fastcontext
Locate the right files and line ranges before answering, editing, reviewing, or debugging unfamiliar code — explore first, then act on a compact set of citations instead of repeating broad searches.
Install
npx skills add https://github.com/intel/dokimi/tree/main/dokimi/data/skills/fastcontext
claude plugin marketplace add https://llmmart.ai/marketplace.json && claude plugin install intel-dokimi@llmmart
git clone https://github.com/intel/dokimi.git
The skills CLI installs just this skill, for any of its supported agents. Claude Code installs the whole intel/dokimi collection as a plugin from our marketplace. Git is the plain clone.
Skill manifest
FastContext: explore before you act
Use this skill whenever a task touches code or docs you have not already read in this session and the answer depends on more than one file or on cross-module logic. Find the relevant locations first, commit to them as a compact list of citations, then read narrow windows around those citations instead of scanning the whole repository.
This skill is about context acquisition, not about the fix itself. Doing it well makes the rest of the task cheaper and more accurate: fewer wasted reads, tighter edits, less frontier-model token spend.
When to use it
- Before answering a "where / how is X implemented?" question.
- Before editing, refactoring, or debugging code you have not localized yet.
- Before reviewing a change whose blast radius you have not mapped.
When to skip it
- You already have the file and line range in hand.
- The task is a single, known file you have just read.
- A one-shot lookup you can answer without cross-referencing.
How to explore
- Search by intent, not by dumping files: grep for the symbol, the error string, the config key, or the API name — scoped to a directory when you can.
- Run independent searches in parallel rather than one-at-a-time.
- Read only the narrow window each hit points at, not the whole file.
- Stop when you can name the locations that matter. Do not keep searching to feel thorough — converge.
Commit to a citation block
When you have located the evidence, write it as a compact, machine-readable
block of path:start-end citations (one per line):
<final_answer>
src/pkg/module.py:42-58
tests/test_module.py:101-119
docs/guide.md:120-180
</final_answer>
Rules for good citations:
- Cite the tightest range that contains the relevant logic — a function or a few lines, not a whole file. Whole-file citations are not localization.
- Include the supporting evidence the task needs: the implementation and its test or caller, the doc and the example.
- Do not pad the list. Extra, loosely-related files lower precision and make the next reader distrust the set.
After exploring
- Read the narrow windows your citations point at; do not re-run the broad search you already did.
- Trust the located set. Re-doing exploration from scratch wastes tokens and is the dominant failure mode this skill exists to prevent.
Files (dokimi)
-
SKILL.md 2.6 KB
--- name: fastcontext description: Locate the right files and line ranges before answering, editing, reviewing, or debugging unfamiliar code — explore first, then act on a compact set of citations instead of repeating broad searches. --- # FastContext: explore before you act Use this skill whenever a task touches code or docs you have not already read in this session and the answer depends on **more than one file** or on **cross-module logic**. Find the relevant locations *first*, commit to them as a compact list of citations, then read narrow windows around those citations instead of scanning the whole repository. This skill is about *context acquisition*, not about the fix itself. Doing it well makes the rest of the task cheaper and more accurate: fewer wasted reads, tighter edits, less frontier-model token spend. ## When to use it - Before answering a "where / how is X implemented?" question. - Before editing, refactoring, or debugging code you have not localized yet. - Before reviewing a change whose blast radius you have not mapped. ## When to skip it - You already have the file and line range in hand. - The task is a single, known file you have just read. - A one-shot lookup you can answer without cross-referencing. ## How to explore 1. Search by **intent**, not by dumping files: grep for the symbol, the error string, the config key, or the API name — scoped to a directory when you can. 2. Run independent searches in parallel rather than one-at-a-time. 3. Read only the **narrow window** each hit points at, not the whole file. 4. Stop when you can name the locations that matter. Do not keep searching to feel thorough — converge. ## Commit to a citation block When you have located the evidence, write it as a compact, machine-readable block of `path:start-end` citations (one per line): ```text <final_answer> src/pkg/module.py:42-58 tests/test_module.py:101-119 docs/guide.md:120-180 </final_answer> ``` Rules for good citations: - Cite the **tightest range** that contains the relevant logic — a function or a few lines, not a whole file. Whole-file citations are not localization. - Include the supporting evidence the task needs: the implementation **and** its test or caller, the doc **and** the example. - Do not pad the list. Extra, loosely-related files lower precision and make the next reader distrust the set. ## After exploring - Read the narrow windows your citations point at; do **not** re-run the broad search you already did. - Trust the located set. Re-doing exploration from scratch wastes tokens and is the dominant failure mode this skill exists to prevent.
Comments (0)
Sign in to join the conversation.
Reviews (0)
No reviews yet.
No comments yet.