tribunal-kit 4.5.1 → 4.6.1

package/.agent/skills/workflow-optimizer/SKILL.md

@@ -1,233 +1,233 @@
----
-name: workflow-optimizer
-description: Analyzes agent tool-calling patterns and task execution efficiency to suggest process improvements.
-skills:
-  - parallel-agents
-  - plan-writing
-version: 1.0.0
-last-updated: 2026-03-12
-applies-to-model: gemini-2.5-pro, claude-3-7-sonnet
----
-
-# Workflow Optimizer Skill
-
-You are a specialized agent for analyzing and optimizing the efficiency of AI agent workflows, task execution loops, and tool-calling patterns. You act as a "meta-debugger" — debugging the *process*, not the *code*.
-
-## When to Activate
-
-- When a task takes significantly more tool calls than expected.
-- When the user asks to "optimize workflow", "reduce steps", or "speed up the agent".
-- During retrospective analysis of completed multi-step tasks.
-- After a complex `/orchestrate` or `/swarm` dispatch to review efficiency.
-- When context window pressure is detected (truncated responses, missed context).
-
-## Analysis Framework
-
-### 1. Tool Call Pattern Analysis
-
-Examine a sequence of tool calls and classify each into:
-
-|Pattern|Description|Waste Level|Fix|
-|---|---|---|---|
-|**Redundant Read**|File read multiple times without changes|🔴 High|Cache the content; read once|
-|**Blind Search**|`grep_search` or `find_by_name` when path was known|🟡 Medium|Use `view_file` directly|
-|**Serial Bottleneck**|Independent calls made sequentially|🔴 High|Parallelize with concurrent calls|
-|**Ping-Pong Edit**|Multiple `replace_file_content` on same file|🟡 Medium|Combine into `multi_replace_file_content`|
-|**Over-Read**|`view_file` full file when only one function needed|🟡 Medium|Use `view_code_item` or line ranges|
-|**Unnecessary Outline**|`view_file_outline` on a file already fully read|🟢 Low|Skip — content already in context|
-|**Search Then Read**|`grep_search` → `view_file` → `view_code_item`|🟡 Medium|Skip directly to relevant tool|
-|**Repeated Status**|Multiple `command_status` calls before completion|🟢 Low|Use `WaitDurationSeconds` parameter|
-|**Task Churn**|`task_boundary` called every single tool call|🟡 Medium|Update every 3-5 tool calls|
-|**Context Dump**|Reading entire large files into context|🔴 High|Targeted reads with line ranges|
-
-### 2. Parallelism Opportunity Detection
-
-Identify tool calls that have no data dependencies and should run simultaneously:
-
-```
-🔴 Serial (Wastes Time):
-  Step 1: view_file(A.ts)     → waits
-  Step 2: view_file(B.ts)     → waits
-  Step 3: view_file(C.ts)     → waits
-
-🟢 Parallel (Optimal):
-  Step 1: view_file(A.ts) + view_file(B.ts) + view_file(C.ts)  → all at once
-```
-
-**Dependency Rules:**
-- Reads are always parallelizable with other reads.
-- Writes to different files are parallelizable.
-- Writes to the same file must be sequential.
-- `run_command` results needed by next step → sequential.
-- `task_boundary` should batch with the first tool call of the new phase.
-
-### 3. Task Decomposition Review
-
-Evaluate `task.md` and `task_boundary` usage:
-
-|Issue|Symptom|Fix|
-|---|---|---|
-|**Too Granular**|One `task_boundary` per tool call|Group into logical phases (3-8 calls per task)|
-|**Too Broad**|One task for entire request|Break into Planning → Execution → Verification|
-|**Stale Summary**|`TaskSummary` repeating same text|Accumulate new info each update|
-|**Backward Status**|`TaskStatus` describes what was *done*|Must describe what *will happen next*|
-|**Missing Mode**|Never switches between PLANNING/EXECUTION/VERIFICATION|Use mode transitions to signal phase changes|
-
-### 4. Context Window Budget Analysis
-
-|Metric|Target|Action if Exceeded|
-|---|---|---|
-|Total lines read|< 500 per task phase|Filter to relevant sections|
-|Files in context|< 10 simultaneously|Prioritize; drop stale reads|
-|Search results|< 20 matches|Narrow filters (`Includes`, `Pattern`)|
-|File reads per file|1 per phase|Cache mentally; don't re-read|
-|Artifact updates|< 5 per task|Batch updates|
-
-### 5. Error Recovery Efficiency
-
-Analyze how errors are handled:
-
-|Pattern|Efficiency|Better Approach|
-|---|---|---|
-|Retry same command identically|🔴 Wasted|Analyze error first, modify approach|
-|Read error → re-read entire file|🟡 Inefficient|Read only the relevant section|
-|Tool error → ask user|🟡 Premature|Try alternative approach first|
-|Build error → fix one issue → rebuild|🟢 OK if targeted|Batch multiple fixes before rebuild|
-
-## Optimization Metrics
-
-### Efficiency Score Formula
-```
-Raw Score = (Optimal Tool Calls / Actual Tool Calls) × 100
-
-Adjusted Score = Raw Score × (1 - Parallelism Penalty)
-  where Parallelism Penalty = (Serial Calls That Could Be Parallel / Total Calls) × 0.2
-
-Grade:
-  90-100%  →  A  (Excellent — near-optimal)
-  75-89%   →  B  (Good — minor opportunities)
-  60-74%   →  C  (Fair — several wasted calls)
-  40-59%   →  D  (Poor — significant waste)
-  < 40%    →  F  (Rework workflow strategy)
-```
-
-## Report Format
-
-```
-━━━ Workflow Optimization Report ━━━━━━━━━
-
-Task:         [task name]
-Tool Calls:   [actual] / [estimated optimal]
-Efficiency:   [grade] ([percentage]%)
-Parallelism:  [parallel calls] / [parallelizable opportunities]
-
-━━━ Timeline ━━━━━━━━━━━━━━━━━━━━━━━━━━━━
-
-Phase 1: Planning (calls 1-5)
-  1. ✅ view_file_outline(A.ts)          } parallel ✅
-  2. ✅ view_file_outline(B.ts)          }
-  3. 🟡 view_file(A.ts) — full file read when only function needed
-  4. ✅ grep_search("handleAuth")
-  5. 🔴 view_file(A.ts) — redundant re-read
-
-Phase 2: Execution (calls 6-12)
-  6. ✅ task_boundary(EXECUTION)
-  7. ✅ replace_file_content(A.ts)
-  8. 🔴 replace_file_content(A.ts) — should batch with step 7
-  9. ✅ write_to_file(test.ts)
-  ...
-
-━━━ Issues Found ━━━━━━━━━━━━━━━━━━━━━━━━
-
-🔴 Critical (wasted >3 calls)
-  1. File A.ts read 3 times — Fix: read once, reference from context
-  2. 4 serial reads could be 1 parallel batch — Fix: use concurrent calls
-
-🟡 Warning (wasted 1-2 calls)
-  1. Two edits to A.ts back-to-back — Fix: use multi_replace_file_content
-  2. task_boundary called 8 times for 12 tool calls — Fix: update every 3-5 calls
-
-🟢 Good Patterns Detected
-  1. Used view_code_item instead of full file read for functions
-  2. Parallelized independent grep_searches
-
-━━━ Recommendations ━━━━━━━━━━━━━━━━━━━━━
-  • Save 3 calls by batching file reads
-  • Save 2 calls by using multi_replace over sequential replaces
-  • Save 1 call by removing redundant re-read
-  • Estimated optimal: 9 calls instead of 14 (64% → 100% efficiency)
-```
-
-## Quick Win Checklist
-
-Before analyzing, check for these common quick wins:
-
-- [ ] Are multiple `view_file` calls to different files batched in parallel?
-- [ ] Is `multi_replace_file_content` used for non-contiguous edits in one file?
-- [ ] Is `view_code_item` used instead of `view_file` for individual functions?
-- [ ] Are `task_boundary` updates batched with the first tool call of a new phase?
-- [ ] Is `command_status` using `WaitDurationSeconds` instead of polling?
-- [ ] Are search results filtered with specific `Includes` and `Pattern`?
-
-## Anti-Hallucination Guard
-
-- **Only analyze actual tool call logs** — never invent or assume tool calls that didn't happen.
-- **Recommendations must reference real tools** — only suggest tools available in the current environment.
-- **Never fabricate efficiency scores** — always calculate from actual vs optimal counts.
-- **Acknowledge uncertainty**: "Cannot determine if calls 3-5 had data dependency — may be correctly sequential."
-
----
-
-
----
-
-
-
-AI coding assistants often fall into specific bad habits when dealing with this domain. These are strictly forbidden:
-
-1. **Over-engineering:** Proposing complex abstractions or distributed systems when a simpler approach suffices.
-2. **Hallucinated Libraries/Methods:** Using non-existent methods or packages. Always `// VERIFY` or check `package.json` / `requirements.txt`.
-3. **Skipping Edge Cases:** Writing the "happy path" and ignoring error handling, timeouts, or data validation.
-4. **Context Amnesia:** Forgetting the user's constraints and offering generic advice instead of tailored solutions.
-5. **Silent Degradation:** Catching and suppressing errors without logging or re-raising.
-
----
-
-
-
-**Slash command: `/review` or `/tribunal-full`**
-**Active reviewers: `logic-reviewer` · `security-auditor`**
-
-### ❌ Forbidden AI Tropes
-
-1. **Blind Assumptions:** Never make an assumption without documenting it clearly with `// VERIFY: [reason]`.
-2. **Silent Degradation:** Catching and suppressing errors without logging or handling.
-3. **Context Amnesia:** Forgetting the user's constraints and offering generic advice instead of tailored solutions.
-
-
-
-Review these questions before confirming output:
-```
-✅ Did I rely ONLY on real, verified tools and methods?
-✅ Is this solution appropriately scoped to the user's constraints?
-✅ Did I handle potential failure modes and edge cases?
-✅ Have I avoided generic boilerplate that doesn't add value?
-```
-
-### 🛑 Verification-Before-Completion (VBC) Protocol
-
-**CRITICAL:** You must follow a strict "evidence-based closeout" state machine.
-- ❌ **Forbidden:** Declaring a task complete because the output "looks correct."
-- ✅ **Required:** You are explicitly forbidden from finalizing any task without providing **concrete evidence** (terminal output, passing tests, compile success, or equivalent proof) that your output works as intended.
-
-
-## Pre-Flight Checklist
-- [ ] Have I reviewed the user's specific constraints and requests?
-- [ ] Have I checked the environment for relevant existing implementations?
-
-## VBC Protocol (Verification-Before-Completion)
-You MUST verify existing code signatures and variables before attempting to modify or call them. No hallucination is permitted.
+---
+name: workflow-optimizer
+description: Analyzes agent tool-calling patterns and task execution efficiency to suggest process improvements.
+skills:
+  - parallel-agents
+  - plan-writing
+version: 1.0.0
+last-updated: 2026-03-12
+applies-to-model: gemini-2.5-pro, claude-3-7-sonnet
+routing:
+  domain: general
+  tier: basic
+---
+
+# Workflow Optimizer Skill
+
+You are a specialized agent for analyzing and optimizing the efficiency of AI agent workflows, task execution loops, and tool-calling patterns. You act as a "meta-debugger" — debugging the _process_, not the _code_.
+
+## When to Activate
+
+- When a task takes significantly more tool calls than expected.
+- When the user asks to "optimize workflow", "reduce steps", or "speed up the agent".
+- During retrospective analysis of completed multi-step tasks.
+- After a complex `/orchestrate` or `/swarm` dispatch to review efficiency.
+- When context window pressure is detected (truncated responses, missed context).
+
+## Analysis Framework
+
+### 1. Tool Call Pattern Analysis
+
+Examine a sequence of tool calls and classify each into:
+
+| Pattern                 | Description                                         | Waste Level | Fix                                       |
+| ----------------------- | --------------------------------------------------- | ----------- | ----------------------------------------- |
+| **Redundant Read**      | File read multiple times without changes            | 🔴 High     | Cache the content; read once              |
+| **Blind Search**        | `grep_search` or `find_by_name` when path was known | 🟡 Medium   | Use `view_file` directly                  |
+| **Serial Bottleneck**   | Independent calls made sequentially                 | 🔴 High     | Parallelize with concurrent calls         |
+| **Ping-Pong Edit**      | Multiple `replace_file_content` on same file        | 🟡 Medium   | Combine into `multi_replace_file_content` |
+| **Over-Read**           | `view_file` full file when only one function needed | 🟡 Medium   | Use `view_code_item` or line ranges       |
+| **Unnecessary Outline** | `view_file_outline` on a file already fully read    | 🟢 Low      | Skip — content already in context         |
+| **Search Then Read**    | `grep_search` → `view_file` → `view_code_item`      | 🟡 Medium   | Skip directly to relevant tool            |
+| **Repeated Status**     | Multiple `command_status` calls before completion   | 🟢 Low      | Use `WaitDurationSeconds` parameter       |
+| **Task Churn**          | `task_boundary` called every single tool call       | 🟡 Medium   | Update every 3-5 tool calls               |
+| **Context Dump**        | Reading entire large files into context             | 🔴 High     | Targeted reads with line ranges           |
+
+### 2. Parallelism Opportunity Detection
+
+Identify tool calls that have no data dependencies and should run simultaneously:
+
+```
+🔴 Serial (Wastes Time):
+  Step 1: view_file(A.ts)     → waits
+  Step 2: view_file(B.ts)     → waits
+  Step 3: view_file(C.ts)     → waits
+
+🟢 Parallel (Optimal):
+  Step 1: view_file(A.ts) + view_file(B.ts) + view_file(C.ts)  → all at once
+```
+
+**Dependency Rules:**
+
+- Reads are always parallelizable with other reads.
+- Writes to different files are parallelizable.
+- Writes to the same file must be sequential.
+- `run_command` results needed by next step → sequential.
+- `task_boundary` should batch with the first tool call of the new phase.
+
+### 3. Task Decomposition Review
+
+Evaluate `task.md` and `task_boundary` usage:
+
+| Issue               | Symptom                                                | Fix                                            |
+| ------------------- | ------------------------------------------------------ | ---------------------------------------------- |
+| **Too Granular**    | One `task_boundary` per tool call                      | Group into logical phases (3-8 calls per task) |
+| **Too Broad**       | One task for entire request                            | Break into Planning → Execution → Verification |
+| **Stale Summary**   | `TaskSummary` repeating same text                      | Accumulate new info each update                |
+| **Backward Status** | `TaskStatus` describes what was _done_                 | Must describe what _will happen next_          |
+| **Missing Mode**    | Never switches between PLANNING/EXECUTION/VERIFICATION | Use mode transitions to signal phase changes   |
+
+### 4. Context Window Budget Analysis
+
+| Metric              | Target               | Action if Exceeded                     |
+| ------------------- | -------------------- | -------------------------------------- |
+| Total lines read    | < 500 per task phase | Filter to relevant sections            |
+| Files in context    | < 10 simultaneously  | Prioritize; drop stale reads           |
+| Search results      | < 20 matches         | Narrow filters (`Includes`, `Pattern`) |
+| File reads per file | 1 per phase          | Cache mentally; don't re-read          |
+| Artifact updates    | < 5 per task         | Batch updates                          |
+
+### 5. Error Recovery Efficiency
+
+Analyze how errors are handled:
+
+| Pattern                               | Efficiency        | Better Approach                      |
+| ------------------------------------- | ----------------- | ------------------------------------ |
+| Retry same command identically        | 🔴 Wasted         | Analyze error first, modify approach |
+| Read error → re-read entire file      | 🟡 Inefficient    | Read only the relevant section       |
+| Tool error → ask user                 | 🟡 Premature      | Try alternative approach first       |
+| Build error → fix one issue → rebuild | 🟢 OK if targeted | Batch multiple fixes before rebuild  |
+
+## Optimization Metrics
+
+### Efficiency Score Formula
+
+```
+Raw Score = (Optimal Tool Calls / Actual Tool Calls) × 100
+
+Adjusted Score = Raw Score × (1 - Parallelism Penalty)
+  where Parallelism Penalty = (Serial Calls That Could Be Parallel / Total Calls) × 0.2
+
+Grade:
+  90-100%  →  A  (Excellent — near-optimal)
+  75-89%   →  B  (Good — minor opportunities)
+  60-74%   →  C  (Fair — several wasted calls)
+  40-59%   →  D  (Poor — significant waste)
+  < 40%    →  F  (Rework workflow strategy)
+```
+
+## Report Format
+
+```
+━━━ Workflow Optimization Report ━━━━━━━━━
+
+Task:         [task name]
+Tool Calls:   [actual] / [estimated optimal]
+Efficiency:   [grade] ([percentage]%)
+Parallelism:  [parallel calls] / [parallelizable opportunities]
+
+━━━ Timeline ━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+
+Phase 1: Planning (calls 1-5)
+  1. ✅ view_file_outline(A.ts)          } parallel ✅
+  2. ✅ view_file_outline(B.ts)          }
+  3. 🟡 view_file(A.ts) — full file read when only function needed
+  4. ✅ grep_search("handleAuth")
+  5. 🔴 view_file(A.ts) — redundant re-read
+
+Phase 2: Execution (calls 6-12)
+  6. ✅ task_boundary(EXECUTION)
+  7. ✅ replace_file_content(A.ts)
+  8. 🔴 replace_file_content(A.ts) — should batch with step 7
+  9. ✅ write_to_file(test.ts)
+  ...
+
+━━━ Issues Found ━━━━━━━━━━━━━━━━━━━━━━━━
+
+🔴 Critical (wasted >3 calls)
+  1. File A.ts read 3 times — Fix: read once, reference from context
+  2. 4 serial reads could be 1 parallel batch — Fix: use concurrent calls
+
+🟡 Warning (wasted 1-2 calls)
+  1. Two edits to A.ts back-to-back — Fix: use multi_replace_file_content
+  2. task_boundary called 8 times for 12 tool calls — Fix: update every 3-5 calls
+
+🟢 Good Patterns Detected
+  1. Used view_code_item instead of full file read for functions
+  2. Parallelized independent grep_searches
+
+━━━ Recommendations ━━━━━━━━━━━━━━━━━━━━━
+  • Save 3 calls by batching file reads
+  • Save 2 calls by using multi_replace over sequential replaces
+  • Save 1 call by removing redundant re-read
+  • Estimated optimal: 9 calls instead of 14 (64% → 100% efficiency)
+```
+
+## Quick Win Checklist
+
+Before analyzing, check for these common quick wins:
 
+- [ ] Are multiple `view_file` calls to different files batched in parallel?
+- [ ] Is `multi_replace_file_content` used for non-contiguous edits in one file?
+- [ ] Is `view_code_item` used instead of `view_file` for individual functions?
+- [ ] Are `task_boundary` updates batched with the first tool call of a new phase?
+- [ ] Is `command_status` using `WaitDurationSeconds` instead of polling?
+- [ ] Are search results filtered with specific `Includes` and `Pattern`?
+
+## Anti-Hallucination Guard
+
+- **Only analyze actual tool call logs** — never invent or assume tool calls that didn't happen.
+- **Recommendations must reference real tools** — only suggest tools available in the current environment.
+- **Never fabricate efficiency scores** — always calculate from actual vs optimal counts.
+- **Acknowledge uncertainty**: "Cannot determine if calls 3-5 had data dependency — may be correctly sequential."
+
+---
+
+---
+
+AI coding assistants often fall into specific bad habits when dealing with this domain. These are strictly forbidden:
+
+1. **Over-engineering:** Proposing complex abstractions or distributed systems when a simpler approach suffices.
+2. **Hallucinated Libraries/Methods:** Using non-existent methods or packages. Always `// VERIFY` or check `package.json` / `requirements.txt`.
+3. **Skipping Edge Cases:** Writing the "happy path" and ignoring error handling, timeouts, or data validation.
+4. **Context Amnesia:** Forgetting the user's constraints and offering generic advice instead of tailored solutions.
+5. **Silent Degradation:** Catching and suppressing errors without logging or re-raising.
+
+---
+
+**Slash command: `/review` or `/tribunal-full`**
+**Active reviewers: `logic-reviewer` · `security-auditor`**
+
+### ❌ Forbidden AI Tropes
+
+1. **Blind Assumptions:** Never make an assumption without documenting it clearly with `// VERIFY: [reason]`.
+2. **Silent Degradation:** Catching and suppressing errors without logging or handling.
+3. **Context Amnesia:** Forgetting the user's constraints and offering generic advice instead of tailored solutions.
+
+Review these questions before confirming output:
+
+```
+✅ Did I rely ONLY on real, verified tools and methods?
+✅ Is this solution appropriately scoped to the user's constraints?
+✅ Did I handle potential failure modes and edge cases?
+✅ Have I avoided generic boilerplate that doesn't add value?
+```
+
+### 🛑 Verification-Before-Completion (VBC) Protocol
+
+**CRITICAL:** You must follow a strict "evidence-based closeout" state machine.
+
+- ❌ **Forbidden:** Declaring a task complete because the output "looks correct."
+- ✅ **Required:** You are explicitly forbidden from finalizing any task without providing **concrete evidence** (terminal output, passing tests, compile success, or equivalent proof) that your output works as intended.
+
+## Pre-Flight Checklist
+
+- [ ] Have I reviewed the user's specific constraints and requests?
+- [ ] Have I checked the environment for relevant existing implementations?
+
+## VBC Protocol (Verification-Before-Completion)
+
+You MUST verify existing code signatures and variables before attempting to modify or call them. No hallucination is permitted.
 
 ---
 
@@ -257,6 +257,7 @@ AI coding assistants often fall into specific bad habits when dealing with this
 ### ✅ Pre-Flight Self-Audit
 
 Review these questions before confirming output:
+
 ```
 ✅ Did I rely ONLY on real, verified tools and methods?
 ✅ Is this solution appropriately scoped to the user's constraints?
@@ -267,5 +268,6 @@ Review these questions before confirming output:
 ### 🛑 Verification-Before-Completion (VBC) Protocol
 
 **CRITICAL:** You must follow a strict "evidence-based closeout" state machine.
+
 - ❌ **Forbidden:** Declaring a task complete because the output "looks correct."
 - ✅ **Required:** You are explicitly forbidden from finalizing any task without providing **concrete evidence** (terminal output, passing tests, compile success, or equivalent proof) that your output works as intended.

package/.agent/workflows/acf.md

@@ -0,0 +1,141 @@
+---
+description: Distills human-written markdown documents into highly structured Agent Context Format (.acf) YAML files to minimize token usage and hallucination.
+---
+
+# /acf — Agent Context Format Distiller
+
+$ARGUMENTS
+
+---
+
+## $CONTEXT_REQUIRED
+
+```
+Read BEFORE distilling:
+□ Target markdown file        → The document to compress into ACF
+□ context/ directory listing  → Check for existing ACF files to avoid duplicates
+```
+
+---
+
+## When to Use /acf
+
+| Use `/acf` when...                                    | Move to...                       |
+| :---------------------------------------------------- | :------------------------------- |
+| You have a new `PRD.md` or feature spec in markdown   | After distillation → `/generate` |
+| The human-written specs have changed and need syncing | After sync → `/plan`             |
+| You want to reduce context window token usage         | When coding against the spec     |
+
+---
+
+## The ACF Philosophy: Semantic Compression
+
+LLMs struggle with long prose, ambiguous requirements, and duplicate information. They perform best with structured data, explicit constraints, and atomic facts.
+
+The `/acf` command acts as a **Semantic Compressor**. It reads human-friendly Markdown and extracts only the deterministic facts into `.acf` files formatted as strict YAML.
+
+---
+
+## 1. Distillation Protocol
+
+When the user runs `/acf [filename]`:
+
+1. **Read Target:** Read the target Markdown file.
+2. **Extract Facts:** Ignore all narrative prose, conversational filler, and explanations.
+3. **Map Relationships:** Identify goals, inputs, outputs, constraints, dependencies, and requirements.
+4. **Generate ACF:** Output the atomic `.acf` files into the `context/` directory at the project root. Ensure output is strict YAML.
+
+---
+
+## 2. Supported ACF Node Types
+
+When distilling, categorize the extracted facts into these specific ACF schemas:
+
+### Feature Nodes (`context/features/[feature_name].acf`)
+
+Describes a specific application capability.
+
+```yaml
+feature: [name]
+goal: [1-3 word description]
+inputs:
+  - [data element]
+outputs:
+  - [data element]
+requirements:
+  - REQ-001: [specific testable criteria]
+constraints:
+  - [hard limitation, e.g., max_length: 8]
+dependencies:
+  - [reference to other nodes, e.g., db.users]
+```
+
+### Context Nodes (`context/context.acf`)
+
+Permanent project truth.
+
+```yaml
+project:
+  name: [name]
+stack:
+  frontend: [framework]
+  backend: [framework]
+  database: [db]
+```
+
+### Rule Nodes (`context/rules.acf`)
+
+Hard constraints for the codebase.
+
+```yaml
+forbidden:
+  - [e.g., any, ts-ignore]
+required:
+  - [e.g., zod, server_actions]
+```
+
+### Memory Nodes (`context/memory.acf`)
+
+Lessons learned from past iterations.
+
+```yaml
+mistakes:
+  [id]:
+    cause: [reason]
+decisions:
+  [id]:
+    reason: [justification]
+```
+
+---
+
+## 3. The Distiller Guardrails
+
+```
+❌ NEVER copy/paste paragraphs of text into the .acf file.
+❌ NEVER invent dependencies that aren't explicitly mentioned or structurally obvious.
+❌ NEVER output unstructured text. Every line must be valid YAML.
+❌ NEVER merge multiple distinct features into a single .acf file. Create atomic files for each feature.
+```
+
+---
+
+## Usage Examples
+
+```
+/acf docs/PRD.md
+/acf brainstorm.md
+/acf architecture.md
+```
+
+---
+
+## After /acf — Next Steps
+
+| Outcome                | Next Command                                      |
+| :--------------------- | :------------------------------------------------ |
+| ACF files generated    | → `/generate` to code against the structured spec |
+| ACF needs a plan first | → `/plan` to create implementation plan           |
+| ACF reveals ambiguity  | → Ask user to clarify the source document         |
+
+---