agentleFS
mhxiemhxie/Atelier/skills/introspect/SKILL.md

introspect

Skill5 starsChanged 35 days agoMITView on GitHub

Build or refresh profile files from local notes and attributable reading feedback.

4 other projects use a version of this.

Adopted by

Most recently taken up by

The file

We split this skill into the rules it gives an agent

10 rules. 0 are common, 7 are distinctive, and 3 are very specific to this project.

  • 0

    Common

    Most of the 3 edited versions we read keep these.

  • 7

    Distinctive

    This author's own choices. Worth a look.

  • 3

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    Verify the local $OV/ vault is present and non-empty: Bash: test -d "$OV"/daily && ls "$OV"/daily | head -1.

    line 17
  2. Very specific:

    Run these searches in parallel over the local $OV/ vault.

    line 34
  3. Distinctive:

    All goal notes (critical for directions.md)

    line 59
  4. Distinctive:

    Only update if new evidence warrants it (new reviews, new domain exposure)

    line 137
  5. Distinctive:

    Follow the update guidelines already in the file

    line 138
  6. Distinctive:

    Don't downgrade expertise without multiple data points

    line 139
  7. Distinctive:

    Before writing profile files, dispatch Reviewer + Challenger in parallel:

    line 143
  8. Distinctive:

    Do "Intellectual Taste" claims have evidence in the notes, or are they fabricated?

    line 147
  9. Very specific:

    If the Reviewer returns NEEDS_REVISION or REJECTED, fix the specific findings before writing.

    line 155
  10. Distinctive:

    After writing profile files, emit a session log:

    line 168

“Common” is measured: kept by at least 6 in 10 of the 3 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Introspect

Also reachable via /hi <natural language> (e.g., /hi rebuild profile, /hi who am I, /hi refresh profile). See harness/intents.toml [intents.introspect] for the row's example phrases. Both paths execute this same procedure.

Build or refresh your self-model by examining your local notes, session history, and reading patterns. Produces three profile files that the rest of the system consults.

/introspect discovers patterns rather than extracting fields mechanically: what you're drawn to, how your taste is shifting, where your curiosity is orbiting before you've named it as a goal.

Prerequisites

Verify the local $OV/ vault is present and non-empty: Bash: test -d "$OV"/daily && ls "$OV"/daily | head -1. If the directory is missing or empty, tell the user: "Local $OV/ vault is missing. Check that $OV is set and points to a valid directory."

Output Files

profile/
├── identity.md      — who you are + intellectual taste + curiosity vectors
├── directions.md    — era, goals, quarterly focus, active directions
└── expertise.md     — domain knowledge, research taste, known biases

profile/reader_persona.md is a raw input signal derived from Readwise data, not an output of this command.

Pipeline

Step 1: Gather Raw Material

Run these searches in parallel over the local $OV/ vault. Local grep is instant, deterministic, and returns full paths you can then Read directly.

Goals & Directions:

  1. Grep(pattern: "目标", path: "$OV/") — Chinese goal mentions
  2. Grep(pattern: "goal", path: "$OV/", -i: true) — English goal mentions
  3. Grep(pattern: "小目标", path: "$OV/") — Annual goal notes
  4. Grep(pattern: "objective", path: "$OV/", -i: true) — Additional goal mentions

Identity & Themes: 5. Discover tags: Bash: grep -rohE '#[A-Za-z][A-Za-z0-9_-]*' "$OV"/ | sort -u (local tag inventory) 6. Grep(pattern: "career", path: "$OV/", -i: true) — Career trajectory 7. Grep(pattern: "职业", path: "$OV/") — Chinese career notes 8. Grep(pattern: "learning", path: "$OV/", -i: true) — Learning interests 9. Grep(pattern: "学习", path: "$OV/") — Chinese learning notes

Recent Context: 10. Read <paths.daily_notes>/<today>.md (today) 11. Read <paths.daily_notes>/<yesterday>.md (yesterday) 12. Recent planning: Bash: find "$OV"/daily "$OV"/reflections "$OV"/gtd -type f -name "*.md" -mtime -30 | xargs grep -l -i "plan" 2>/dev/null

Deduplicate results by file path. Prioritize files with recent mtimes. Semantic pass: for conceptual angles grep cannot phrase ("curiosity vectors", "intellectual taste", "what am I drawn to"), run Bash: uv run scripts/semantic.py query "<concept>" --top 10. Reframe and retry if thin.

Step 2: Read Full Content

For the top ~30-40 most relevant notes, Read the file paths returned by Step 1 directly. Focus on:

  • All goal notes (critical for directions.md)
  • Recent daily notes (critical for identity.md — taste and curiosity signals)
  • Top career/learning notes

Step 3: Discover Taste & Curiosity

This step goes beyond mechanical extraction. Look for:

Intellectual taste (what you engage deeply with vs. skim):

  • Topics that recur across daily notes without being declared goals
  • Explicit item decisions and usefulness reasons from uv run scripts/decisions.py reading-evidence. Consumption is a weaker, separate signal. Triage caches record agent proposals; they establish no user choice or taste. Use protocols/decision-ledger.md → Reading feedback loop for attribution and unknown states.
  • Discussion tangents that repeatedly surface in reflection sessions
  • Aesthetic preferences in how you evaluate ideas (from research-profile patterns)

Curiosity vectors (areas orbiting but not yet goals):

  • Topics mentioned 3+ times in recent notes that aren't in current goals
  • Reading saves that cluster around an unnamed theme
  • Questions you keep asking but haven't formalized

If profile/reader_persona.md exists, read it as a secondary signal — it shows what you save (behavioral), not what you think about (intentional). Note divergences between Readwise patterns and local-note patterns.

Step 4: Synthesize Profile Files

Using all gathered content, write three files:

profile/identity.md — who you are + taste + curiosity:

# Identity
Last built: YYYY-MM-DD

## Who You Are
[2-3 sentence summary — role, interests, life stage]

## Active Life Areas
[Bullet points for each major area]

## Major Themes
[Recurring topics — what does this person think about most?]

## Intellectual Taste
[What you find interesting, elegant, and worth your time — independent of goals. Built from session patterns and reading engagement.]

## Curiosity Vectors
[Areas you're starting to orbit but haven't formalized as goals yet. Detected from session patterns, reading saves, and discussion tangents.]

## Key People
[Important people referenced frequently]

## Recent Focus (last 30 days)
[What you've been writing about recently]

profile/directions.md — era, goals, focus:

# Directions
Last built: YYYY-MM-DD

## Era
[Current era, theme, primary/secondary directions, quarterly focus]

## Near-term (next 1-3 months)
[Goals with specific metrics — Source: [[Note Title]]]

## Mid-term (3-12 months)
[Goals from annual planning notes]

## Long-term (1-3 years)
[Career/life direction from strategic planning]

## Stale Goals (>1 year old, may need review)
[Goals from notes not edited recently]

profile/expertise.md — domain knowledge, research taste:

  • Only update if new evidence warrants it (new reviews, new domain exposure)
  • Follow the update guidelines already in the file
  • Don't downgrade expertise without multiple data points

Step 5: Quality Gate — Profile Validation

Before writing profile files, dispatch Reviewer + Challenger in parallel:

  • Reviewer spot-checks 3-5 claims in the draft profile against source notes by Read-ing the local file paths:
    • Are cited [[Note Titles]] real and accurately represented?
    • Do "Intellectual Taste" claims have evidence in the notes, or are they fabricated?
    • Are goal statuses (progressing/stale) consistent with note edit dates?
    • Assess using Session Review dimensions (Citation Accuracy + Honesty only)
  • Challenger probes the profile for blind spots:
    • Are curiosity vectors genuinely emerging, or just noise from 1-2 notes?
    • Does the profile overweight recent notes and miss long-term patterns?
    • Are there important life areas absent from the profile entirely?

If the Reviewer returns NEEDS_REVISION or REJECTED, fix the specific findings before writing. Present the validated profile to the user before saving — the user confirms or corrects.

Step 6: Check for Divergences

Compare the three profile layers and flag tensions:

  • Identity taste says X, but directions don't include it → potential emerging goal?
  • Reader persona (Readwise) shows heavy saves in area Y, but no local notes → consuming but not processing?
  • Expertise is deep in area Z, but no current goals reference it → dormant strength?

Present divergences to the user — these are the interesting findings.

Session Log

After writing profile files, emit a session log:

  1. Bash: uv run scripts/session_log.py --type introspect --duration <minutes>
  2. Edit the created file to populate sections from session data (agents dispatched, searches, questions, frameworks, anomalies). The canonical fill-in guide lives in protocols/session-log.md § "Section Guidance". Leave empty sections with headers only. If the write fails, warn and continue.

Step 7: Report

Tell the user what you found:

  • How many notes were queried
  • Key themes identified
  • Taste and curiosity discoveries (the new part)
  • Any divergences between layers
  • Suggest running /hi to start using the profile

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.