AI Development Methodology

E09 — External landscape pass over six repos (skills-spec conformance + item-level clarification marker)

Pillar (primary): P4 — Tool compatibility Pillar (secondary): P1 — Doc completeness Status: done (2026-08-19) Phase: Phase 1 — Foundation Started: 2026-08-19 (chartered at maintainer direction) Closed: 2026-08-19 — all three items shipped in v1.31.0 Owner: maintainer + AI coding agent

Frozen intent — Outcome and exit criteria approved by maintainer on 2026-08-19. Agents do not edit them; halt and renegotiate instead.

Outcome (jobs-to-be-done)

When the maintainer evaluates a batch of external projects, they want the pass to end with the smallest set of changes that survive scrutiny — so that the review is not itself a reason to add conventions.

Six repos were reviewed on 2026-08-19 at maintainer request: addyosmani/agent-engineer, agentskills/agentskills, obra/superpowers, NousResearch/hermes-agent, msitarzewski/agency-agents, github/spec-kit.

Most of the value was already extracted by earlier passes. The CHANGELOG records superpowers, Spec Kit and Hermes Agent each being mined before: v1.20.0 memory archival (Hermes), v1.21.0 EARS (Spec Kit/Kiro), v1.22.0 constitution check (Spec Kit) plus two-stage review and rule pressure-testing (superpowers/BMAD). Four of the six yield nothing new.

Two things do, and they are deliberately asymmetric in kind:

1. A compatibility signal, not a rule. The Agent Skills format is an open specification. skills/ai-dev-methodology/SKILL.md already conforms to it; nothing says so. Declaring it costs one sentence and serves the pillar that is actually binding.
2. One convention that closes a hole the last pass named and did not fix. BL-0033 (v1.30.0) identified the unattended-mode failure — an agent answering its own clarifying questions — and gave it a caution but no mechanism. Spec Kit’s inline clarification marker is that mechanism, and it fits an item-level gap that already existed: epic charters have an “Open questions” section with a closing gate (03); items have no equivalent.

This epic adds one named convention. That is the point of it, not a limitation of it.

Exit criteria (binary)

• Each of BL-0034 through BL-0036 is resolved — shipped, deferred, or rejected with reasoning.
• The conformance claim was verified before it was written — the normative spec read, the frontmatter checked field by field — and the check is recorded, not asserted.
• The conformance wording is a format claim, not peer positioning (see “Out of scope”).
• The clarification marker reuses the existing frozen-intent marker shape rather than inventing a second one, and is distinguished in one sentence from Principle 1 in 06.
• Every change is markdown-only — no scripts, no runtime dependencies.
• Repo-wide rendering-link check clean; every touched methodology doc under the 1,050-line cap. Counts taken after the final edit, per the v1.30.1 correction.

KPIs

• 04_backlog_items.md grows by ≤ 25 lines (it is the longest doc at 1,020 of a 1,050 cap).
• Zero new named concepts beyond the one marker.
• Conformance declared in exactly two places (skill + README); no third copy to drift.

Out of scope

Rejected — persona agents (msitarzewski/agency-agents). Fourth consideration, fourth decline. 230+ agent personalities across 15 divisions is the most complete version of the idea we have seen, and it does not change the reasoning from E06, E07 and E08: token cost per turn, coupling to persona-capable tooling, and the fact that our challenge-prompt, cross-AI and decision-ownership mechanisms already cover the goal. templates/ROLE_BRIEFS.md is the deliberate alternative — a brief describes a stance for a phase, a persona describes a character for an agent. Recorded here so the fifth pass does not re-litigate it.

Rejected — bundling references/ into the skill (agentskills/agentskills). The spec’s progressive-disclosure model would let the skill carry the 14 methodology docs and work offline. Declined by maintainer decision on 2026-08-19: it duplicates ~17,000 lines into a second location and creates a permanent sync burden against P3. Link-out stays the single source of truth. The symptom that would justify revisiting is concrete: an adopter reporting that an agent without network access got the operating contract and none of the depth.

No adoption — NousResearch/hermes-agent. Its curator memory pattern was already imported in v1.20.0 (archive-don’t-destroy). The remainder — messaging gateways, provider adapters, terminal backends, session search, cron scheduling — is runtime-framework machinery with no markdown-methodology analogue. Its agent-authored executable skills conflict directly with the no-code stance.

No adoption — addyosmani/agent-engineer. A course on building agents (Vertex AI, ADK, MCP, A2A). This methodology governs projects that use agents — a different axis, not a competing answer on the same one. Its authoring principles (“link, don’t duplicate”; “honest about trade-offs”) are already our doc altitude, the memory admission test, and STATUS.md.

Deferred, not rejected — two ideas, recorded in FUTURE.md: superpowers’ baseline-before-rule sequencing, and Spec Kit’s cross-artifact coverage check.

Linked docs

• Pillars: P4, P1, P5
• Parent thread: E08’s BL-0033 — named the failure this epic gives a mechanism to.
• Precedent: CHANGELOG v1.25.0 (E06), v1.29.0 (E07), v1.30.0 (E08) — three prior landscape passes.
• Sources: the six repos named above, all public. Full internal comparison is held local-only in self-development/brief/03_competitive_landscape.md (gitignored by policy — see .gitignore).

Item roster

See ARCHIVE.md — all three closed. BACKLOG.md is empty.

Risks

• Convention fatigue, fourth pass. v1.25.0 added five conventions, v1.29.0 five, v1.30.0 a template plus two sharpenings — with zero external adopters exercising any of them, a fact EPICS.md already names against itself. This epic’s mitigation is structural: it ships one convention, and its primary item is not a convention at all but a compatibility signal aimed at the constraint that is actually binding (adopters, not rules). If the pass had produced five more rules, chartering it would have been the wrong call.
• A conformance claim that stops being true. The spec can change; our skill can drift. The claim names the format, not a version, and the frontmatter check is cheap enough to repeat — the semi-annual currency pass (P3, next due 2026-11-25) is the natural place. Recorded in the item’s resolution so the check is reproducible rather than remembered.
• A marker that becomes a to-do list. In Spec Kit the marker sits in a generated spec that a human then answers. Ours sits in a hand-authored item, where the risk is that it accumulates instead of blocking. Mitigated by binding it to the ready gate: an item carrying an unresolved marker cannot be picked up, so the marker has teeth or it has none.

Closing note

One exit criterion is met by waiver rather than by passing. Cross-AI findings-verification was waived by maintainer decision on 2026-08-19; the criterion above covers the mechanical checks, which did run and are reproducible. The three landscape passes before this one each cleared that gate. Recorded in ARCHIVE.md and in the v1.31.0 CHANGELOG entry rather than folded into a checkbox, so the next evaluation can weigh it.

KPIs met. 04_backlog_items.md grew 16 lines against a ≤25 budget, landing at 1,036 of a 1,050 cap. One new named concept, as chartered. Conformance declared in exactly two places.

This site is open source. Improve this page.