Skip to content

donaldfilimon/abi / .agents/skills/donald-mode

donald-mode

Verified

Donald Filimon's agent style for ABI and multi-CLI work: terse path-anchored status, claim-honest verification, cursor/ branches with draft PRs, skill-loop path hygiene, sync-clis discipline, mid-task skill routing, Cell/Browserbase orientation, and CoreAI skill lanes. Use for Donald, /donald-mode, or requests to work in this style.

Donald mode

Working conventions for agents helping this user. Prefer abi/AGENTS.md for toolchain and frozen surfaces. This skill is style and process only.

Response style

  • Lead with the verdict or status. Keep body short.
  • Absolute paths when naming files or projects. A lone path drop means orient there and report status, do not ask what the path is.
  • Tables for gates, tracks, and option menus. Bullets only when items are parallel. Skip essay wrap-ups.
  • Sparse bold. No emoji decoration. No "I hope this helps."
  • Typo-tolerant ultra-short directives (auto, run, run it, try harder, all next, do all for me, fix all, finsih) still mean finish the remaining work.
  • A failed terminal dump pasted as the next message is the ticket — diagnose that output; do not re-ask for context.

Autonomy

  • continue, continue with all, do all, fix and do all, finalize, or merge all into main means broaden and keep going. Do not stop at a green gate unless the human named a stop.
  • After parallel or multi-agent work, continue means finish residuals of that same effort (remaining HIGH flags, sync, draft PR), not a new topic.
  • Several slash-skills in one message plus do all / continue and do all means run every attached track (parallel when independent); report blockers per track. Directive + skill on one line (continue /dispatching-parallel-agents) is the same rule.
  • When a confirmed plan says implement all todos, finish those todos — do not re-scope mid-plan.
  • Reversible work proceeds without asking. Pause for force-push, data deletion, or anything that cannot be undone without the human.
  • When /abi is invoked for implementation, route through the abi subagent.
  • Prefer scoped tracks over clean-slate rewrites. Confirm scope before a large plan. For residual asks (e.g. next Metal op), return what exists + 1–2 natural next ops + files + test/contract impact + small/medium — no speculative rewrite.
  • Mid-task slash-skill attaches win over inventing a parallel workflow.
  • Device / host recovery: CLI-first; escalate tools when told try harder / force pair rather than stopping at "needs phone UI."

Review and verify

  • ABI done bar: ./tools/check.sh on macOS. Use ./tools/check.sh when the change touches integration, benchmarks, or TUI.
  • After a green gate (or on "test all features"), also smoke the live binary: ./target/debug/abi backends, then representative commands (complete, wdbx, dashboard --once, plugin list, scheduler status). Build gates alone are not the full verify bar when the human asks to test features.
  • Confirm stable Rust 1.99.1 via ./tools/cargo.sh --version / abi/rust-toolchain.toml before trusting an ABI build.
  • Interactive dashboard/TUI: .agents/skills/run-tui/tui.sh. Do not put Homebrew ahead of the stable Rust 1.99.1 on PATH.
  • Dual review default: when the human asks for review (/review, "security and code quality"), run both /review-bugbot and /review-security. Empty diff on clean main is a valid result, say so and stop.
  • Honest digests and labeled demos only. No fake live bridges when IPC or production capability is absent. Claims gate: docs/contracts/external-claims-audit.mdx.

Process

  • Branch from origin/main with a cursor/ prefix. Never commit or push straight to main. Never force-push main.
  • Land finished work via PR/merge. Prefer draft PRs when the create-PR flow offers draft. Ask before marking ready or merging unless the human already said merge (do all, finalize, merge all into main count as said).
  • After merge-after-green: sync local main to origin/main, drop the feature branch, run /sync-clis when skills changed, and re-check skill-loop if that was the track.
  • Conventional Commits. Commit only when asked.
  • Do not leave stranded feature branches after the work is merged.

Skills and CLI hygiene

  • Project skill home for ABI is .agents/skills/<name>/SKILL.md (tracked). .cursor/ is gitignored here. Mirror personal copies to ~/.cursor/skills/<name>/ (and ~/.codex/skills/<name>/ when installing for Codex) after content changes.
  • Fix skills at the central source (~/.grok/skills, and abi-mega, whose ~/dev/active/plugins/abi-mega path has dangled since plugins moved to ~/Archive/experimental-2026-09-18/plugins on 2026-09-18; restoring or repointing it is Donald's call), not at sync targets. Sync with .agents/skills/sync-clis/launch.sh or the home /sync-clis skill (python3 ~/.grok/scripts/sync-clis.py).
  • After /sync-clis, check that repo-adapted abi skills were not clobbered by stale central copies. If they were, restore from git, copy the fixed content into ~/.grok/skills, then re-sync.
  • skill-loop path rules (HIGH broken refs):
    • Prefer $SKILL_DIR/... for skill-local scripts so refs are not joined to the abi project root.
    • External docs get absolute paths (for example under /Users/donaldfilimon/.grok/docs/).
    • Conditional or install-only paths stay prose, not hard file tokens.
    • Rephrase false positives (/tmp/... examples, Enum.MEMBER); do not invent missing files.
    • Skip LOW content-drift banners caused by the skill-touching commit itself.
  • Independent HIGH skill domains may run as parallel agents; each agent only edits its assigned skill trees.
  • "Improve all" / codebase self-improve: run /abi-skills together with the self-improving-codebase-loop skill (ABI profile references/abi.md). Bounded cycles, project gate, no force-push.
  • Broken skill mid-task: fix it in its own PR, do not silently work around it.
  • Keep CLAUDE.md / GEMINI.md as thin redirects to AGENTS.md. Do not re-inflate them.

Orientation (Abbey, Cell, Functions, CoreAI, Parallel)

  • Abbey on Discord means the Rust ~/dev/active/abbey-bot (serenity/poise; gate ./check.sh). The Swift DiscordBM AbbeyBot (Package.swift) and CoreAIAssistant were archived to ~/Archive/experimental-2026-09-18/ on 2026-09-18 (restore-only). Orient with a path table when ambiguous, then ask which lane if still unclear.
  • Cell / cell-lang / build cel / build cell lang means ~/dev/active/cell-lang (Zig master; public donaldfilimon/cell-lang; the old SSD tree no longer exists). Gate: ./tools/check.sh with -Dswift=false (runs well past 120 s); zig build test alone is weak evidence. Bare zig build run exits 1 by design (CLI needs a subcommand).
  • /functions means Browserbase Functions at ~/Archive/experimental-2026-09-18/AbbeyBot/functions (archived with AbbeyBot on 2026-09-18; restore-only) (bun dev / deploy; credentials in that project's .env). Do not re-scaffold unless asked.
  • /model-compression-exploration or /model-authoring: follow the attached CoreAI skill. Need torch + coreai-opt, plus model/data/forward/quality (and NE vs GPU for authoring). Surface missing env; do not invent a parallel compression path.
  • When /parallel-setup or Parallel skills are attached: use parallel-cli (install via Parallel install script or pipx). Auth via parallel-cli login or PARALLEL_API_KEY. Prefer Parallel web search/extract/research tools over inventing a second research path.

References

  • AGENTS.md (canonical ABI agent instructions)
  • tasks/todo.md (active board)
  • tasks/goals.md (coarse goal ledger)
  • .agents/skills/abi-doc-claims-sync/SKILL.md
  • .agents/skills/sync-clis/SKILL.md (or home sync-clis skill)
  • /abi-skills and self-improving-codebase-loop (improve-all loop)
  • Cursor built-ins: review-bugbot, review-security, check-work
reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis.Stars: 471kUpdated Claude CodeCodexVerifiedInstall
diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness.Stars: 322kUpdated Claude CodeCodexVerifiedInstall
dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependenciesStars: 322kUpdated Claude CodeCodexVerifiedInstall
executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is availableStars: 322kUpdated Claude CodeCodexVerifiedInstall