observe-skill

Logs the outcome of a skill execution to observations/<skill-name>/runs.md. Trigger on: "log this run", "skill worked", "skill failed", "this didn't work", "log the outcome", "record this", "note that

SKILL.md

GitHub

Observe Skill

Purpose

Track how skills perform in practice. A skill that works in theory may miss steps, use outdated payloads, or fire on the wrong triggers in real-world conditions. This skill captures that gap by recording execution outcomes to a per-skill log file.

Trigger Conditions

Activate this skill when the user says any of:

  • "log this run"
  • "skill worked" / "that worked"
  • "skill failed" / "that didn't work" / "this didn't work"
  • "log the outcome"
  • "record this run"
  • After any skill completes and the user describes a result (success, partial, fail)

Methodology

  1. Identify which skill was just used — ask if unclear. The skill name must match a directory name under skills/.
  2. Determine the outcome — success (goal achieved), partial (some steps worked, some didn't), or fail (skill produced no useful result).
  3. Collect details:
    • Date (today's date)
    • Task description (what was the user trying to do?)
    • What specifically worked (payloads, steps, tools that produced results)
    • What specifically failed (steps that produced no result, wrong output, or errors)
    • Any error messages or unexpected responses verbatim
  4. Write the entry — append to observations/<skill-name>/runs.md. Create the file and directory if they don't exist.
  5. Confirm — show the user the appended entry.

Entry Format

---
date: YYYY-MM-DD
skill: <skill-name>
task: <one-line description of what was attempted>
outcome: success | partial | fail
what_worked: |
  <describe what steps, payloads, or approaches produced results>
what_failed: |
  <describe what steps or payloads did not work>
errors: |
  <paste any error messages or unexpected responses>
notes: |
  <any additional context — target type, environment quirks, etc.>
---

Example Interaction

User: "log this run — bola-idor worked, I found a sequential integer ID on /api/orders and was able to read other users' orders using my own session token"

Response: Creates or appends to observations/bola-idor/runs.md:

---
date: 2026-03-14
skill: bola-idor
task: Test /api/orders endpoint for horizontal IDOR using sequential integer ID
outcome: success
what_worked: |
  Sequential integer IDs confirmed. Substituting victim's order ID (attacker ID +/- 1-10)
  returned full order details including name, address, and payment last4. Tested with
  attacker's Bearer token — no ownership check.
what_failed: |
  UUID endpoint /api/orders/uuid was not vulnerable — UUIDs not guessable.
errors: |
  None.
notes: |
  REST API, no GraphQL. /api/v1/ also vulnerable. /api/v2/ returns 403 on cross-user ID.
---

Fix Patterns

This is a meta-skill — it has no fix patterns. It generates data that other skills (especially /amend-skill) consume.