Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIMuseMuseDotsDots+21 more

    Investigate: Root Cause Without Guessing

    1

    Makes your coding agent prove a bug's root cause with file:line evidence and a reproduction before it proposes a fix.

    Free

    1 useSecurity scanned
    Investigate: Root Cause Without Guessing

    Investigate: Root Cause Without Guessing

    Example session with this skill installed

    Search seems broken: I can't find some of my notes even though I typed their title. Can you figure out why?

    • Read your context and instructions
    • Compiled the investigate
    • Generated the document

    Real run of this skill on the example Notes app in poteto-kit, with a bug seeded for the demo. The prompt did not mention it.

    Investigation: search returns nothing when the typed text matches a note title (capitalised titles only)

    • Reproduction: REPRODUCED. Steps: npm ci, then npm run verify -- --feature search --scenario shortcut-filters-and-selects (build c0ab390, repo's verify-app). Result FAIL; failing row: a title match leaves only that note | expected ["Grocery list"] | observed [] | FAIL. Artifacts: artifacts/2026-10-19T04-30-32-957Z/search-shortcut-filters-and-selects/ (report.md, screenshot.png, console.json = []).
    • Hypotheses:
      • H1 Title is compared without case folding: SUPPORTED. If true, any title containing an uppercase letter misses and body matches still work; refuted if lowercase queries on capitalised titles match. Evidence: src/features/search/filter-notes.ts:8 note.title.includes(needle) while the body goes through normalize(...); needle is lowercased at :12 via normalize (:3-5).
      • H2 The palette mangles or truncates the query before it is filtered: REFUTED. search-palette.tsx:59 passes event.target.value to setQuery unchanged, search-store.ts:23-25 stores it as is, and search-palette.tsx:20 hands it straight to filterNotes. Body-only queries in capitals still match (below), so the query path is fine.
      • H3 Note data in the store has altered titles: REFUTED. Seed titles are intact (src/electron/main/index.ts:5-7), and the same data matches once H1 is toggled.
    • Root cause: the title is not lowercased before being compared with the lowercased query (OBSERVED) at src/features/search/filter-notes.ts:8, introduced in c0ab390 refactor search matching (git log -p -- src/features/search/filter-notes.ts; the previous code was note.title.toLowerCase().includes(needle)).
    • Confirmation: probe npx tsx probe.ts calling filterNotes on the 3 seed notes.
      Before: "Grocery list" -> [], "grocery list" -> [], "groc" -> [], "Meeting" -> [], "verification" -> ["Meeting agenda"], "REPORT" -> ["Meeting agenda"].
      Toggled line 8 to normalize(note.title).includes(needle): all of the first four return their note, and the scenario goes PASS search/shortcut-filters-and-selects, 1/1 scenarios passed.
      git checkout src/features/search/filter-notes.ts reverted it: [] again for "Grocery list". Symptom follows the toggle both ways.
    • Observed facts:
      1. All three seed titles start with a capital letter and none is found by its own title ("Welcome to Notes" -> [], "Grocery list" -> [], "Meeting" -> []).
      2. Body matching is case-insensitive and unaffected ("REPORT" finds Meeting agenda, filter-notes.ts:8).
      3. A single commit touched the search feature after the initial import (git log --oneline -- src/features/search shows c0ab390 and 2ad2643).
      4. The browser console is clean (console.json = []), so this is wrong filtering, not a runtime error.
    • Inferences:
      1. A note whose title is entirely lowercase and typed in lowercase would still be found, so the user sees "some" notes missing (INFERRED from line 8; no such seed note exists to run).
      2. Notes whose body happens to contain the title words appear to work, which explains the "some" in the report (INFERRED; e.g. "notes" returns Welcome to Notes only through its body text).
    • Unknowns / not checked: other callers of filterNotes (only search-palette.tsx:3,20 imports it, via grep); no title containing non-ASCII characters tested.
    • Proposed fix: filter-notes.ts:8 becomes normalize(note.title).includes(needle) || normalize(note.body).includes(needle) (not applied; the repo is back at c0ab390 after the toggle).

    screenshot.png

    PDF · document

    Generated

    Example file from a real run - the skill writes it into your workspace.

    Connects securely to your tools. The creator never sees your data.

    About this skill

    Coding agents love to name a cause the moment you describe a bug, often without opening the code that would be affected. Every wrong guess costs a fix attempt, a review round and your trust.

    investigate turns "it has to be this" into an evidence-backed finding. Before your agent states a root cause or proposes a fix, it must:

    • Reproduce first: run the app, a failing command or a test, and save the artifacts. "Not reproduced" is a valid result.
    • Hold at least two hypotheses, each with what would confirm or refute it.
    • Read the real code path and cite every claim as path/to/file:line.
    • Fan out read-only subagents (at most 4 at once) for the legwork.
    • Label every statement as OBSERVED or INFERRED.
    • Earn the word "confirmed": toggle the suspected cause and show the symptom change.
    • Check recent changes with git log -p / blame on the implicated lines.

    It ends with a fixed write-up format (reproduction, hypotheses, root cause, confirmation, unknowns, proposed fix), so you can review an investigation in a minute.

    Works with any language or stack. Pairs well with an app-verification skill such as create-verification (also free on Agensi).

    Free and open source. Part of poteto-kit (MIT). Unofficial: poteto-kit distills public talks and posts by Lauren Tan (poteto) and is not affiliated with or endorsed by her.

    Needs a way to reproduce the problem: a runnable app, a failing command or a test. Without one, the agent reports "not reproduced" and an inferred cause instead of a confirmed one.

    How to use

    1. 1

      Connect your agent to the Agensi MCP

      Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup

    2. 2

      Sign in to Agensi

      Your agent shows a sign-in link. Sign in and click Allow.

    3. 3

      Ask for what you need

      Your agent loads this skill and uses it on your task. It always asks before any payment.

    Reviews

    No reviews yet

    Be one of the first to try it. Every listed skill goes through the trust checks below.

    Security scanned

    Checked by our 8-point scan before listing

    1 use

    Times used to date

    Free forever

    No account required to browse

    Trust & safety

    Security scan

    Security scan passed 29 days ago

    • Free with an Agensi account

    Requires

    Git

    Permissions

    Terminal / Shell
    Read Files
    Write Files
    Listed29 days ago

    What's inside

    Frequently Asked Questions

    It makes the agent reproduce the problem first, write down at least two hypotheses, read the code path it blames and cite it as file:line, and label every statement OBSERVED or INFERRED. It may only call something the root cause after toggling the suspected cause and watching the symptom change.

    Any agent that loads SKILL.md files and can read files and run terminal commands (Claude Code, Cursor, Codex and others). It is language-agnostic. If your agent cannot spawn subagents, it does the same reads one after another.

    Investigation is read-only: subagents only read, search and reproduce. The agent toggles the suspected cause briefly to confirm it, then writes up the finding and proposes the smallest fix. It applies a fix only if you ask.

    Yes. It is part of poteto-kit, MIT licensed, at https://github.com/EPISTEX0/poteto-kit. The repo also explains the ideas behind it and has a runnable example app.

    No. poteto-kit is an unofficial community kit that distills her public talks and posts. It is not affiliated with or endorsed by her.

    New in Testing & Debugging