This is a self-improving skill collection. The meta-agent (root SKILL.md) discovers, devours, and creates new skills autonomously.
- Read
CONTEXT.mdfirst — it defines the project vocabulary - Skills live in
skills/— each has a SKILL.md with YAML frontmatter - The meta-agent's memory lives in
memory/— state files that persist between sessions - Run
/setupon first use to configure user preferences - Quality bar: every skill must pass
node scripts/validate-skill.js
| Skill | Slash Command | Solves |
|---|---|---|
| setup | /setup | First-time configuration |
| ai-debt-detector | /ai-debt-detector | "AI code has hidden bugs nobody catches" |
| grill | /grill | "Building the wrong thing" |
| prove-it | /prove-it | "Agent says 'done' but it's NOT" |
| diagnose | /diagnose | "Something broken, don't know why" |
| self-review | /self-review | "Agent shows buggy first drafts" |
| zoom-out | /zoom-out | "Lost in unfamiliar code" |
| handoff | /handoff | "Need to continue later" |
| session-guard | /session-guard | "Long sessions degrade" |
| context-builder | /context-builder | "Agent doesn't speak my language" |
| dynamic-workflow | /dynamic-workflow | "Complex task needs parallel work" |
| mcp-conductor | /mcp-conductor | "Agent uses one tool when it should chain five" |
| git-workflow | /git-workflow | "Drowning in git complexity" |
| web-perf | /web-perf | "Site is slow, don't know why" |
| error-resilience | /error-resilience | "Code breaks in production" |
| arch-from-code | /arch-from-code | "Nobody knows how codebase works" |
| db-schema | /db-schema | "Database breaks at scale" |
| Script | Purpose |
|---|---|
scripts/route-task.js |
TF-IDF semantic routing with compound skill matching |
scripts/index-memory.js |
Build indexed learnings (category + keyword + apply_to) |
scripts/skillopt.js |
SkillOpt self-improvement loop (record → analyze → propose → apply) |
scripts/orchestrate.js |
End-to-end project orchestration (7-phase guided plan) |
scripts/discover.js |
Multi-strategy skill discovery |
scripts/self-improve.js |
Monitor repos + weakness detection |
scripts/rl-update.js |
Reinforcement learning signal processing |
scripts/validate-skill.js |
Quality validation for skills |
After every action, the agent runs a mandatory self-check (see root SKILL.md Phase 5). This is not optional.
- No lazy implementation — throw compute at quality, never cut corners
- Use installed skills as a pipeline (route → grill → create → prove-it → self-review)
- README frames skills as PROBLEMS not features
- Git push requires HTTP/1.1:
git -c http.version=HTTP/1.1 push(HTTP/2 fails intermittently) - Permanent Cursor install uses directory junction:
~/.cursor/skills/skill-forge→ this repo - Tests must pass (104 tests):
node tests/run-tests.js - All skills must validate with 0 warnings:
node scripts/validate-skill.js skills/<name> - Marketplace install:
/plugin marketplace add Adit-Jain-srm/skill-forge - Memory index must be rebuilt after changing learnings:
node scripts/index-memory.js - SkillOpt records outcomes:
node scripts/skillopt.js --record --skill <name> --task "..." --outcome <good|poor|mixed> - Fresh installs: run
skill-forge initor scripts auto-initialize frommemory/defaults/ - Repo ships 70 files to users;
.agents/,.claude/, runtime state are gitignored - Publishing target: GitHub under
Adit-Jain-srmorg - Context compaction survival: hooks are ineffective post-compaction; behavioral self-enforcement (recite rules at checkpoints, re-read key files) is the only reliable countermeasure
- Personal Cursor rules at
~/.cursor/rules/: max-effort.mdc, skill-orchestration.mdc, prompt-amplifier.mdc (allalwaysApply: true, not in repo — machine-level config)