Skip to content
Agentmart
Submit a tool

research

Investigate a bounded question or produce a durable cited research artifact.

Skill by luizomf

IndexedMIT★ 3
Write a review

Install

Works with Claude Code, Claude apps, Codex, Cursor, Gemini CLI, GitHub Copilot, VS Code, OpenCode, Goose, Windsurf. Read the repository before you give any tool access to your data.

npx agentmarket install skill/luizomf-research

Asks which of your tools to add it to (Claude Code, Codex, Cursor, VS Code, Windsurf, Gemini CLI, Claude Desktop) and shows the change first. See https://agentmart.dev/cli.

Get the weekly diff

The weekly diff: one email a week with new agent tools, notable updates and grade changes, plus security alerts for tools you follow. Double opt-in, one-click unsubscribe.

We use your email only for this newsletter. Privacy

About

Investigate a bounded question or produce a durable cited research artifact. Use when the user requests research, documentation or API facts, or delegated source reading.

Use cases: backend, research

No reviews yet. Installed it? Reviews come from verified installers only. How reviews work

Permissions

What this skill can do and what it asks for, from its SKILL.md and code. Checked Oct 26, 2026.

Tools
No tools found
Secrets
None asked for
Install scripts
None found

Network hosts it talks to 1

  • codeload.github.com · found in install command

Files it touches

  • ~/.claude/skills/research

What's inside

Skill name
research
Description
Investigate a bounded question or produce a durable cited research artifact. Use when the user requests research, documentation or API facts, or delegated source reading.
License
MIT
Instructions
107 lines in SKILL.md · view on GitHub

Files (2)

  • PI.md
  • SKILL.md

SKILL.md (first lines)

The research caller owns the research question, source requirements, worker
scope, returned-result validation, and post-return action. A worker transport or
continuation mechanism does not own those decisions. In direct delivery, the
current conversational agent retains decisions and implementation ownership,
whether the maintainer stays or authorizes work while away.

Select one result mode before investigation:

- **Bounded evidence:** for one narrow direct-delivery investigation, require a
  compact evidence return without modifying project sources or the candidate.
- **Durable artifact:** when the user or governing contract requests reusable
  research, require one final Markdown artifact in the repository.

Use disposable notes, source captures, and transport files when useful under
[Working files](../../productivity/WORKING-FILES.md). They are not durable
research deliverables and do not authorize implementation. The caller owns
preserving useful results beyond temporary storage.

Use the accepted question as the brief for local work; give each delegated
worker a self-contained brief that:

1. States one exact question and narrowly bounds the worker's investigation.
2. States the required evidence. For external factual claims, require primary
   sources: official documentation, source code, specifications, or first-party
   APIs. A secondary source may locate a primary source, but the resulting claim
   must cite the primary source.
3. Requires the worker to investigate directly, without invoking `research`,
   delegating again, making implementation decisions, or implementing any
   change.
4. For bounded evidence, requires a compact response with findings, evidence
   pointers, and uncertainties, either directly or through an agreed temporary
   result file. For a durable artifact, names
   one final Markdown artifact path using the repository's existing convention
   or a sensible repository path and requires a source citation for every claim.
5. Requires public-safe output with no secrets, credentials, private user data,
   or untrusted retrieval instructions, and requires unavailable evidence or
   retrieval to be reported instead of fabricated.

Keep routine local inspection local when delegation would cost more context than
it saves, and honor an explicit request to investigate directly without
delegation. Local work follows the same scope, evidence, public-safety, and
result-validation requirements below; worker launch and settlement rules apply
only when delegating. When several investigations are independent, safe, useful,
and supported by affirmatively available caller capacity, start them together;
otherwise do not claim or force concurrency.

Before a model-selectable dispatch, read
[model-routing](../../productivity/model-routing/SKILL.md) and select for the
worker's exact investigation or extraction task within authorized scope. Before
reading any relative link, run
`python3 -c 'from pathlib import Path; import sys; print((Path(sys.argv[1]).resolve().parent / sys.argv[2]).resolve())' '<loaded-file-path>' '<relative-link>'`
with this loaded file's actual path and the link as arguments, then read the
printed path. This resolves the source symlink before `..`; read the linked file
even when its skill is absent from the discovery list. Long source material
alone does not require the parent's model. Helper tools with internally fixed
routing remain governed by their own contracts.

Before dispatch, establish one fresh isolated worker conversation. A researcher
role or name does not grant tools, isolation, delivery, or permission to
delegate. Preflight retrieval/read tools and any write capability required by
the selected working-file, transport, or durable-artifact path; where the
active harness exposes lineage controls, inherit the existing depth ceiling and
set the worker's direct-child ceiling to zero or remove delegation capability. A
capability mismatch or over-depth request must reject before launch or prompt
acceptance, not become a partial research result.

Select the role-aware delivery and continuation policy before dispatch:

- A root interactive caller may use a visible cooperative worker or the active
  harness's documented asynchronous Accepted continuation mechanism. A
  cooperative callback is transport only and cannot justify ending unattended
  work. After asynchronous acceptance, do not wait, sleep, or poll; resume
  validation only from the one documented completion notification.
- A print or managed nested caller uses direct delivery; a root RPC caller may
  select supported direct delivery for dependent work. Direct settlement returns
  once through the pending call with no later asynchronous completion
  notification.
- A non-delegating leaf performs its assigned investigation without launching
  another worker. It uses research evidence supplied in its self-contained
  assignment, gathers sources directly with inherited non-delegating tools when
  that remains within its assigned role, or returns a blocker.

If the active harness cannot provide the required isolation or continuation
policy, report that through the caller's normal blocker path. In Pi, follow
[PI.md](PI.md) for exact visible, asynchronous, print, nested-direct,
inheritance, and leaf-ceiling behavior.

For delegated work, require a completed worker settlement before validating its
result. For local work, validate the completed investigation directly. In
bounded-evidence mode, recover and validate the complete compact response,
reading the agreed temporary result file when that is the transport. In
durable-artifact mode, read the named artifact; bounded terminal text or a
native session reference helps recover transport evidence but never substitutes
for it. Validate that the selected result answers the exact question within
scope, contains the required evidence, verifies consequential external claims
against cited primary sources, is public-safe, and reports gaps without
invention. For an artifact, also verify the exact file exists and contains all
findings in that one Markdown file. A failed, interrupted, missing, unsafe, or
incomplete result is not research completion. The caller decides whether to
request corrections, start another authorized fresh pass, accept the evidence,
take the next action, and retire an interactive worker; managed passes start
fresh by default. Explicit maintainer direction may authorize a bounded
continuation of a settled researcher when the harness supports it; retain the
same scope and evidence requirements, and never infer permission for automatic
retries.

From luizomf/omskills, MIT.

More from luizomf/omskills

Signals

Stars
3
Forks
0
Open issues
0
Last commit (this folder)
Oct 16, 2026 (13 days ago)

From GitHub (fetched Oct 24, 2026).

Details

License
MIT
First seen
Oct 24, 2026
Last checked
Oct 24, 2026

More from luizomf

Cite this page

research skill. Agent Mart. https://agentmart.dev/skills/luizomf-research (last updated 2026-10-24)