Getting started

Ten minutes from install to the first scouting run.

  1. Requirements
  2. 1. Install the workspace
  3. 2. Describe yourself, briefly
  4. 3. Start Claude Code
  5. 4. Scout
  6. 5. From signals to an idea
  7. 6. Run one experiment
  8. 7. Check and commit

Requirements

  • Claude Code, signed in.
  • git.
  • Python 3.10 or newer, reachable as python3. The engine uses the standard library only; there is nothing to pip install.
  • macOS or Linux. The installer is plain bash and also runs on the bash 3.2 that ships with macOS. On Windows, use WSL.

1. Install the workspace

Pick a new, empty folder: the workspace refuses product code, so it never shares a folder with a code project. Two ways to fill it, with the same result.

With the Claude Code plugin (Claude Code 2.1.295 or later). In any Claude Code session:

/plugin install panlode --marketplace milosCvetkovicDev/panlode
/panlode:init ~/ideas

/panlode:init checks that the folder is not a code project, shows the plan, then runs the installer. The plugin itself adds only /panlode:init and /panlode:update; it installs no hooks, so other projects are not affected.

With git. From your shell:

git clone https://github.com/milosCvetkovicDev/panlode.git
panlode/install.sh ~/ideas

Then put the workspace under version control:

cd ~/ideas && git init && git add -A && git commit -m "chore: start the idea workspace"

The installer copies template/ into the folder you name and never overwrites a file that already exists. It prints a line for each file it adds (add) or leaves alone because yours differs, it is a symlink or it sits in a symlinked folder (skip), and a summary; it generates INDEX.md only when the folder has none. It refuses your home folder, the root folder and any parent of the home folder, in any spelling, because a workspace’s .claude/ and CLAUDE.md would then apply to every project below it. It also refuses a path with a newline, a .. after a folder that does not exist yet, or a dangling symlink. Before it writes anything, dry run included, it stops when the engine or .claude/settings.json is missing behind a symlink in the folder (a dangling scripts/ link, say), because the hooks would then do nothing. Useful options:

Option Effect
--dry-run Print the plan; write nothing.
--force In an existing workspace only, refresh the engine files: scripts/, .claude/agents/, .claude/rules/, .claude/skills/, templates/. Your CLAUDE.md, founder/ files, .claude/settings.json and records are never touched. Commit first: an engine file you edited is replaced too, and git brings it back. /panlode:update runs this for you.

2. Describe yourself, briefly

Open the three files in founder/ and replace every <placeholder>:

File What to write What it feeds
profile.md what you can build and deliver fast score_fit, score_speed
constraints.md capital, budget, capacity, and vetoes: the kinds of business you will not do every session (imported into CLAUDE.md) and the scouts’ skip list
channels.md where you can actually reach prospects score_reach

The founder files filter ideas; they never generate them. “I could build this” is not a signal.

If you want Claude to converse in another language, change the first sentence of the Language section in CLAUDE.md. Records stay in English unless you decide otherwise; file names are always ASCII.

3. Start Claude Code 

claude

The SessionStart hook gives Claude the brief. In an empty workspace it reads:

Idea engine — 2026-10-24
Empty pipeline — start with /scout-online.
Minimum loop: one /scout-online per week; ≥2 prospect contacts logged per running experiment.

4. Scout 

/scout-online
/scout-online accounting firms e-invoicing     # optional focus

Three signal-scout subagents run in parallel, one per recipe in .claude/skills/scout-online/sources/: freelance and job posts, review sites, Reddit and forums. Each returns at most eight candidates with a source link, plus the searches that led nowhere. Claude removes duplicates and anything that hits a veto, scores the evidence (taking the lower score when unsure), writes each kept candidate to signals/ and prints a digest with suggested clusters.

Every write goes through the hooks: a malformed record is reported at once and fixed before the next.

5. From signals to an idea

When two or more signals share a segment or a pain, mark them clustered with the same cluster slug. When a cluster has enough evidence (the digest suggests it once three signals score 3 or more), shape an idea:

python3 scripts/engine.py next-id idea          # e.g. idea-0001
python3 scripts/engine.py slug "Invoice import for small accounting firms"

Ask Claude to write ideas/<slug>/idea.md from templates/idea.md. The rules for ideas load automatically when Claude works in ideas/.

6. Run one experiment

Write the experiment from templates/experiment.md, in this order so every reference resolves:

  1. the idea with status: shaped;
  2. the experiment with idea: idea-NNNN, a success_metric (numeric where possible), a budget_eur and a deadline;
  3. the idea set to validating, with next: exp-NNNN and dated kill_criteria.

Log every contact and result as it happens:

/log-experiment exp-0001 sent 10 cold emails to accounting firms, 2 replies, 1 call booked

When the deadline passes or the metric is met, /log-experiment asks for continue, pivot or kill and updates both records.

7. Check and commit 

python3 scripts/engine.py check     # exit 0 means every invariant holds
git add -A && git commit -m "docs(signals): first scouting run"

INDEX.md is regenerated by the hooks; never edit it by hand. Next: the weekly loop.