Superpowers Ported Skills

URL: https://mkdocs.justinsforge.com/memory/general/reference_superpowers_ported_skills/

Forge already lifted brainstorming + writing-plans into /feature-plan (2026-05-03, see memory/research/superpowers-context-mode-claudemem-vs-forge-stack-2026-05-03.md). This round surveyed the rest of obra/superpowers for what still cleanly complements it, ported the highest-value pieces, and left the rest out on purpose.

Ported

/tdd (atomic)

Purpose: strict red-green-refactor TDD (write the failing test, watch it fail, minimal code to pass, refactor) for a single feature or bugfix in code that already has a real test runner.

Trigger: "TDD this", "write a failing test first", "red-green-refactor", "test-driven".

Source: superpowers/skills/test-driven-development (371 lines). Trimmed to the mechanics (iron law, cycle, verification-mandatory, red flags) and cut the persuasion essays (rationalization tables, "why order matters" prose) the source repo needs for a team enforcing process on each other. Solo builder doesn't need to be argued into it twice.

Adaptation: explicit skip list matching /feature-plan's own carve-outs (one-line fixes, doc/config edits, hotfixes) plus a forge-specific one: automation scripts with no test framework and no plan to add one use /learnskill's Step 7 smoke test instead, which is deliberately lighter (run once, check exit code and idempotency, not test-first). References real forge test surfaces: pytest (scripts/forge_finances_tests.py), vitest/jest (creatortrack).

Path: .claude/skills/tdd/SKILL.md

/execute-plan (orchestrator)

Purpose: executes an approved /feature-plan output by dispatching a fresh Worker subagent per task, reviewing each with the harness's native /code-review, and logging progress to a resumable ledger.

Trigger: "execute the plan", "run the plan", "dispatch workers for this plan", said after /feature-plan produces a plan file.

Source: superpowers/skills/subagent-driven-development (418 lines, the largest skill in the repo). This was the biggest adaptation of the three.

What was cut, and why: the source skill reimplements its own review loop (task-reviewer-prompt.md, scripts/review-package, a full spec-compliance-vs-quality rubric) because superpowers assumes no native review tool exists on the host. Forge's harness already ships /code-review, so /execute-plan calls that instead of duplicating a second reviewer. Same for the source's verification-before-completion dependency: /execute-plan's final pass hands off to native /verify rather than re-specifying a verification checklist. The source's using-git-worktrees and finishing-a-development-branch dependencies were dropped too: the harness's native EnterWorktree/ExitWorktree tools and subagent isolation: worktree cover parallel worktrees (forge_worktree.sh retired 2026-07-09), and Justin's direct-commit workflow on this repo doesn't run a PR-merge-menu ceremony per plan.

What was kept: the load-bearing insight is fresh-context-per-task plus a mandatory review gate plus a durable ledger (comms/results/plan-ledgers/<slug>.md) so a compacted or resumed session doesn't silently re-dispatch completed tasks. Dispatch uses forge's existing worker infrastructure directly: the Task tool for same-session subagents, or /spawn-headless-agent for long-running or compaction-surviving tasks, both already following the .claude/agents/worker.md result-file contract. No new dispatch mechanism was invented.

Path: .claude/skills/execute-plan/SKILL.md

/debug-systematic (atomic)

Purpose: four-phase root-cause-first debugging (investigate, pattern-match against a working reference, hypothesize and test minimally, implement) for any bug, test failure, or unexpected behavior, before a fix is proposed.

Trigger: "why is this broken", "debug this", "investigate the bug", "root cause", "this isn't working".

Source: superpowers/skills/systematic-debugging (296 lines). Trimmed the rationalization table and "your human partner's signals" section (written for a live pairing session with constant partner interjection, forge sessions are mostly autonomous).

Why it was added even though it wasn't named in the initial ask: it's the one gap that showed up while surveying the repo that cleanly complements /feature-plan (which plans new work) and /tdd (which builds it test-first) without overlapping native /code-review (reviews a diff already written) or native /verify (confirms a fix already works). Forge doctrine already says "robust over quick, no band-aids" (Section 9.5) but doesn't give an enforcement mechanism for the investigation step; this skill is that mechanism, same role the HARD-GATE plays for /feature-plan.

Adaptation: ties Phase 4's "create a failing reproduction" step to /tdd's RED step when a test framework exists, hands off the post-fix confirmation to native /verify instead of re-specifying one, and points repeated (3+) failed hypotheses at LESSONS.md per doctrine Section 10 rather than inventing a separate escalation path.

Path: .claude/skills/debug-systematic/SKILL.md

Deliberately skipped

Source skill	Why skipped
brainstorming, writing-plans	Already ported into /feature-plan (2026-05-03). Re-porting would duplicate.
receiving-code-review, requesting-code-review	The harness's native /code-review already covers this. Porting would build a second, competing review skill.
verification-before-completion	The harness's native /verify already covers this.
using-git-worktrees	Native EnterWorktree/ExitWorktree tools and subagent isolation: worktree cover this (forge_worktree.sh retired 2026-07-09). Same job, no gap to fill.
finishing-a-development-branch	Built for a PR-merge-vs-discard menu ceremony per feature branch. Forge's central repo workflow is direct commits with an explicit "create a new commit, don't amend" rule (CLAUDE.md); the source skill's 4-option menu doesn't map onto how Justin actually closes out work here. Revisit if forge moves to a PR-per-feature model.
dispatching-parallel-agents	Generic parallel-subagent-dispatch pattern. Its one clear use case (N independent bug/test failures dispatched concurrently) is already covered by /debug-systematic (single bug, sequential phases) plus ad hoc parallel Task tool calls, which this session's own tool-use conventions already support natively. Not enough net-new value to justify a fourth skill.
using-superpowers	Meta-skill for navigating the superpowers plugin's own skill library (a discovery/dispatcher skill for skills). Forge's skill discovery is ~/.claude/skills/ plus this reference doc; no meta-navigation layer needed for three skills.
writing-skills	Meta-skill for authoring superpowers-style skills. Forge already has this convention documented inline in CLAUDE.md (Section 3 naming, atomic vs orchestrator classification) and demonstrated by /learnskill. A second, competing "how to write a skill" doc would fragment the convention.

How the three chain together

/feature-plan          -> writes memory/plans/<slug>-YYYY-MM-DD.md, HARD-GATE on Justin approval
  /tdd                  -> optional, invoked by a task brief when the task touches tested code
  /debug-systematic     -> invoked ad hoc, any time, independent of a plan
/execute-plan           -> consumes the plan, dispatches Worker subagents per task
  native /code-review   -> per-task and final quality gate (not reimplemented)
  native /verify        -> end-to-end confirmation after the final pass (not reimplemented)

/tdd and /debug-systematic are atomic (one job each, callable standalone or from inside a dispatched task). /execute-plan is an orchestrator (chains /feature-plan's output, worker dispatch, and native review/verify; no logic duplicated from any of them). Per the atomic-vs-orchestrator rule, none of the three reimplement what another forge tool already does.

Related

• /feature-plan, upstream planning skill these three extend
• Skill systems pattern, atomic vs orchestrator classification rule
• Robust over quick, the doctrine /debug-systematic and /tdd enforce
• .claude/agents/worker.md, the result-file contract /execute-plan dispatches against
• Native worktree tools (EnterWorktree/ExitWorktree), why using-git-worktrees wasn't ported