Plugin Claude Code
Turns your agent into a trainable employee. Learns rules from your corrections (with your approval), and skills earn autonomy through demonstrated reliability — proposed → verified → trusted → autonomous. No daemon. No cloud.
Instructions file
Instructions for ankitkr3/compounded, covering claude.md (compounded repository), what this repo is, conventions, how the lifecycle works and what you can change.
Agent
A subagent dispatched by compounded's Stop hook to decide whether a proposed skill should graduate to .verified or move to .rejected/. Reads the proposed SKILL.md and the just-completed task transcript, applies the rubric in the compounded-verifier skill, and returns a JSON verdict.
Command
Manually demote a skill one tier (.autonomous → .trusted → .verified → .proposed). Pass the skill name as $ARGUMENTS.
Command
Export compounded state (USER.md + verified skills) to a portable .tar.gz archive. Pass the output path as $ARGUMENTS, e.g. /compounded:export /Desktop/compounded.tar.gz. Add --include-pending to also include .proposed/ and .rejected/.
Command
Import a compounded archive (.tar.gz produced by /compounded:export). Pass the input path as $ARGUMENTS. Add --overwrite to replace any colliding local skills.
Command
Pin a skill at its current trust state, preventing automatic demotion. Pass the skill name as $ARGUMENTS. Pinning is reversed by /compounded:unpin.
Command
Show compounded status — USER.md size, skill counts by trust tier, recent transitions, recent rejections.
Command
Show the compounded trust ladder. Lists every skill at every tier with use/correction counts, highlights skills approaching promotion or demotion.
Command
Manually promote a skill (skip the trust gradient). Usage /compounded:trust --to verified|trusted|autonomous. Use sparingly — manual overrides defeat the verification system.
Command
Remove a pin from a skill, allowing automatic demotion to resume. Pass the skill name as $ARGUMENTS.
Command
Manually run the verifier subagent against a proposed skill. Pass the skill name as $ARGUMENTS. Use this when you want to verify a proposal immediately rather than waiting for an applicable task to come up naturally.
Hook
Runs when a session starts on startup, resume, clear and compact, executing memory_inject.py via python3. From ankitkr3/compounded.
Hook
Runs when the agent finishes a response, executing skill_verify.py and auto_propose.py via python3 with --check-pending (2 commands). From ankitkr3/compounded.
Hook
Runs when a session ends, executing trust_ladder.py via python3 with --process-session. From ankitkr3/compounded.
Skill Claude CodeCodex
Use this skill when you have just completed a non-trivial task (more than 3 tool calls, novel approach, recovery from a dead end) and are considering whether to save the procedure as a reusable skill via compounded; OR in RULE MODE when the user corrected your approach and the correction encodes a generalizable…
Skill Claude CodeCodex
Use this skill when working with skills already at .verified, .trusted, or .autonomous trust levels — particularly when deciding whether to invoke a verified skill on a task, when applying a trusted skill (which requires confirmation), or when an autonomous skill auto-applies. Also use when a skill produces output the…
Skill Claude CodeCodex
Use this skill when the user reports a typo in their codebase or documentation and asks for it to be fixed across the project, or when reviewing code and noticing a misspelled identifier, comment, or string that should be corrected consistently across all occurrences.
Skill Claude CodeCodex
Use this skill ONLY when invoked as the skill-verifier subagent by compounded's Stop hook. This is the rubric for deciding whether a proposed skill should graduate to .verified or move to .rejected/ based on its replayability against a just-completed task. Do not invoke this skill in normal sessions.
Skill Claude CodeCodex
Use this skill at the start of any session where compounded is installed, or whenever the user mentions earning trust, verifying skills, the trust ladder, compounded, the trust gradient, autonomous skills, or asks how compounded works. Explains the four trust states (.proposed, .verified, .trusted, .autonomous) and…