agentic-coding-methodology-composition

Agentic Coding Methodology Composition

Compare, evaluate, and compose AI coding agent methodologies and skills frameworks. Class of work: "should we use methodology X?" / "how do X and Y compare?" / "can Z and W coexist?" / "why is my agent misbehaving in a way that maps to a methodology design?".

When to load this skill

• User asks about a specific AI coding methodology (Superpowers, GSD Core, grill-me, OpenSpec, BMAD, Spec-Kit, Kiro, gstack, GSTACK, custom /skill packs).
• User asks "should we install X?" / "is X worth it?" / "how do X and Y differ?".
• User asks for a combo analysis ("can we run X with Y?" / "what's the best of both worlds?").
• User asks about an emerging methodology and wants to know how it relates to existing ones.
• User reports an agent failure that maps to a methodology design (e.g., "it skipped TDD" → check Superpowers' HARD-GATE; "context bloat after 2 hours" → check GSD Core's context-engineering rationale).

Methodology landscape (snapshot 2026-07-30)

Methodology	Author	Canonical source	Local status
Superpowers	Jesse Vincent (obra)	https://github.com/obra/superpowers	Installed v6.2.0
GSD Core	TÂCHES	https://github.com/open-gsd/gsd-core	Installed v1.9.0 (--profile=full) — see --profile pitfall
grill-me / grilling	Matt Pocock	https://github.com/mattpocock/skills	Installed (21 Matt Pocock skills at ~/.claude/skills/{grill-me,...})
OpenSpec	Fission-AI	https://github.com/Fission-AI/OpenSpec	Not installed
BMAD Method	bmadcode	https://github.com/bmadcode/BMAD-METHOD	Not installed
Spec-Kit	GitHub	https://github.com/github/spec-kit	Not installed
Kiro Specs	AWS	https://kiro.dev	Not installed
GSTACK	garrytan	https://github.com/garrytan/gstack	Not installed

(See references/ for per-methodology profiles with verified installation paths, distinctive features, and source-of-truth URLs.)

Workflow — methodology-composition analysis

Phase 0: clarify scope

Single-methodology or comparison? If comparison, name candidates or "compare all major frameworks" (≥3).

Phase 1: canonical-source verification

For each methodology:

1. Find the canonical repo — GitHub org, official docs site, npm package, plugin marketplace. Reject mirrors/forks unless explicit.
2. Verify recency — confirm the canonical repo is still maintained. Watch for "this repo has moved" redirects (e.g., gsd-build/get-shit-done archived → open-gsd/gsd-core).
3. Verify local install — see references/installation-check.md.

Phase 2: profile each methodology

Use references/methodology-profile-template.md to capture:

• Author + canonical source
• One-line positioning
• Pipeline / loop / workflow
• Distinctive features (what it does that others don't)
• Invocation model (auto-fire vs opt-in)
• Artifact tree (where state lives)
• Multi-harness support
• Known anti-patterns

Phase 3: composition analysis

Identify:

• Overlap zones (where two methodologies do similar things — pick one as authoritative).
• Complement zones (where one extends the other).
• Conflict zones (where one explicitly fights the other, e.g., Superpowers' HARD-GATE vs opt-in frameworks).
• Outer-loop / inner-loop splits (GSD Core for milestone continuity, Superpowers for per-task TDD).

Output: a composition-recommendation matrix with concrete install paths.

Phase 4: anti-pattern audit

Check against references/composition-patterns.md for known anti-patterns. Flag any that apply to the user's setup.

Phase 5: deliver

Slack-native format per SOUL.md colored-icons-in-status-reports:

• ? Healthy (sources confirmed, install status)
• ? What each does that others don't
• ? How they compose
• ? Risky / anti-patterns
• ? Recommended combo
• ? Next actions

Apply LLM-provenance caveat footer per SOUL.md llm-provenance-caveat. Hyperlink every URL per no-trailing-asteri[REDACTED_OPENAI_KEY] and pr-hyperlink rules.

Phase 5.5: execute the install (don't post the list and stop)

Load-bearing rule, verified 2026-07-30. When the recommended-combo or Next-actions section enumerates installable commands (npx @opengsd/gsd-core@latest --claude --global, npx skills@latest add mattpocock/skills --skill=grill-me -y -g, etc.) AND the user said anything like "install them", "set it up", "go ahead", "ship it", "let's do X" — execute the install in the same turn, do not post the list and wait. The user will read the install summary in the next reply, not a static plan.

If the user asked for analysis only ("compare X and Y", "how do they differ"), do not install — analysis-only is a valid request.

If the user gave a guarded install directive ("install them all but don't add default guidance"), execute the install but honor the guard:

• Default-install flags that auto-enable hooks → downgrade to lean profile (GSD: --profile=core not full; grill-me keeps disable-model-invocation: true).
• Default guidance files → don't touch (SOUL.md, CLAUDE.md, plugin enable flags).
• Verify post-install what got auto-configured and call it out honestly in the next reply. Don't hide it.

Anti-patterns (load-bearing)

1. Don't recommend installing all global frameworks. Superpowers' HARD-GATE will fight GSD's /gsd:* namespace and grill-me's disable-model-invocation: true. Pick one as the per-session default; bring others in explicitly per-phase.
2. Don't accept GSD's --profile=full default when the user said "no default guidance." Full installs ~15 always-on hooks to settings.json (verified v1.9.0). Use --profile=core (~700 tokens eager-load) instead. See references/gsd-core.md for the full hook table.
3. Don't duplicate state trees. GSD → .planning/. Superpowers → docs/superpowers/{specs,plans}/. grill-me → nothing or CONTEXT.md. Designate one artifact tree as authoritative.
4. Don't use a pure interview primitive (grill-me) inside an execution loop. Grill-me is for upstream alignment only.
5. Don't recommend a framework whose canonical repo has moved without flagging the redirect. Always verify the live repo.
6. Don't treat star counts as signal of correctness. GSD's 48.3k stars ≠ better than Superpowers' smaller community. They solve different problems.
7. Don't claim a methodology is "installed" without probing the filesystem. See references/installation-check.md.
8. Don't create status-cron jobs that ping the user about a choice they're still making. When the user says "install X but don't add default guidance," a 20m cron pinging them about "did you pick a combo?" violates the directive. Wait for the user's own next message; don't schedule a nudge.

Support files

• references/methodology-profile-template.md — capture template for one methodology
• references/superpowers.md — Superpowers profile (obra, installed v6.2.0)
• references/gsd-core.md — GSD Core profile (TÂCHES, not installed)
• references/grill-me.md — grill-me profile (Matt Pocock, not installed)
• references/composition-patterns.md — verified combinations with anti-patterns
• references/installation-check.md — how to verify local install status per methodology

How to extend this skill

When a new methodology emerges:

1. Verify the canonical source (live repo, official docs).
2. Add references/<name>.md profile using the template.
3. Update composition-patterns.md if it composes with existing methodologies.
4. Update the landscape table in this SKILL.md.

When a methodology moves (e.g., gsd-build/get-shit-done → open-gsd/gsd-core):

1. Update the profile's canonical source.
2. Note the redirect history in the profile.
3. Grep existing references for stale URLs.