# Context Pack

> Prepare files, folders, code, logs, and structured data as a bounded, traceable context pack for Codex. Use when source material must be converted, deduplicated, selected, budgeted, or handed off with exact evidence anchors.

- Skill: `tikazi/context-pack` (Agent Skill, multi-file: 2 files)
- Install (CLI): `npx skillmds@latest add tikazi/context-pack`
- Raw SKILL.md: https://api.skillmd.com/api/skills/tikazi/context-pack/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- Author: TIKAZI (https://skillmd.com/u/tikazi)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/tikazi/context-pack

---


# Context Pack

Designed, integrated, independently refactored, and continuously maintained by **TIKAZ**.

## Inputs and routing

Accept one or more files, folders, code trees, logs, structured data, or converter-produced Markdown plus a concrete task. Use `text` for confident text-first material, `hybrid` for bounded task-relevant visuals or complex tables, and `source` when extraction cannot preserve important evidence.

## Workflow

Own canonical ingestion, fidelity profiling, exact deduplication, evidence selection, and final pack size. First run `profile` or let `pack` profile automatically:

- `text`: canonical Markdown is sufficient;
- `hybrid`: use Markdown for text and a bounded visual-evidence queue for informative images or complex tables;
- `source`: keep the original asset/page path when safe extraction cannot preserve task-relevant information.

Do not trigger vision for a logo, repeated icon, background, or every image merely because it exists. When the queue contains `pending-vision` items and the host can inspect images, resolve the referenced item, record an anchored observation plus uncertainty, and keep the original reference. When the capability is unavailable, leave it pending or recommend the source file; never invent a description.

Profile first, protect literal facts and anchors, deduplicate only exact or formatting-only repetition, select task-relevant evidence, and assemble one task-ready artifact in this order:

1. task and expected output;
2. selected mode and estimated budget;
3. confirmed constraints and protected facts;
4. exact evidence excerpts with source anchors;
5. decisions, completed work, and current state;
6. conflicts and open questions;
7. omitted-anchor inventory and verification limits.

The pack must distinguish exact source text, structured state, and inference. It must remain useful without the surrounding chat. Count the complete artifact against the budget. If essential protected evidence cannot fit, return a visible budget conflict instead of silently exceeding the limit.

## Output contract

Return `<directory>/packs/current-task.context.md` with `profile.json`, `visual-evidence.json`, `context-cost-ledger.json`, canonical files, indexes, and `savings-report.md`. Distinguish exact excerpts, structured state, inference, omissions, and pending visual evidence.

## Validation and fallback

Count the complete artifact against the budget and verify protected facts plus selected anchors. If a converter or vision host is unavailable, leave evidence pending or preserve the source reference. Never invent a visual description. If essential evidence cannot fit, report a visible budget conflict.

## Example

```text
Use context-pack on these release notes and logs. Build an 800-token pack for regression review, keep commands and versions exact, and list omitted anchors.
```

Run `python scripts/tikaz_context.py pack --input <path> --query <task> --budget <tokens> --visual-budget <items> --output <directory>`.