shikhu-inquiry
Answer a question about how or why something in THIS repository works, by showing which tracked files answer it and why, then recording the exchange so it can seed a quiz later. Use it whenever the user asks a genuine question whose subject belongs to this repository — its source
Install
npx skills add https://github.com/arjunpatel7/shikhu/tree/main/src/shikhu/_skill/shikhu-inquiry
claude plugin marketplace add https://llmmart.ai/marketplace.json && claude plugin install arjunpatel7-shikhu@llmmart
git clone https://github.com/arjunpatel7/shikhu.git
The skills CLI installs just this skill, for any of its supported agents. Claude Code installs the whole arjunpatel7/shikhu collection as a plugin from our marketplace. Git is the plain clone.
Skill manifest
shikhu-inquiry — answer from the codebase, then bank the question
Why this exists
Shikhu tracks how well the user understands their own code. Historically that meant generating quizzes and asking them to sit through them, which people don't do. But people do ask their agent to explain their code, constantly. This skill turns that existing habit into the same signal: the question they already asked becomes the seed of a quiz question, and the files you cite become the record of what it was about.
Two things follow from that, and they are the whole point:
- Show your work. Don't just answer — say which files answer it and why those files. The user should end the turn knowing where the answer lives, not just what it is.
- Ask once, cheaply. One short confirm closes the loop. Silence earns nothing.
0. Invoked explicitly?
If the user ran /shikhu-inquiry directly, skip any judgment about whether this is a
conceptual question — they've already made that call. Use their message as the question, or
ask them for one if they gave none, then continue from step 1.
1. Get the packet
shikhu inquiry-packet "<the user's question>"
If shikhu is not on PATH or is an older install without this command, invoke the module
directly instead: python -m shikhu inquiry-packet "<question>", using the interpreter for
this project.
Three possible outcomes:
| output | what it means | what you do |
|---|---|---|
START HERE + files |
files were found | continue to step 2 |
not answerable from this codebase |
the question is about a library, a tool, or general knowledge | answer normally, record nothing, skip the confirm |
No cached file summaries |
the repo hasn't been summarized | say one line: "shikhu has no summaries for this repo yet — shikhu summarize would let me point at files." then answer normally |
Never invent a packet. If the command fails for any reason, just answer the question normally and say nothing about shikhu.
2. Answer, leading with the files
Structure the answer so the files come first and carry their justification:
Two files answer this.
src/shikhu/store.pyis where it actually lives — it's the SQLite persistence layer, and thequestionstable is defined there.src/shikhu/commands/generate_from_study.pymatters because it's what writes new questions in the first place.[then the actual explanation, citing
path:line]
Rules:
- Lead with the strong file(s). The packet's scores are calibrated; a 0.93 is a confident hit and a 0.35 is a maybe. Present them with that asymmetry, don't flatten them into a list.
- Read the files before explaining them. The packet is built from cached summaries, not from current source. Confirm the code still says what you're about to claim.
- Cite
path:lineso the user can jump there. - Don't show raw scores unless the user asks. "This is the main one, and this one's related" carries the same information without the noise.
- If the packet's top file looks wrong to you after reading it, say so and name the file you think is actually right. You have the source; the packet only had summaries.
3. Ask the confirm — one line, at the end
After answering, ask exactly one short question:
"Was that the right place to be looking?"
Keep it to one line. Do not offer a menu, do not explain the credit system, do not ask it in the middle of the answer.
4. Record their reply
When they answer, record it and say nothing further about it:
# they said yes / "yeah" / "that's it" / moved on satisfied
shikhu record-inquiry "<top file from the packet>" "<their original question>" --confirmed --score <top score>
# they said no / named a different file / said it was the wrong area
shikhu record-inquiry "<top file from the packet>" "<their original question>" --rejected --score <top score>
Pass --runner-up and --runner-up-score when the packet listed a second file.
Pass --also "<path>" once for each other START HERE file in the packet. A question that
spans several files should be recorded with all of them, so the quiz it seeds can span them
too and credit each file; skip it for RELATED files, which only sit near the answer.
A rejection is valuable, not a failure. It records that the question was real but the file was wrong, which is exactly the correction that improves ranking. Record it with the same lack of ceremony as a confirmation.
If the user ignores the confirm and moves on to something else, record nothing and drop it. Don't re-ask, and don't record a confirmation they didn't give.
5. Don't oversell what was earned
Confirming validates the question, not the user's knowledge — they just asked it, which is mild evidence they didn't know. Knowledge credit comes only from answering it later in a quiz. So never say "that's covered now" or "coverage went up."
At most, one line if it feels natural:
"Banked —
shikhu generate-from-study src/shikhu/store.pywill turn that into a quiz question."
Usually say nothing at all. The recording is bookkeeping, not an achievement.
When not to use this
- The user asked you to do something (write, fix, run, refactor). That's not an inquiry.
- The question is about a library, a tool, the agent itself, or general programming. The gate in step 1 catches most of these; trust it when it says not answerable.
- The user is mid-task and a detour would break their flow. Answer, skip the confirm.
Files (shikhu)
-
SKILL.md 6.6 KB
--- name: shikhu-inquiry description: >- Answer a question about how or why something in THIS repository works, by showing which tracked files answer it and why, then recording the exchange so it can seed a quiz later. Use it whenever the user asks a genuine question whose subject belongs to this repository — its source, CLI behavior, config, CI, hooks or build setup — and answering it well means reading files here. These arrive casually and mid-task, often elliptical or typo-laden: "how are we storing quizzes rn?", "what does that for loop do?", "what's the pre-push hook?", "why do I gotta do that venv thing?", "explain how verifying/decoding works". Use it for those. Nothing else detects these, so a question you skip is lost. Do NOT use it for: instructions to do work; short remarks, fragments or acknowledgements; session or tool boilerplate; requests to restate or clarify something already said; questions about the assistant itself; questions about external products, services or libraries; or product, pricing and strategy questions. The user can also invoke it explicitly as /shikhu-inquiry. allowed-tools: - Read - Grep - Glob - Bash(shikhu *) - Bash(*python -m shikhu *) --- # shikhu-inquiry — answer from the codebase, then bank the question ## Why this exists Shikhu tracks how well the user understands *their own* code. Historically that meant generating quizzes and asking them to sit through them, which people don't do. But people *do* ask their agent to explain their code, constantly. This skill turns that existing habit into the same signal: the question they already asked becomes the seed of a quiz question, and the files you cite become the record of what it was about. Two things follow from that, and they are the whole point: - **Show your work.** Don't just answer — say which files answer it and why those files. The user should end the turn knowing where the answer lives, not just what it is. - **Ask once, cheaply.** One short confirm closes the loop. Silence earns nothing. ## 0. Invoked explicitly? If the user ran `/shikhu-inquiry` directly, skip any judgment about whether this is a conceptual question — they've already made that call. Use their message as the question, or ask them for one if they gave none, then continue from step 1. ## 1. Get the packet ```bash shikhu inquiry-packet "<the user's question>" ``` If `shikhu` is not on PATH or is an older install without this command, invoke the module directly instead: `python -m shikhu inquiry-packet "<question>"`, using the interpreter for this project. Three possible outcomes: | output | what it means | what you do | |---|---|---| | `START HERE` + files | files were found | continue to step 2 | | `not answerable from this codebase` | the question is about a library, a tool, or general knowledge | answer normally, **record nothing**, skip the confirm | | `No cached file summaries` | the repo hasn't been summarized | say one line: *"shikhu has no summaries for this repo yet — `shikhu summarize` would let me point at files."* then answer normally | Never invent a packet. If the command fails for any reason, just answer the question normally and say nothing about shikhu. ## 2. Answer, leading with the files Structure the answer so the files come first and carry their justification: > Two files answer this. **`src/shikhu/store.py`** is where it actually lives — it's the > SQLite persistence layer, and the `questions` table is defined there. **`src/shikhu/commands/generate_from_study.py`** > matters because it's what writes new questions in the first place. > > [then the actual explanation, citing `path:line`] Rules: - **Lead with the strong file(s).** The packet's scores are calibrated; a 0.93 is a confident hit and a 0.35 is a maybe. Present them with that asymmetry, don't flatten them into a list. - **Read the files before explaining them.** The packet is built from cached summaries, not from current source. Confirm the code still says what you're about to claim. - **Cite `path:line`** so the user can jump there. - **Don't show raw scores** unless the user asks. "This is the main one, and this one's related" carries the same information without the noise. - If the packet's top file looks wrong to you after reading it, say so and name the file you think is actually right. You have the source; the packet only had summaries. ## 3. Ask the confirm — one line, at the end After answering, ask exactly one short question: > *"Was that the right place to be looking?"* Keep it to one line. Do not offer a menu, do not explain the credit system, do not ask it in the middle of the answer. ## 4. Record their reply When they answer, record it and say nothing further about it: ```bash # they said yes / "yeah" / "that's it" / moved on satisfied shikhu record-inquiry "<top file from the packet>" "<their original question>" --confirmed --score <top score> # they said no / named a different file / said it was the wrong area shikhu record-inquiry "<top file from the packet>" "<their original question>" --rejected --score <top score> ``` Pass `--runner-up` and `--runner-up-score` when the packet listed a second file. Pass `--also "<path>"` once for each *other* START HERE file in the packet. A question that spans several files should be recorded with all of them, so the quiz it seeds can span them too and credit each file; skip it for RELATED files, which only sit near the answer. **A rejection is valuable, not a failure.** It records that the question was real but the file was wrong, which is exactly the correction that improves ranking. Record it with the same lack of ceremony as a confirmation. If the user ignores the confirm and moves on to something else, **record nothing** and drop it. Don't re-ask, and don't record a confirmation they didn't give. ## 5. Don't oversell what was earned Confirming validates the **question**, not the user's knowledge — they just asked it, which is mild evidence they *didn't* know. Knowledge credit comes only from answering it later in a quiz. So never say "that's covered now" or "coverage went up." At most, one line if it feels natural: > *"Banked — `shikhu generate-from-study src/shikhu/store.py` will turn that into a quiz question."* Usually say nothing at all. The recording is bookkeeping, not an achievement. ## When not to use this - The user asked you to *do* something (write, fix, run, refactor). That's not an inquiry. - The question is about a library, a tool, the agent itself, or general programming. The gate in step 1 catches most of these; trust it when it says not answerable. - The user is mid-task and a detour would break their flow. Answer, skip the confirm.
Comments (0)
Sign in to join the conversation.
Reviews (0)
No reviews yet.
No comments yet.