Context Pack Architect

---
name: context-pack-architect
description: Build or repair a reusable context pack for recurring AI-assisted work. Use when outputs are generic, inconsistent, stale, or repeatedly require the same corrections and source material.
license: All-Rights-Reserved
metadata:
  author: Zain Haseeb
  version: 1.0.0
  category: ai-workflows
  tags: [context, sources, quality, reusable-work]
---

# Context Pack Architect

Turn scattered instructions, examples, source material, and judgment into a compact context pack that can support repeated work without silently inventing missing facts.

## Critical behavior

1. Separate durable context from facts that change on every assignment.
2. Preserve the source hierarchy and name conflicts instead of quietly choosing a winner.
3. Mark every important statement as supplied, derived, inferred, or unknown.
4. Keep decisions requiring human authority with the human.
5. Add freshness dates to facts that can drift.
6. Test the pack on one real assignment before calling it reusable.

## When to use

Use this skill when a user asks to build a context pack, improve repeated AI output, document how good work gets made, transfer team knowledge, or diagnose why an agent keeps missing the same nuance.

Do not use it for a one-off prompt that has no repeated context, or to bypass privacy and rights restrictions on source material.

## Build the pack

### 1. Define the recurring job

Capture:

- job to be done;
- person who will use the result;
- visible artifact produced;
- frequency and expected variation;
- decisions the human retains;
- consequence of a wrong result.

If any missing answer changes the design materially, ask one bundled question. Otherwise state a conservative assumption and continue.

### 2. Inventory the context

Classify each input as one of:

- governing source;
- domain authority;
- current fact;
- quality standard;
- positive example;
- failure example;
- operating boundary;
- historical context;
- task-specific input.

Record owner, date, scope, and rights boundary for every governing or private source. If two sources conflict, name the conflict and stop before resolving it without authority.

### 3. Extract invisible judgment

Ask for one strong example, one acceptable example, and one polished-looking failure. For each, identify:

- cues the expert noticed;
- tradeoff accepted;
- shortcut rejected;
- threshold that changed the decision;
- evidence that would reverse the judgment.

Do not flatten taste into a score alone. Preserve paired examples and exception cases.

### 4. Assemble the context pack

Use this exact structure:

1. Purpose and audience
2. Definitions and boundaries
3. Source hierarchy
4. Accepted facts with freshness dates
5. Quality bar
6. Positive and negative examples
7. Decision and approval boundaries
8. Unknowns and conflicts
9. Task intake block
10. Verification checklist
11. Maintenance triggers

Keep durable material in the pack. Put changing details in the task intake block.

### 5. Test and repair

Run one representative assignment with the pack. Compare the result with the same assignment using the old setup or no pack. Check factual accuracy, specificity, source use, audience fit, human correction required, and whether the agent respected boundaries.

Repair the smallest context gap that caused each failure. Do not add more material merely because the test was imperfect. Stop when the pack clears the named quality bar twice on representative inputs.

## Output format

Return:

- a Markdown context pack;
- a short task intake template;
- a source register;
- test results and remaining gaps;
- the next review date or change event.

## Example

Input: "Build a context pack for our monthly executive update. The AI keeps making it sound like a marketing newsletter."

Output: a source-ranked pack separating durable executive preferences from monthly metrics, paired examples showing decision-ready versus promotional writing, an approval boundary for forward-looking claims, and a test using last month's update.

## Troubleshooting

- If the pack grows into a document dump, retain only material that changes the output or constrains a decision.
- If examples disagree with stated rules, treat the conflict as evidence and ask which source governs.
- If the result remains generic, inspect missing audience consequence, positive examples, and expert thresholds before changing tools.
- If facts drift quickly, move them out of durable context and require them in each task intake.

## What not to do

- Do not copy private or licensed material into a shareable pack without permission.
- Do not present inferred judgment as a rule the expert approved.
- Do not hide contradictions to make the pack look complete.
- Do not treat a long pack as a good pack.
- Do not let the pack authorize external, destructive, or consequential actions.