Read the AXIS miss. Edit one kit skill. Re-run the same scenario. Then stop.
Stay in this repo. Do not copy to scalekit-inc/skills unless the user says ship.
Guardrails
- MUST load this skill only after an AXIS report exists.
- MUST change one
SKILL.md(plus itsreferences/if the miss lives there). - MUST keep the scenario prompt unchanged. The skill is the code. AXIS is the test.
- MUST NOT empty the workspace. Skill-loaded runs only.
- MUST NOT rewrite the skill. Surgical fix.
- MUST look up live product facts with Exa before you change an API, env name, dashboard path, or connector model. Prefer
https://docs.scalekit.com. Do not cache CLI help orllms.txt.
Then read authoring-skills (Improving an existing skill) and skill-design-principles. Apply them. Do not paste them here.
Step 1 — Name the miss
From the latest report:
- Scenario key
- AXIS score and Goal score
- Each goal check that scored under full weight
- The assistant text that failed the check
Done when: one failing check is named in one sentence.
Step 2 — Name the skill
Map the check to one folder under kits/ or skills/setup-scalekit.
If the agent never opened the skill, the miss is the description. That is authoring-skills trigger work. Still edit only that file.
Done when: one path is named.
Step 3 — Classify
| Symptom | Fix |
|---|---|
| Skill never fired | Add the customer phrase to description. |
| Fired, then skipped a fact | Put the fact in Guardrails or the step that runs before the wait. |
| Wrong API or old model | Exa-fetch the live docs page. Then fix that step. Keep one current model. |
| Over-broad MUST NOT | Scope the rule. Prompt the positive. |
| Invented secrets | Keep wait-for-user. Do not add sample values that look real. |
Done when: one row is picked.
Step 4 — Look up, then patch
If the miss is a product fact (method name, env var, callback path, Virtual MCP model, host rule), call Exa first:
exa__web_search_exafor the live docs page.exa__web_fetch_exaon that URL.- Patch only what the live page supports.
If the miss is trigger, stop line, or wait-for-user, skip Exa.
Smallest edit that makes the failing check pass on the next run.
Follow this repo writing bar: one job, Use when, ≤200 lines, live lookups stay live.
Done when: git diff shows only that skill.
Step 5 — Check and re-run
scripts/validate.sh
dora review --quick <skill-folder>
npx axis run --agent grok-local
If Goal is still short, one more surgical pass. Then stop and report.
If Goal is good: npx axis baseline set. Commit. Do not push unless asked.
Done when: the same scenario is re-run and the result is written down.