Build
A curated collection of autonomous AI agent skills powered by Vercel's `skills` npm cli package. Specialized tools for productivity, document management, and custom LLM workflows.
npx -y skills add arendon1/agent-skills --skill buildAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
2 things to look at
- no licenseNo license file was found in the repository. Code published without one is not open source by default, so using it at work is a question for whoever answers licensing questions where you are.
- 0 stars0 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
What its author says it does
Copied from the file, not written here
Execute a plan against the active plan folder. Native single-thread loop: per task, flip status, implement via TDD, verify, commit. Auto-invokes tdd, verify, lessons, git, and dispatch on failure or when parallelizable sets appear. Use when the user asks to build, implement, or tackle a task, says "build §T.3", "run the build", or "do the next task".
SKILL.md
5.8 KB, as published. Nobody here has run it
build — execute the plan
Single-thread native plan-then-execute. No swarm. You are the main agent; you may
dispatch parallel work to subagents only when the plan flags an independent set
(via the dispatch discipline).
OWNERSHIP
Owns: PLAN.md task status cells (flips [ ] -> ~ -> [x]) — and nothing else
in the plan artifacts.
Reads: PLAN.md (required), PRD.md/ARD.md/SPEC.md (whatever exists),
CONTEXT.md, the codebase.
MUST NOT rewrite PRD.md, ARD.md, SPEC.md, LESSONS.md, or PLAN.md task
content. Only status cells.
WHEN (user-invoked)
- User says "build", "implement", "execute the plan", "run the build".
- User names a task: "build §T.3", "do T2".
- User says "do the next task", "--next", "implement next".
STEP 0 — LOAD
- Resolve the active plan per §6. If none, tell the user to run
grill+planfirst. Stop. - Read
PLAN.md. If absent, tell the user to runplan. Stop. - Read whatever else exists:
PRD.md,ARD.md,SPEC.md,CONTEXT.md,§R(research) if present — external facts the build MUST honor. - Parse the invocation:
§T.n/T<n>→ that task only.--next/ "next" → lowest-numbered task with status[ ]or~.--all/ empty → every[ ]task in§Torder.
Right-size gate
High blast radius (shared module, auth, data, money, public interface)? Run the
review loop first. Trivial and reversible? Skip planning ceremony, go to EXECUTE.
STEP 1 — PLAN (per chosen task)
Native plan mode — the agent's own task breakdown, not a new artifact. For each chosen task:
- Cite every
§Vinvariant that applies. The plan MUST respect all. - Cite every
§Iinterface touched. The plan MUST preserve shape. - List files to create / edit.
- Verification contract — name the EXACT test(s) that prove each
§Vtouched. Which test, not "add tests". Each§Vtouched -> a named test that fails first. (This is what thetdddiscipline will write.) - Name the verification command (test, build, lint) — the external oracle. Green = done; NEVER "looks done".
Show the plan. Wait for user OK unless the user said "auto" / "go".
Independent-set detection
If the chosen tasks include a parallelizable set (flagged in PLAN.md §N), and
the work would benefit from isolation, invoke the dispatch discipline to farm
the independent tasks to subagents. Otherwise, single-thread.
STEP 2 — EXECUTE (per task, in dependency order)
For each task in order:
- Flip
§T.nstatus cell[ ]->~. Write only that cell. - Invoke the
tdddiscipline: write the failing test named in the verification contract, watch it fail, write minimal code to pass, watch it pass. - Run the verification command (the oracle).
- Pass → flip
~->[x]. Commit (invoke thegitdiscipline:T<n>: <goal line>+§Vcites). Next task. - Fail → invoke the
debugdiscipline, then thelessonsreflex. Do NOT retry blindly. See FAIL below.
STEP 3 — FAIL → DEBUG → LESSONS
On test/build failure:
- Invoke the
debugdiscipline — build a tight feedback loop, reproduce, hypothesize, instrument, fix. Do not guess. - When the root cause is found, invoke the
lessonsreflex: ask "would a new§Vinvariant catch this class of bug?"- If yes →
lessonsappends the§Brow + the new§VtoLESSONS.md/SPEC.md, and writes the regression test. Then resumebuildagainst the updated spec. - If no (one-time typo, external dep) →
lessonsstill appends the§Brow.
- If yes →
Rule: NEVER silently fix root-cause without considering lessons. LESSONS.md
is the memory that stops recurrence (see the lessons skill).
STEP 4 — VERIFY (before declaring done)
A task is [x] only if (invoke the verify discipline):
- The verification command (the oracle) exits 0, run fresh in this session.
- Every
§Vtouched has its named test from the verification contract, and it passes. - No
§Vinvariant regressed (run the full test suite at the end of the task, and again at the end of the whole plan).
STEP 5 — COMPLETE
After all chosen tasks are [x]:
- Run the full test suite + lint + build. All green.
- Invoke the
verifydiscipline on the whole plan. - Invoke the
gitdiscipline: present options (merge, PR, keep branch). - Tell the user: plan complete. Optionally archive the plan folder to
docs/plans/archive/.
WRITE POLICY
- Only flip
§Tstatus cells inPLAN.md. No other artifact edits frombuild. - Other spec edits → invoke
spec. Other invariant additions →lessons. - Commit after each
§Tcompletes. Message:T<n>: <goal line>+§Vcites.
NON-GOALS
- No speculative work beyond chosen task scope.
- No progress dashboards.
grep '|~|' PLAN.mdis the dashboard. - No silent fixes. Every failure routes through
debug+lessons. - No completion claims without fresh verification evidence (the
verifyIron Law).
BOUNDARIES
- MUST NOT rewrite
PRD.md,ARD.md,SPEC.md,LESSONS.md, orPLAN.mdtask content. Status cells only. - MUST NOT declare a task
[x]without running the verification command fresh. - MUST NOT retry a failure blindly. Route through
debugthenlessons. - MUST commit after each task (via the
gitdiscipline). - MUST use
cavemanfor any artifact touches andCONTEXT.mdcanonical terms.