calibrate

# Calibrate: In-Session Self-Improvement ## What This Does Reviews the current conversation and suggests specific updates to skills, instruction files, memory, or workflows based on what just happened. The user picks which suggestions to apply. ## Two modes - **Full (default)** — deep scan of the whole conversation, max 7 suggestions. Use at end of long sessions or after multiple corrections. - **Lite** — triggered by `calibrate lite` or `/calibrate lite`. Fast 1-3 suggestion sweep for shorter sessions. Skips deep pattern matching. Just surfaces the loudest 1-3 things from the most recent 10-20 turns. ~80% fewer tokens. If a lite sweep returns "nothing to calibrate", that's a valid output — say so in one line and stop. ## When to Use - End of a work session - After a task with multiple corrections - When the user says "calibrate", "what can you improve", "tune up", "update your skills" ## Lite mode shortcut If trigger is `calibrate lite` / `/calibrate lite`: 1. Scan only the last 10-20 turns (not the whole session) 2. Look for at most 3 things — corrections that happened, preferences revealed, repeated patterns 3. Output max 3 numbered suggestions in the same `[TARGET: file] — [what]` format 4. If nothing meaningful surfaced: respond `Clean — nothing to calibrate` in one line. Don't pad. Lite is the right call when: - Session was short (under ~20 turns) - Most of the work was unmemorable (routine queries, lookups) - You just want a quick "anything I should fix?" without a deep audit ## Process ### Step 1: Scan the Conversation Review everything in the current session. Look for: - **Corrections** — where the user said "no", "wrong", "not that", "I meant X" - **Preferences revealed** — format choices, tone adjustments, workflow preferences - **Process gaps** — steps you missed that a skill should encode - **Repeated patterns** — same type of correction more than once = systemic issue - **Things that worked** — approaches the user confirmed or accepted without pushback ### Step 2: Match to Updateable Targets For each finding, identify WHERE the fix belongs: - **Skill file** — if the issue is about how a specific task is executed - **Instruction file (CLAUDE.md or equivalent)** — if the issue is about general agent behavior or rules - **Memory** — if it's a preference or context that should persist across sessions - **Workflow / scheduled task** — if it's about when or how automated tasks run - **Tone / style file** — if it's about voice or writing style (rare, check with the user first) ### Step 3: Present Suggestions Format as a numbered list. Each suggestion should be: ``` [number]. [TARGET: filename] — [what to change] Brief why: [one sentence explaining the pattern that triggered this] ``` Example: ``` 1. SKILL: thumbnail-gen — Add rule to always generate 3 variations, not 1 Brief why: User asked for alternatives twice this session after I gave a single option. 2. CLAUDE.md — Add "never suggest routing to other agents" to red lines Brief why: User corrected me when I suggested handing off. 3. MEMORY — Save preference: user prefers kebab-case for all file names Brief why: User renamed two files I created in camelCase. ``` Rules for suggestions: - **Max 7 suggestions.** If you find more, prioritize by impact (repeated corrections > one-offs). - **Be specific.** Not "improve writing quality" but "add rule: no sentences over 20 words in posts." - **Only suggest changes the data supports.** Don't pad with generic improvements. - **If nothing meaningful to update, say so.** "Clean session — nothing to calibrate" is a valid output. ### Step 4: Apply Wait for the user to respond with which numbers to apply (e.g. "do 1 and 3", "all", "skip 2"). Then: 1. Make the changes to the target files 2. Confirm each change with the file path and a one-line summary 3. If updating a skill, re-read it after editing to verify it's coherent 4. If updating memory, follow your memory system's rules (write file + update any index) ### Step 5: Done No follow-up needed. Changes are live for the next time the skill/rule is invoked. ## What NOT to Suggest - Changes that are already in the relevant skill or instruction file (check first) - Generic "best practices" not grounded in this session's data - Temporary fixes for one-off situations - Changes to files outside your write scope