:** standalone project has `system-prompt.md` but missing `space-instructions-perplexity.md` or `CLAUDE.md`.
 
 Read the full `system-prompt.md` as source.
 
 ### 3a. Generate Perplexity Space instructions
 
-**Target:** `space-instructions.md` | **Budget:** under 4,000 chars
+**Target:** `space-instructions-perplexity.md` | **Budget:** under 4,000 chars
 
-**Include:** identity, domain expertise, research domains, source priority, answer format, key frameworks (by name only)
+**Include:** identity, domain expertise, research domains, source priority, answer format, key frameworks (by name only), reasoning approach, accuracy standards, anti-hallucination rules
 **Exclude:** internal templates, scoring matrices, reference file content, deployment details, full tables
 
 **Output format:**
@@ -235,7 +379,27 @@ Read the full `system-prompt.md` as source.
 ## Research Domains
 ## Source Priority
 ## How to Answer
-## Accuracy Standards
+## Reasoning Approach
+## Quality & Accuracy Standards
+```
+
+**Mandatory reasoning section for Perplexity outputs:**
+```markdown
+## Reasoning Approach
+Think step-by-step: Understand the question → search sources → analyze findings → recommend with rationale → verify claims are sourced. Lead with the answer, then the evidence. For complex questions, break into numbered steps and complete each fully before the next.
+```
+
+**Mandatory quality section for Perplexity outputs:**
+```markdown
+## Quality & Accuracy Standards
+- Flag confidence level: HIGH (multiple sources confirm), MEDIUM (single source), LOW (inferred)
+- Never fabricate version numbers, statistics, citations, or URLs
+- If sources disagree, cite both and explain the discrepancy
+- When information may be outdated (>12 months), note the publication date
+- If you cannot find reliable sources, state that clearly rather than speculating
+- Distinguish verified facts from analytical inferences
+- Lead with the bottom line, then supporting evidence
+- Structure every response: answer first, reasoning second, sources third
 ```
 
 **Perplexity guardrails:**
@@ -248,7 +412,7 @@ Read the full `system-prompt.md` as source.
 
 **Target:** `CLAUDE.md` | **Budget:** under 250 lines
 
-**Include:** identity, behaviors, domain expertise, frameworks (one-line each), operating modes, quality controls
+**Include:** identity, behaviors, domain expertise, frameworks (one-line each), operating modes, quality controls, anti-hallucination rules
 **Exclude:** full scoring matrices, templates, reference file content, corporate data tables
 
 **Output format:**
@@ -269,10 +433,18 @@ Read the full `system-prompt.md` as source.
 - Self-check block for quality-critical outputs
 - Reference knowledge files by filename only
 
-### 3c. Validate budgets
+### 3c. Generate Claude Desktop project instructions (if Claude Projects is a target platform)
+
+**Target:** `project-instructions-claude-desktop.md` | **Budget:** under 2,500 chars
+
+**Include:** role, behavioral constraints, domain expertise (condensed), output format, quality controls, when uncertain
+**Exclude:** full domain details (those go in knowledge files), reference content, deployment details
+
+### 3d. Validate budgets
 
 After generating, check:
-- `space-instructions.md` under 4,000 chars
+- `space-instructions-perplexity.md` under 4,000 chars
+- `project-instructions-claude-desktop.md` under 2,500 chars
 - `CLAUDE.md` under 250 lines
 
 If over budget: flag and suggest sections to trim.
@@ -293,26 +465,26 @@ node bin/prompt-compile.js <project-name>
 ### 4b. Show results table
 
 ```
-| Output               | Chars  | Budget | Status |
-|----------------------|--------|--------|--------|
-| system-prompt.md     | X      | —      | Source |
-| project-instructions | X      | 2,500  | OK     |
-| space-instructions   | X      | 4,000  | OK     |
-| CLAUDE.md            | X lines| 250 ln | OK     |
-| references/          | N files| —      | Upload |
+| Output                                   | Chars  | Budget | Status |
+|------------------------------------------|--------|--------|--------|
+| system-prompt.md                         | X      | —      | Source |
+| project-instructions-claude-desktop.md   | X      | 2,500  | OK     |
+| space-instructions-perplexity.md         | X      | 4,000  | OK     |
+| CLAUDE.md                                | X lines| 250 ln | OK     |
+| references/                              | N files| —      | Upload |
 ```
 
 ### 4c. Deployment instructions
 
-**Claude Projects (claude.ai):**
+**Claude Desktop / Projects (claude.ai):**
 1. Open project at claude.ai/projects → "Set project instructions"
-2. Standalone: paste `system-prompt.md` | Compiled: paste `project-instructions.md`
+2. Standalone: paste `system-prompt.md` | Compiled: paste `project-instructions-claude-desktop.md`
 3. If `references/` exists: upload each `.md` file as project knowledge
 4. Save
 
 **Perplexity Spaces:**
 1. Open Space → Settings → Answer Instructions
-2. Paste `space-instructions.md`
+2. Paste `space-instructions-perplexity.md`
 3. Save
 
 **Claude Code:**
@@ -321,6 +493,8 @@ node bin/prompt-compile.js <project-name>
 ### 4d. Offer next actions
 - "Edit the prompt? I'll regenerate platform outputs after."
 - "Want to regenerate? Run `/px-prompt <project-name>` again."
+- "Generate knowledge files from research? I'll create reference docs from Perplexity."
+- "Scan project folder for recommended changes? Run `/px-prompt --scan <project-name>`."
 
 ---
 
@@ -334,10 +508,10 @@ node bin/prompt-compile.js --all --diff
 
 Show summary table:
 ```
-| Project | CLAUDE.md | Project Instr. | Space Instr. | Changes |
-|---------|-----------|----------------|--------------|---------|
-| praxis  | 3,534     | 1,316 ✓        | 1,529 ✓      | none    |
-| maximus | —         | —              | 3,977 ✓      | standalone |
+| Project      | CLAUDE.md | Claude Desktop    | Perplexity       | Changes    |
+|--------------|-----------|-------------------|------------------|------------|
+| praxis       | 3,534     | 1,316 ✓           | 1,529 ✓          | none       |
+| maximus      | —         | —                 | 3,977 ✓          | standalone |
 ```
 
 For standalone projects, report validation status instead of compilation status.
@@ -346,18 +520,654 @@ Print deployment reminders for any project with changes.
 
 ---
 
+## Step 6 — SCAN & EDIT: Analyze and update existing project prompts
+
+**Triggered when:** `/px-prompt --scan <project-name>` or user asks to review/update existing prompts.
+
+### 6a. Read all project files
+1. Read `prompt-config.yaml` for project metadata
+2. Read `system-prompt.md` (source of truth)
+3. Read all platform outputs (`space-instructions-perplexity.md`, `project-instructions-claude-desktop.md`, `CLAUDE.md`)
+4. Read all files in `references/` directory
+5. List any other files in the project folder
+
+### 6b. Analyze for issues
+Check each file against these criteria:
+
+**Quality checks:**
+- Does system-prompt.md include Quality Controls section?
+- Does system-prompt.md include Anti-Hallucination Protocol?
+- Does system-prompt.md include When Uncertain section?
+- Are platform outputs in sync with system-prompt.md content?
+- Are file naming conventions correct (platform suffixes)?
+- Are all referenced knowledge files present in references/?
+
+**Budget checks:**
+- `space-instructions-perplexity.md` under 4,000 chars?
+- `project-instructions-claude-desktop.md` under 2,500 chars?
+- `CLAUDE.md` under 250 lines?
+
+**Currency checks (via Perplexity):**
+- Are domain-specific terms, standards, and versions still current?
+- Have any referenced frameworks been updated or renamed?
+
+### 6c. Present findings and offer edits
+Show a structured report:
+```
+| Check                     | Status | Details                           |
+|---------------------------|--------|-----------------------------------|
+| Quality Controls section  | PASS   | Present in system-prompt.md       |
+| Anti-Hallucination        | FAIL   | Missing — will add                |
+| Budget: Perplexity        | PASS   | 2,392 / 4,000 chars              |
+| File naming               | FAIL   | Uses old convention — will rename |
+| Knowledge files           | WARN   | 0 reference files — consider adding |
+```
+
+For each FAIL:
+- Show the specific fix needed
+- Apply the fix (with user confirmation for content changes)
+- Regenerate affected platform outputs
+
+### 6d. Reference existing files for future changes
+After scanning, maintain awareness of:
+- Which sections of the prompt cover which domains
+- What knowledge files exist and what they contain
+- What the version history looks like (from prompt-config.yaml)
+
+Use this context when the user asks for changes — edit in place rather than regenerating from scratch.
+
+---
+
+## Step 7 — DEAL: Fast-path for Maximus capture deals
+
+**Triggered when:** `/px-prompt --deal <deal-name>`
+
+This is the fastest path for the most common task: creating a new Maximus capture deal.
+**Zero questions asked** — Perplexity researches the deal from the name alone.
+
+### 7a. Deal name + optional type
+
+The deal name IS the research query. Optional `--type` sets the deal template.
+
+```
+/px-prompt --deal irs-masterfile
+/px-prompt --deal irs-masterfile --type recompete
+/px-prompt --deal benefeds --type idiq
+/px-prompt --deal dha-tricare --type new-start
+```
+
+If no `--type` given, Perplexity research determines the type automatically (incumbent found = recompete, no incumbent = new-start, IDIQ/BPA language = task-order).
+
+If the user provides additional context alongside the name (e.g., `/px-prompt --deal irs-masterfile "IRS Individual Masterfile modernization, current incumbent Accenture"`), use it to seed the research. Otherwise, research from the name alone.
+
+### Deal Type Templates
+
+Each deal type activates different emphasis in the compiled output and research priorities:
+
+**Recompete** (`--type recompete`)
+- Incumbent defense: Remind → Reveal → Reimagine activated
+- Ghost theme matrix pre-seeded for incumbent weaknesses
+- Research priority: incumbent CPARS, protest history, pain points
+- Transition emphasis: continuity of operations, Day-1 readiness, retained mission knowledge
+- Scoring weight: Past Performance and Customer Relationship weighted higher
+- Phase start: Mid Capture (incumbent already has intel advantage)
+
+**New Start** (`--type new-start`)
+- Discovery-first: OSINT-heavy, agency mission deep dive
+- No incumbent to ghost — focus on innovation and fresh approach
+- Research priority: agency strategic plan, IG/GAO findings, budget trends, technology landscape
+- Solution emphasis: differentiated approach, not transition continuity
+- Scoring weight: Technical Approach and Solution Readiness weighted higher
+- Phase start: Shaping
+
+**Task Order** (`--type task-order`)
+- Vehicle-constrained: must map to existing IDIQ/BPA/GWACs scope
+- Fast timeline: typically 30-60 day turnaround
+- Research priority: task order history on the vehicle, rates, ceiling remaining
+- Emphasis: price competitiveness, rapid staffing, past performance on same vehicle
+- Scoring weight: Price and Past Performance weighted higher
+- Phase start: Mid Capture (vehicle already won)
+
+**IDIQ/BPA Pursuit** (`--type idiq`)
+- Long-game: win the vehicle, compete TOs later
+- Research priority: vehicle scope, evaluation criteria, small business targets
+- Emphasis: broad capabilities, teaming strategy, rate structure
+- Scoring weight: Technical Approach and Management Plan weighted higher
+- Phase start: Shaping
+
+The deal type is stored in `prompt-config.yaml` as `deal_type` and influences:
+1. Which Perplexity queries run (research priority above)
+2. The initial capture phase set in the deal config
+3. Which sections of the scorecard are weighted in gate reviews
+4. The Perplexity Space research domains (type-specific sources)
+
+### 7b. Research the deal via Perplexity (mandatory)
+
+Run these queries in parallel to build the deal intelligence:
+
+**Query 1 — Contract & procurement data:**
+```
+perplexity_research: "<deal-name> federal contract. Search SAM.gov, 
+USASpending.gov, FPDS.gov, GovWin. Find: agency, program office, 
+contract type, vehicle, NAICS, set-aside, estimated value, 
+period of performance, current status, solicitation number."
+```
+
+**Query 2 — Incumbent & competitive landscape:**
+```
+perplexity_research: "<deal-name> incumbent contractor. Who currently 
+holds this contract? Contract value, CPARS ratings if available, 
+key subcontractors, known competitors pursuing recompete. 
+Search GovConWire, Washington Technology, Bloomberg Government, 
+10-K filings, protest history on GAO.gov."
+```
+
+**Query 3 — Agency mission & pain points:**
+```
+perplexity_ask: "What is the mission context for <deal-name>? 
+Agency strategic plan priorities, relevant IG/GAO findings, 
+congressional testimony, budget trends, technology modernization 
+initiatives. What problems is the agency trying to solve?"
+```
+
+**Query 4 — Maximus positioning:**
+```
+perplexity_search: "Maximus Inc <agency from Q1> past performance 
+contracts site:usaspending.gov OR site:sam.gov"
+```
+
+**Source priority for deal research:**
+1. SAM.gov — active solicitations, contract awards, vendor registrations
+2. USASpending.gov — contract values, spending history, awardees
+3. FPDS.gov — detailed procurement records, contract modifications
+4. GovWin IQ — opportunity tracking, competitive intelligence (if accessible)
+5. Bloomberg Government (BGOV) — contract analytics, market intelligence
+6. GovConWire / Washington Technology — industry news, award announcements
+7. GAO.gov — protest decisions, audit findings
+8. Agency strategic plans & budget justifications — mission context
+9. SEC 10-K filings — competitor revenue, backlog, strategy statements
+
+If Perplexity cannot find data for a field, mark it `[RESEARCH NEEDED]` — not TBD.
+
+### 7c. Build deal context from research
+
+Structure Perplexity findings into a deal intelligence file:
+
+```markdown
+---
+deal: <deal-name>
+generated: <today>
+source: perplexity-research
+confidence: <HIGH if SAM.gov/FPDS data found, MEDIUM if news only, LOW if inferred>
+---
+
+# Deal Intelligence: <deal-name>
+
+## Opportunity Profile
+| Field | Value | Source |
+|-------|-------|--------|
+| Agency | <from research> | <source> |
+| Program | <from research> | <source> |
+| Solicitation # | <if found> | SAM.gov |
+| Contract Type | <from research> | <source> |
+| Vehicle | <from research> | <source> |
+| NAICS | <from research> | <source> |
+| Set-Aside | <from research> | <source> |
+| Estimated Value | <from research> | <source> |
+| POP | <from research> | <source> |
+| Status | <from research> | <source> |
+
+## Incumbent & Competition
+| Competitor | Role | Contract Value | CPARS | Strengths | Weaknesses |
+|-----------|------|---------------|-------|-----------|------------|
+| <incumbent> | Incumbent | <value> | <rating> | <from research> | <from research> |
+| <competitor 2> | Challenger | — | — | <from research> | <from research> |
+
+## Agency Mission Context
+<Pain points, strategic priorities, IG/GAO findings from Query 3>
+
+## Maximus Positioning
+<Existing contracts at this agency, relevant past performance, platform fit from Query 4>
+
+## Research Gaps
+<Fields marked [RESEARCH NEEDED] — what couldn't be found and where to look manually>
+
+## Sources
+<Citations from all Perplexity queries>
+```
+
+Write to `prompts/projects/<deal-name>/references/<deal-name>-intel.md`
+
+### 7d. Auto-scaffold with maximus-sa
+
+1. `mkdir -p prompts/projects/<deal-name>/references`
+2. Write `prompt-config.yaml`:
+   - Profile: `maximus-sa`
+   - `deal_type`: from `--type` flag or auto-detected from research (incumbent found → recompete, none → new-start)
+   - `capture_phase`: initial phase based on deal type (see template table above)
+   - `knowledge_packs`: deal-context + corporate-reference, vars populated from research
+   - `overrides.perplexity_space_append.research_domains`: **deal-specific** research domains (see 7f)
+3. Write deal intel to `references/<deal-name>-intel.md`
+4. Compile: `node bin/prompt-compile.js <deal-name>`
+5. Render knowledge packs: `node bin/prompt-knowledge.js <deal-name>`
+
+### 7e. Show confirmation card with research findings
+
+```
+DEAL: irs-masterfile
+━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+Agency:      IRS (from USASpending)
+Program:     Individual Masterfile Modernization
+Incumbent:   Accenture Federal Services (from FPDS)
+Value:       $2.1B ceiling (from SAM.gov)
+Type:        recompete (auto-detected — incumbent found)
+Phase:       Mid Capture
+NAICS:       541512
+Vehicle:     GSA Alliant 2
+Status:      Active — recompete expected FY2026
+Confidence:  HIGH (SAM.gov + FPDS data)
+
+Maximus PP:  2 contracts at IRS (from USASpending)
+Gaps:        Key personnel [RESEARCH NEEDED]
+━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+Files created:
+  ✓ CLAUDE.md (31,646 chars)
+  ✓ space-instructions-perplexity.md (3,976 chars) — deal-specific research domains
+  ✓ references/irs-masterfile-intel.md
+  ✓ knowledge/deal-context.md
+  ✓ knowledge/maximus-corporate.md
+
+[Deploy] [Advance phase] [Edit intel] [Re-research]
+```
+
+### 7f. Per-deal Perplexity Space instructions
+
+The compiled `space-instructions-perplexity.md` includes the standard Maximus SA persona, but the **Research Domains** section is customized per deal using `perplexity_space_append.research_domains` in the project config.
+
+**Auto-generate research domains from deal type + agency:**
+
+For **recompete** deals:
+```yaml
+research_domains: |
+  - <agency> <program> contract history, modifications, and performance (USASpending, FPDS)
+  - <incumbent> federal performance, CPARS, protest history, recent awards
+  - <agency> strategic plan, IG reports, GAO findings, budget justifications
+  - <agency> technology modernization initiatives, cloud migration, IT spend
+  - Maximus past performance and existing relationships at <agency>
+  - Competitive landscape: who else is pursuing <program> recompete
+```
+
+For **new-start** deals:
+```yaml
+research_domains: |
+  - <agency> mission priorities, strategic plan, leadership agenda
+  - <agency> IG/GAO findings related to <program area>
+  - <program area> technology landscape, vendor market, innovation trends
+  - Congressional testimony and budget trends for <agency>
+  - Maximus capabilities and past performance in <program area>
+  - Similar programs at other agencies — lessons learned, best practices
+```
+
+For **task-order** deals:
+```yaml
+research_domains: |
+  - <vehicle> task order history, award patterns, ceiling status
+  - <agency> recent task orders on <vehicle>, evaluation preferences
+  - Competitor rates and staffing patterns on <vehicle>
+  - <program> scope, deliverables, and performance standards
+  - Maximus existing task orders on <vehicle>
+```
+
+This means each deal's Perplexity Space is tuned to research THAT deal's specific intelligence needs.
+
+**Total: 0 questions → deal name only → Perplexity fills everything → compiled project + deal-specific intel + per-deal research space + all outputs.**
+
+---
+
+## Step 8 — EDIT: Targeted edit with auto-regeneration
+
+**Triggered when:** `/px-prompt --edit <project-name> "<change description>"`
+
+### 8a. Read current state
+
+1. Read `prompt-config.yaml` to determine mode (compiled vs standalone)
+2. Read the source file:
+   - Compiled: read the relevant block files based on what the change targets
+   - Standalone: read `system-prompt.md`
+3. Read all platform outputs for comparison
+
+### 8b. Apply the targeted edit
+
+Based on the change description, identify which section(s) to modify:
+
+**For standalone projects:**
+1. Edit `system-prompt.md` — apply the change to the specific section
+2. Auto-regenerate all platform outputs (Step 3)
+3. Show diff of what changed in each file
+
+**For compiled projects:**
+1. Identify which block file contains the content to change
+2. Edit the block file (both FULL and CONDENSED variants)
+3. Re-compile: `node bin/prompt-compile.js <project-name> --diff`
+4. Show diff of what changed
+
+### 8c. Report
+
+```
+EDIT APPLIED
+━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+Changed: <file(s) modified>
+Regenerated: <platform outputs updated>
+Budget: perplexity ✓ | claude-desktop ✓ | CLAUDE.md ✓
+━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+```
+
+**Examples:**
+- `/px-prompt --edit elect-azure "Add VITA SEC530 cybersecurity standard to domain expertise"`
+- `/px-prompt --edit maximus "Update CEO to new name"`
+- `/px-prompt --edit praxis "Add Python to tech stack"`
+
+---
+
+## Step 9 — DASHBOARD: Project index with status
+
+**Triggered when:** `/px-prompt --dashboard`
+
+### 9a. Gather data
+
+```bash
+node bin/prompt-compile.js --dashboard
+```
+
+### 9b. Show dashboard
+
+```
+PROMPT ENGINE DASHBOARD
+━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+Project       Mode       Perplexity    Claude Proj   CLAUDE.md   Refs  Updated     Stale?
+─────────────────────────────────────────────────────────────────────────────────────────
+maximus       compiled   3,976 ✓       4,261 ⚠       30,665 ✓    4     2026-04-04  No
+elect-azure   standalone 2,392 ✓       —             3,098 ✓     0     2026-04-04  No
+praxis        compiled   1,626 ✓       1,404 ✓       3,417 ✓     0     2026-04-04  No
+━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+
+Staleness: projects not updated in >30 days marked stale.
+Budgets: ✓ under budget, ⚠ over soft budget, ✗ over hard budget.
+```
+
+### 9c. Offer actions for flagged projects
+
+For stale projects: "Run `/px-prompt --refresh <name>` to update research."
+For over-budget: "Run `/px-prompt --edit <name>` to trim."
+For missing outputs: "Run `/px-prompt <name>` to generate."
+
+---
+
+## Step 10 — REFRESH: Re-run Perplexity research and update
+
+**Triggered when:** `/px-prompt --refresh <project-name>`
+
+### 10a. Read current project
+
+1. Read `prompt-config.yaml` for research domains and mode
+2. Read current `system-prompt.md` (standalone) or relevant block files (compiled)
+3. Note the current domain expertise content
+
+### 10b. Re-run Perplexity research
+
+For each research domain (from config or inferred from prompt content):
+
+```
+perplexity_ask: "What are the current best practices and standards for [domain] 
+in 2025-2026? Focus on changes since [last updated date]. 
+Key terminology updates, new frameworks, deprecated standards."
+```
+
+### 10c. Diff and propose updates
+
+Compare research findings against current prompt content:
+
+```
+REFRESH REPORT: <project-name>
+━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+Domain: cloud-infrastructure
+  Current: "Azure Well-Architected Framework (5 pillars)"
+  Updated: "Azure Well-Architected Framework (6 pillars — Sustainability added 2025)"
+  → SUGGEST: Update Domain Expertise section
+
+Domain: govcon
+  Current: "CMMC 2.0"
+  Updated: "CMMC 2.0 — Final Rule effective Dec 2024, Level 2 assessments active"
+  → SUGGEST: Add CMMC enforcement timeline
+━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+```
+
+### 10d. Apply updates
+
+For each suggested update, show the change and apply:
+- Standalone: edit `system-prompt.md`, regenerate platform outputs
+- Compiled: edit relevant block file(s), recompile
+- Knowledge files: regenerate from fresh research
+
+Update the `version` field in `prompt-config.yaml` and `last_updated` timestamp.
+
+---
+
+## Step 11 — DEPLOY: Copy outputs with deployment URLs
+
+**Triggered when:** `/px-prompt --deploy <project-name>`
+
+### 11a. Read project config
+
+Determine which platforms are targets from `prompt-config.yaml`.
+
+### 11b. Deploy sequence (per platform)
+
+**For Claude Projects / Desktop:**
+1. Read `system-prompt.md` (standalone) or `project-instructions-claude-desktop.md` (compiled)
+2. Copy content to clipboard: `cat <file> | pbcopy`
+3. Print: "Copied to clipboard. Paste at: claude.ai/projects → Set project instructions"
+4. List knowledge files to upload: all `.md` files from `references/` AND `knowledge/`
+5. Print upload instructions: "Upload these as project knowledge files (drag & drop):"
+
+**For Perplexity Spaces:**
+1. Read `space-instructions-perplexity.md`
+2. Copy content to clipboard: `cat <file> | pbcopy`
+3. Print: "Copied to clipboard. Paste at: perplexity.ai → Space Settings → Answer Instructions"
+4. List knowledge files to upload as Space sources: all `.md` files from `references/` AND `knowledge/`
+5. Print: "Upload these as Space sources (Add Sources → Files):"
+6. **Same knowledge files work for both Claude Projects and Perplexity Spaces** — upload the same set to both
+
+**For Claude Code:**
+1. If project has a `repo_root` in vars: `cp CLAUDE.md <repo_root>/CLAUDE.md`
+2. Print: "CLAUDE.md copied to repo root."
+3. If no repo_root: "Copy CLAUDE.md to your project repo root manually."
+
+### 11c. Deploy one platform at a time
+
+Since clipboard can only hold one thing, deploy sequentially:
+
+```
+DEPLOY: dha-tricare
+━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+
+[1/3] Claude Projects
+  → Copied system-prompt.md to clipboard (5,824 chars)
+  → Paste at: claude.ai/projects → Set project instructions
+
+  Upload these as project knowledge files:
+    references/dha-tricare-intel.md      (deal intelligence)
+    knowledge/deal-context.md            (deal context)
+    knowledge/maximus-corporate.md       (corporate reference)
+  Press Enter when done...
+
+[2/3] Perplexity Spaces
+  → Copied space-instructions-perplexity.md to clipboard (3,965 chars)
+  → Paste at: perplexity.ai → Space Settings → Answer Instructions
+
+  Upload these as Space sources (Add Sources → Files):
+    references/dha-tricare-intel.md      (deal intelligence)
+    knowledge/deal-context.md            (deal context)
+    knowledge/maximus-corporate.md       (corporate reference)
+  Press Enter when done...
+
+[3/3] Claude Code
+  → CLAUDE.md → /path/to/repo/CLAUDE.md
+  Done.
+
+━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+DEPLOYED to 3 platforms.
+Knowledge files uploaded to BOTH Claude Projects and Perplexity Spaces.
+```
+
+---
+
+## Step 12 — ADVANCE: Move deal to next capture phase
+
+**Triggered when:** `/px-prompt --advance <deal-name>`
+
+Deals progress through the PAMASI lifecycle. This command advances the deal to the next phase, updates scoring expectations, runs phase-appropriate research, and flags what's missing.
+
+### 12a. Read current deal state
+
+1. Read `prompt-config.yaml` — get `capture_phase` and `deal_type`
+2. Read `references/<deal-name>-intel.md` — current intelligence
+3. Read latest compiled outputs — current scoring context
+
+### 12b. Determine next phase
+
+| Current Phase | Next Phase | Gate Criteria |
+|--------------|------------|---------------|
+| Shaping | Mid Capture | PAMASI at P→A. Sections I + XI GREEN. |
+| Mid Capture | Pre-Proposal | PAMASI at A→M→S. All 11 GREEN at mid-capture level. |
+| Pre-Proposal | Pre-Submission | PAMASI at S→I. All 11 GREEN at proposal level. Red Team complete. |
+| Pre-Submission | Orals | Zero deficiencies. All findings resolved. Compliance verified. |
+| Orals | Post-Submit | Q&A matrix complete. Presentations rehearsed. |
+
+### 12c. Run phase-transition Perplexity research
+
+Each phase transition triggers different research:
+
+**Shaping → Mid Capture:**
+```
+perplexity_research: "<deal-name> latest developments. New solicitation 
+activity on SAM.gov, draft RFP, industry day announcements, Q&A responses, 
+amendment history. Search SAM.gov, GovConWire, agency procurement forecasts."
+```
+
+**Mid Capture → Pre-Proposal:**
+```
+perplexity_research: "<deal-name> RFP release. Final solicitation, 
+evaluation criteria, Section L/M structure, proposal due date, 
+Q&A deadline, amendment status. Search SAM.gov, beta.sam.gov."
+```
+
+**Pre-Proposal → Pre-Submission:**
+```
+perplexity_research: "<deal-name> competitor activity. Who submitted 
+questions? Protest signals? Teaming announcements? Rate benchmarks 
+for <NAICS>. Search GovConWire, Bloomberg Government, GAO.gov."
+```
+
+**Pre-Submission → Orals:**
+```
+perplexity_ask: "For <agency> <program>, what are the most common 
+oral presentation evaluation criteria and question patterns? 
+Recent oral presentations at <agency> — format, duration, topics."
+```
+
+### 12d. Update deal config and outputs
+
+1. Update `capture_phase` in `prompt-config.yaml`
+2. Update deal intel with new research findings
+3. Recompile: `node bin/prompt-compile.js <deal-name> --diff`
+4. Update Perplexity Space research domains for the new phase
+5. Regenerate knowledge packs if deal context changed
+
+### 12e. Show phase transition report
+
+```
+PHASE ADVANCE: irs-masterfile
+━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+Previous:    Shaping
+Current:     Mid Capture
+PAMASI:      P → A (Approach stage)
+Deal Type:   recompete
+
+Gate Check (Shaping → Mid Capture):
+  I  Customer & Mission    ✓ GREEN
+  XI Competitive           ✓ GREEN
+  II Architecture          ○ Not yet scored (expected at Mid Capture)
+
+New Intel:
+  • Draft RFP posted on SAM.gov (2026-03-15)
+  • Industry day scheduled for April 20
+  • Amendment 1 added CMMC L2 requirement
+
+Updated Files:
+  ✓ prompt-config.yaml (phase: Mid Capture)
+  ✓ references/irs-masterfile-intel.md (new findings appended)
+  ✓ space-instructions-perplexity.md (research domains updated)
+
+Next Gate: Pre-Proposal — requires all 11 GREEN at mid-capture level
+━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+```
+
+### 12f. Flag missing items for next gate
+
+After advancing, check what's needed for the NEXT gate and flag gaps:
+
+```
+NEXT GATE CHECKLIST: Pre-Proposal
+  □ Solution architecture (Section II) — not started
+  □ Staffing model (Section X) — not started
+  □ Risk register (Section VII) — not started
+  □ Win themes with FBP proof (Section XI) — 1 of 3 complete
+  □ Competitive ghost matrix — not started
+  
+  5 items needed before Pre-Proposal gate.
+  Run /px-prompt --advance irs-masterfile when ready.
+```
+
+---
+
 ## Rules
 
 ### Always
 - `_base` behaviors (no-flattery, verify, recommend, handle-uncertainty) are included in every project — non-negotiable
+- Quality Controls section with Anti-Hallucination Protocol is **mandatory** in all generated prompts
 - Accuracy Standards section is mandatory in all Perplexity outputs
-- When Uncertain section is mandatory in all Claude Project outputs
+- When Uncertain section with confidence levels is mandatory in all outputs
 - Never ask for repo URL, vault path, or git email unless Claude Code is a target platform
+- **Never hardcode project names in menus or options** — discover dynamically
 
 ### Platform-specific
 - **Perplexity**: no few-shot examples, no URLs, conditional language, search-friendly terms
 - **Claude Code**: positive framing, no "CRITICAL YOU MUST", self-check blocks
-- **Claude Projects**: 5-layer skeleton (Role, Constraints, Format, Knowledge Rules, Failure Handling)
+- **Claude Desktop / Projects**: 7-layer skeleton (Role, Constraints, Expertise, Format, Knowledge Rules, Quality Controls, When Uncertain)
+
+### File naming
+- **Always** use platform-suffixed filenames: `space-instructions-perplexity.md`, `project-instructions-claude-desktop.md`
+- `CLAUDE.md` keeps its name (convention for Claude Code)
+- `system-prompt.md` keeps its name (it's the source of truth, not platform-specific)
+
+### Quality defaults (mandatory in all generated prompts)
+- Anti-Hallucination Protocol: never fabricate, cite sources, flag confidence, distinguish fact from inference
+- Source Verification: cross-reference knowledge files, flag contradictions
+- Output Quality: rationale with recommendations, structured outputs, self-check before delivery
+- Confidence Levels: HIGH (verified), MEDIUM (corroborated), LOW (inferred/speculative)
+
+### Knowledge file generation
+- Always generate knowledge files from Perplexity research by default — don't ask
+- Each knowledge file under 10,000 chars (Claude Projects upload limit)
+- Structure: Key Concepts, Standards, Best Practices, Workflows, Tools, Sources
+- Write to `references/<domain-slug>.md`
+
+### Project folder scanning
+- Read all existing project files before suggesting changes
+- Edit in place rather than regenerating when the user requests specific changes
+- Track what each section covers so future edits are surgical
+- Flag outdated terminology, missing quality sections, budget overruns
 
 ### Generation
 - ALWAYS run Perplexity research before generating system prompts (Step 2)
@@ -365,9 +1175,9 @@ Print deployment reminders for any project with changes.
 - Generated prompts are starting points — tell user to review and refine
 - If Perplexity unavailable: proceed with training data, flag for review
 
-### Block matching (compiled mode)
-- Role → identity block: "architect" → `solutions-architect`, "engineer" → `senior-engineer`, "researcher" → `research-partner`
-- Domain keywords → domain blocks: "cloud/azure/aws" → `cloud-infrastructure`, "federal/govcon" → `govcon`, "web/react" → `web-development`
-- If no match: create custom block on the fly
-- Auto-add `official-docs-first` + `flag-confidence` for Perplexity targets
-- Auto-add `vault-integration` + `mcp-servers` + `praxis-workflow` for Claude Code targets
+### Inference and block matching
+- See Step 1b for full inference rules (role, domain, platform, mode, research domains)
+- If no domain blocks match → standalone mode with AI generation, not compiled with empty blocks
+- Auto-add `official-docs-first` + `flag-confidence` for Perplexity targets (compiled mode)
+- Auto-add `vault-integration` + `mcp-servers` + `praxis-workflow` for Claude Code targets (compiled mode)
+- **Maximum 1 question** for new projects (description). Show confirmation card, not a questionnaire.