Outside

outside -- external grounding pass (prior art + workspace + notes)

The gap this fills: /breakdown decomposes, /fanout diverges internally, /redteam attacks -- none of them LOOK OUTSIDE the session. This one does. It forces "shipped solutions first": adopt or adapt what exists is the default; building from scratch needs a stated reason.

Copy this checklist into your response and check items off as you complete them:

• [ ] 0. Oriented (project state file read, if in a project)
• [ ] 1. Tunnel named (problem restated + current in-head approach stated)
• [ ] 2. Shipped-solutions sweep (web)
• [ ] 3. Workspace sweep (sibling projects)
• [ ] 4. Notes check
• [ ] 5. Report (verdict first, one next action last)

0 -- orient

If in a project with a state/checkpoint file not fully read this session, read it to EOF first. Its recorded decisions and failed approaches may already contain the answer -- or be the tunnel itself.

1 -- name the tunnel

Two sentences, no more:

• The problem, restated stripped of any approach ("we need X", not "we need to build Y

for X").

• The current in-head approach, named explicitly. This is the thing on trial.

2 -- shipped-solutions sweep

Web search for what already exists for this exact ask: GitHub repos, published skills/MCP servers, community writeups, a library, a product someone already ships. 1-3 searches. Prefer free search tools over metered scraping services.

3 -- workspace sweep

Has a sibling project already solved this problem shape? Search the state files, decision docs, and specs of the 2-3 most plausible sibling projects, not the whole tree. Gotcha: if the workspace uses a deny-by-default .gitignore, ripgrep-backed search tools silently skip most files -- use plain grep -rin instead.

4 -- notes check

If the user keeps a notes system (Obsidian vault, wiki) and the project points at a relevant note, read that ONE note. Otherwise skip -- do not trawl the notes system.

5 -- report (this cap BINDS)

Order is fixed:

1. Verdict line first: current approach survives / dies / mutates -- plus the

one-clause why.

1. What exists: 2-3 genuinely different ways to attack the problem, numbered, each

with its strongest source (link or file:line). Max 5 items total. Adopt/adapt beats greenfield; recommending greenfield requires saying what WAS found and why it doesn't fit.

1. One next action, last line. Exactly one.

Source links go inline in the report; write a findings doc only if this feeds a real milestone.

Rules

• One pass, no subagents. A dry sweep is a RESULT ("nothing shipped does this") -- report

it plainly, don't pad.

• If the sweeps confirm the current approach, say so and stop; do not manufacture

alternatives to look useful.

• This is reconnaissance, not a research milestone. If it uncovers a rabbit hole worth

real research, put THAT in the report as the one next action -- don't dive inline.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: dtiger1889-ops
• Source: dtiger1889-ops/claude-harness-toolbox
• License: MIT

Install and usage instructions live in the source repository linked above.