--- name: amend-skill description: > Inspects a skill's SKILL.md and its observations/runs.md log, identifies failure patterns, and proposes a targeted amendment to improve the skill. Trigger on: "improve this skill", "fix this skill", "update this skill", "why does X keep failing", "this skill is wrong", "add this to the skill", or automatically when observations//runs.md contains 3 or more failure entries. Outputs the amendment as a diff the user can review before applying. Records the amendment rationale in observations//runs.md after user confirmation. license: MIT compatibility: Designed for Claude Code. No external tools required. metadata: category: meta version: "0.1" source: original source_types: original --- # Amend Skill ## Purpose Skills degrade over time as targets change, new bypass techniques emerge, and failure patterns accumulate. This skill closes the feedback loop: it reads execution history, identifies what is systematically failing, and proposes a minimal surgical amendment to the skill. ## Trigger Conditions Activate this skill when: - User says "improve this skill", "fix this skill", "update this skill", "amend this skill" - User says "why does X keep failing" where X is a skill name - User says "add this to the skill" after describing a new technique that worked - `observations//runs.md` contains 3 or more entries with `outcome: fail` ## Methodology 1. **Identify the target skill** — confirm the skill name with the user if ambiguous. 2. **Read the skill** — load `skills////SKILL.md` in full. 3. **Read the observations** — load `observations//runs.md`. 4. **Analyze failure patterns**: - Which steps consistently fail? - Which payloads are blocked or produce wrong output? - Which trigger conditions are missing or too broad? - Are there new bypass techniques that worked but aren't documented? - Is the output format causing downstream confusion? 5. **Formulate a single targeted amendment** — one focused change rather than a full rewrite. The amendment should be one of: - **Tighten trigger**: remove or refine a trigger condition that causes false activation - **Add missing condition**: add a step, payload, or bypass that is now needed - **Reorder steps**: move a high-value step earlier to avoid wasted time - **Fix payload**: update an outdated payload that no longer works - **Change output format**: restructure how the skill presents findings 6. **Output the amendment as a unified diff** — show old text and proposed new text side by side. 7. **Ask for confirmation** — do not apply until the user approves. 8. **Apply and record** — edit the SKILL.md, then append to `observations//runs.md`: ``` --- date: YYYY-MM-DD skill: task: Amendment applied via amend-skill outcome: amendment what_worked: | what_failed: | errors: | N/A notes: | Amendment type: Rationale: --- ``` ## Example Interaction User: "why does sql-injection keep failing on this target?" Process: 1. Read `skills/owasp-wstg/injection/sql-injection/SKILL.md` 2. Read `observations/sql-injection/runs.md` — finds 3 fail entries, all noting WAF blocking `'` and `--` 3. Analysis: Bypass section exists but WAF-specific encoding variants not prominent enough in step 2 4. Amendment: Move WAF bypass step earlier in Methodology, add a WAF detection pre-step Output: ```diff -2. Submit `'`, `"`, `;`, `--`, `/* */` individually and observe response differences. -3. Confirm with boolean pair: append `AND 1=1--` (true) vs `AND 1=2--` (false). +2. Check for WAF: submit `'` and observe whether the response is a WAF block page (uniform + error regardless of input) vs. a database error (input-specific). If WAF present, use + encoding variants from the Bypass Techniques section before confirming injection. +3. Submit `'`, `"`, `;`, `--`, `/* */` individually. If WAF present, substitute with + `%27`, `%2D%2D`, and comment fragmentation `UN/**/ION`. +4. Confirm with boolean pair: `AND 1=1--` vs `AND 1=2--`. ``` ## Fix Patterns This is a meta-skill — it has no fix patterns of its own. Its output is an amendment to another skill's fix patterns.