| name | th-projects |
| description | Use for project-level engineering governance in any repository — bootstrap or audit the knowledge architecture (doctrine, specs, topology, standards), concretize a feature request into a spec delta, amend a spec found wrong mid-implementation, reconcile specs against implementation, run a repo-wide health audit, decide what to work on next via spec-driven planning, milestone synthesis, and doctrine-grounded feature ideation, or run a recurring vision-pursuit audit that generates the next best moves toward the project's ideal. Route to exactly one subskill per task. Triggers: "project shape", "bootstrap docs", "what should we work on next", "should we build this", "I want to add X", "spec this feature", "the spec is wrong", "break this down", "review this project", "audit the codebase", "does the code match the spec", "what's the next milestone", "brainstorm features", "run the vision pursuit", "relentlessly improve this project".
|
| metadata | {"owner":"tze","authors":["tze","Claude Fable 5","OpenAI Codex"],"status":"active","last_reviewed":"2026-07-27"} |
| compatibility | Subskill scripts require bash, git, grep, find. project-direction additionally assumes the bd (beads) CLI and an OpenSpec-capable environment for changeset synthesis. |
TH Projects
Superskill router for spec-driven project governance. Five subskills under
subskills/; each a complete standard package (own SKILL.md, references/,
scripts/). Subskills not in global catalog — discover lazily here, load
at most one subskill body per task. A subskill may route you to a sibling
(e.g. project-review runs project-shape's scanner); follow that link from the
subskill, not here.
Five subskills, one lifecycle:
- Shape — set normative baseline: what project believes, how designed,
what must be built, where things live, engineering bar for changing it.
- Feature request — run one concrete proposal through idea funnel against
baseline → signed-off spec delta.
- Review — audit implementation vs baseline + generic health criteria
(incl. exhaustive spec↔code reconciliation) → confirmed findings.
- Direction — baseline + spec deltas + findings → prioritized, spec-linked
work plan; hand execution to beads.
- Pursuit — recurring generative loop against the baseline's ideal:
fan-out surface audits + ideation lenses → ranked moves, tier-board
movement, gated work plan; pursued moves re-enter via feature request.
VISION is a continuous constraint, not a one-time gate: proposal, spec,
allocation, implementation discovery, reconciliation, and milestone closeout
must still serve cited doctrine. Implementation discoveries re-enter at the
earliest affected feature-request gate; milestone closeout returns uncovered
mandates to direction's milestone synthesis. The lifecycle stops only at a
hard human gate or when the approved spec is achieved and verified.
Discover subskills
PKG="$(dirname "$SKILL_PATH")"
find "$PKG/subskills" -maxdepth 2 -name SKILL.md
rg -n "^name:|^description:" "$PKG"/subskills/*/SKILL.md
Routing table
| Task intent | Subskill | Typical trigger |
|---|
| Bootstrap or audit the project's knowledge architecture (five pillars: heart-and-soul, legends-and-lore, openspec, lay-and-land, craft-and-care); decide where an idea should be documented; generate a layman overview. | subskills/project-shape/SKILL.md | "set up project structure", "bootstrap docs", "where should this go", "audit documentation health" |
| Concretize ONE fuzzy feature/project request into a signed-off spec delta: motif, doctrine gate, topology placement, design sketch, WHEN/THEN scenarios. Also amendment mode: fix a spec found wrong/ambiguous mid-implementation. | subskills/project-feature-request/SKILL.md | "I want to add X", "spec this feature", "turn this idea into requirements", "what would it take to build X", "the spec is wrong" |
| Decide what to work on next; evaluate competing priorities; check roadmap alignment; turn approved specs into a prioritized beads work plan. Also the vision-generative modes when nothing is proposed: milestone synthesis (derive work doctrine already implies) and ideation (brainstorm new feature candidates grounded in doctrine). | subskills/project-direction/SKILL.md | "what's highest leverage", "what should we work on next", "break this down", "is this roadmap aligned", "what's the next milestone", "brainstorm features for this project" |
| Repo-wide health audit: code quality, reliability, security, docs, maintainability — scored, evidence-based, with a planning handoff packet. Includes the exhaustive spec-reconciliation mode (bidirectional spec↔code gap audit + remediation). | subskills/project-review/SKILL.md | "review this project", "audit the codebase", "assess project health", "reconcile spec vs implementation", "what's implemented but undocumented" |
| Recurring generative pursuit of the project's ideal: fan-out audits of every surface against the applicable bar + vision-grounded ideation lenses → ranked move list, tier-board movement across runs, gated (never auto-released) work plan. | subskills/relentless-vision-pursuit/SKILL.md | "run the vision pursuit", "what's the next best step toward the vision", "relentlessly improve this project", "audit the whole project against the ideal" |
Routing rules
- Baseline before judgment: review and direction both consume the shape
baseline. No shape artifacts at all + real ask is "make this project legible"
→ project-shape, even if user said "review".
- Audit vs. plan vs. pursue: review classifies (scores, risks, confirmed
findings); direction decides (sequencing, specs, beads); pursuit
generates (gap-to-ideal moves, recurring, orchestrated). "What's wrong
with this repo" → review. "What to do about it / next" → direction.
"Relentlessly push this toward its vision" → pursuit. Full review hands off
to direction inside the subskills.
- Feature vs. direction: one concrete proposal → feature-request; many
competing priorities or "what next" → direction; no proposal at all and the
ask is to invent candidates → direction (ideation), whose pursued picks
feed the funnel. A feature request surviving its funnel hands its spec delta
to direction for sequencing.
- Scope guard: single-PR/diff review →
/code-review, not here.
Change-level engineering-quality judgment (engineering bar, readability, test
rigor, dependency hygiene, cruft cleanup, skill reviews, diagrams) →
/th-engineering (project-shape's craft-and-care adopts that bar by
reference). Backlog mechanics without direction analysis →
/beads-orchestration (beads-writer). AI-tooling harness hygiene (installed
skills, dotfiles, snapshot state) → /th-tooling.
- Design and engineering bars: user-facing behavior or interaction contracts
→
/th-design; implementation quality, tests, dependencies, diagnosis, docs,
or diagrams → /th-engineering. Load the router only when the concern exists,
then its one relevant subskill; carry conclusions into the spec or bead rather
than copying whole skill bodies.
- Fallback: project-adjacent but no row fits → answer from router context or
ask. Do not load a subskill to browse.
Shared invariants (all subskills)
-
Five-pillar shape model (about/heart-and-soul, about/legends-and-lore,
openspec/, about/lay-and-land, about/craft-and-care) is the single
normative vocabulary; subskills must not redefine it.
-
Specs are source of truth for work planning; implementation without spec
coverage is a finding, not a baseline. The OpenSpec file format subskills
read/write is one shared contract,
references/spec-format.md at package root, not
owned by any single subskill. Its mechanical validator,
scripts/spec-trace-check.py, runs before any
subagent spends tokens on semantic spec review.
-
Per-change spec sync: an implementation task that changes observable behavior
amends its governing spec in the same task (feature-request amendment mode).
Episodic reconciliation is the backstop, never the mechanism.
-
VISION/doctrine remains binding after sign-off. Record the baseline commit in
spec handoffs; if it changed, revalidate affected mandates before allocation
or resumption.
-
Gap, TODO, unknown, and adjacent idea discovery is proactive but never silent
scope expansion. Classify it as current-spec correction, boundary/design
change, evidence unknown, new scope, or local debt; route to the earliest
governing gate and record it durably.
-
Work graphs follow
references/work-allocation.md: one bead and
primary agent per cohesive independently verifiable outcome, with enough work
to amortize context/worktree/CI/review overhead and no overlapping ownership.
-
Autonomy contract — what agents may do without the human:
| Artifact | Agent may | Human must |
|---|
Doctrine (about/heart-and-soul/) | draft, review, blast-radius sweep | adopt every amendment |
| Spec delta (proposed behavior) | draft + reconcile changeset | sign off before implementation |
| Main-spec bookkeeping of [Observed] behavior | edit directly | — |
| Beads/planning graph | generate from approved changeset | approve the changeset it derives from |
-
Every major claim cites evidence, labeled [Observed], [Inferred], or
[Unknown].
Package maintenance
After changing any subskill, script, or fixture in this package, run
scripts/validate-th-projects.sh — shell
syntax, project-shape self-tests, fixture invariants, and the spec-trace
fixtures — and fix every FAIL before committing.