blindspot

Blindspot

A reflective mirror for Claude Code — and, via the portable prompts, for any LLM.

Blindspot surfaces patterns in your own thinking, decisions, and work that are hard to see from the inside — where you're making progress, where you're stuck, and any signal that the work is starting to cost you. Every observation is tied to evidence from your own words. It is not a therapist, coach, or diagnostician: it observes and names patterns; you decide what they mean.

You run it with /blindspot, ideally at the end of a work session. An optional, gentle nudge can remind you so you don't have to remember.

/blindspot

A mirror, not a clinician. Blindspot is designed to stop analyzing and respond plainly the moment it detects distress, and to set the tool down entirely at any sign of self-harm. It is not a substitute for professional support. See docs/privacy.md.

---

How it works

Two pieces, deliberately separate:

Piece	What it is	When it runs
The skill	/blindspot — interactive, visible, you read the output	Only when you type it
The nudge (optional)	A SessionStart hook that reminds you to run /blindspot	At most once every 72h

The skill is manual by design (disable-model-invocation: true in its frontmatter), so it never fires on its own and never reacts to every message. The nudge only ever reminds you — it does not run the analysis for you. That keeps you in control of when the mirror happens, which for a self-awareness tool is the point.

Each run reads the current session's conversation plus your running log at ~/.claude/blindspot/sessions.md, and appends to it. The log keeps EVIDENCE (your actual words, timestamped — treated as fact) separate from READS (interpretations, re-derived from scratch each time). A pattern only becomes a trend when evidence across multiple dates supports it. See examples/sessions-example.md for what a populated log looks like.

Install

Marketplace (recommended)

/plugin marketplace add isas1/blindspot
/plugin install blindspot@blindspot

Then restart Claude Code or run /reload-plugins. This installs the skill and the optional nudge.

Local (development)

git clone https://github.com/isas1/blindspot
cd blindspot
bash scripts/link-skill.sh   # symlinks the skill into ~/.claude/skills (no hook)

Full walkthrough — including the recommended "test by hand before automating" ritual — is in docs/getting-started.md.

The nudge (optional, semi-automatic)

Installed automatically with the marketplace plugin. It is conservative on purpose: a single one-line reminder, at most once every 72 hours, surfaced through Claude's first reply.

Configure it with environment variables:

export BLINDSPOT_INTERVAL_HOURS=48   # minimum hours between nudges (default 72)
export BLINDSPOT_NUDGE=off           # disable the nudge entirely

Force it while testing:

rm -f ~/.claude/blindspot/.last-nudge   # clear the gate, then start a new session

Want it fully hands-free instead?

By default the skill is manual and the hook only reminds you. If you'd rather Claude run the check-in itself once per interval, two small edits:

1. In skills/blindspot/SKILL.md, remove the disable-model-invocation: true line.
2. In hooks/session-start.sh, change the reminder message to instruct Claude to run /blindspot rather than mention it.

This trades control for convenience. Test the manual version first; the trap of automation is producing reflections you never actually read.

Use it with any LLM

Blindspot is, at its core, a prompt and a log — neither needs Claude Code. The portable/ folder has two self-contained versions:

• portable/blindspot-prompt.md — the full mirror. Paste it at the end of any chat conversation (the paste itself triggers the check-in), or install it as ChatGPT Project / Custom GPT / Claude.ai Project / Gemini Gem instructions, or as a local model's system prompt.
• portable/blindspot-compact.md — a ~1.6k-char card for small custom-instruction fields.

Without a filesystem, you carry the memory: keep a private note, paste it in at check-in, and copy the fenced log entry the mirror emits back into the note. Same EVIDENCE/READS discipline, same strain tier, same hard stop. Per-platform setup: docs/any-llm.md.

Privacy

Everything stays on your machine in ~/.claude/blindspot/. Nothing is uploaded, synced, or sent anywhere. The file accumulates personal — sometimes sensitive — signals, so treat it like a journal. Details and deletion instructions: docs/privacy.md.

Repository layout