Absolute Spec Skill for Claude

> Start your first response with the 📋 emoji.

Absolute Spec

Absolute Spec turns a fuzzy intent into a reviewed design-spec document — then

stops. It is the planning artifact of work, lifted out of the full lifecycle: scan

the codebase, ask only the questions the code can't answer, write the spec, run an

independent scored review, and hand back a doc you can discuss, hand off, or feed into

/absolute work to build.

It is deliberately lightweight. No relentless interview, no task board, no

decompose, no execution. One deliverable: a spec good enough that an unfamiliar

developer could build from it.

When to use this command

Use spec when:

• You want a design spec to discuss, review, or hand off — not build right now.
• You need to think a feature through on paper before committing to implementation.
• A teammate or another agent will do the building from your spec.
• You want a fast, reviewed design doc without the full phase-gated lifecycle.

Do NOT use spec when:

• You want to design AND build in one flow → use /absolute work (spec is its

Phase 2, followed by decompose + safe-wave execution).

• You're documenting code that already exists / already shipped → use

/absolute docs (reference/explanation docs, not a forward design).

• The change is a one-line obvious fix needing no design.

The line vs work: **spec produces a doc and stops; work produces a doc and then

builds it.** If you're unsure whether you'll build it now, start with spec — you can

chain into work afterward and it will pick the spec up.

Key Principles

• Codebase before questions. Read what exists first; ask only what code can't answer.
• Bounded, not relentless. A short clarify pass (3–5 questions), not the depth-first

grill work runs. Batching questions is fine.

• The spec is the deliverable. Quality matters more here than anywhere — keep the

independent scored review.

• Reuse, don't reinvent. Template, scaling rules, and review rubric come from

absolute-work's references/spec-writing.md. This command is the thin flow around them.

• Stop after review. No code, no board, no execution. Hand off cleanly.
• Never auto-commit. Write the spec file and report; the user commits.

The Flow

SCAN ─→ CLARIFY ─→ WRITE ─→ REVIEW ─→ HANDOFF

No hard gates between steps (that's work). The one checkpoint is the clarify pass —

ask, get answers, then proceed straight through to a reviewed spec.

Step 1 — SCAN

Ground the spec in reality before asking anything.

• Convention detection — read cached config first: if .absolute.config.json or

~/.absolute/config.json exists (from /absolute init), resolve the effective config

(project file → global projects["<cwd>"] → global defaults) and use its conventions,

detecting only what's missing. No config → soft-suggest init and run work's Codebase

Convention Detection table (package manager, language/runtime, test runner, linter, build

system, directory conventions; see absolute-work's SKILL.md under *Codebase Convention

Detection*). The spec must speak the project's real stack, scripts, and paths.

• Deep context scan — read what exists: docs/ (README first), root README.md,

CLAUDE.md, CONTRIBUTING.md, docs/plans/ (overlapping/related designs), recent

commits (last 10–20), package manifests, and the directories the feature touches.

• Synthesize — state what you learned in 2–4 lines (stack, relevant existing code,

any overlapping prior design). Do not dump a file listing.

Detect the work type (feature / refactor / greenfield / migration) only enough to

shape the spec — spec does not run the full per-type question banks.

Step 2 — CLARIFY (bounded pass)

Ask only the questions the codebase genuinely cannot answer — the preference and

scope forks. Cap at 3–5 questions; batching is allowed (one AskUserQuestion call

with multiple questions where available).

Good questions to ask (when code can't answer them):

• Scope boundary — what's in v1 vs explicitly deferred.
• Behavior forks — real-time vs batch, sync vs async, the genuine "which way" decisions.
• Data/contract shape — only where the design hinges on it and code gives no signal.
• Non-goals — what this explicitly will NOT do.

Do NOT ask: anything a config, manifest, test file, or existing pattern already states.

When the code answers it, say so in the spec instead of asking.

If the user said "no questions, just draft it" — skip to WRITE and capture every guess

in an ## Open Questions / assumptions section instead.

Step 3 — WRITE

Write the spec to <specDir>/YYYY-MM-DD-<topic>-design.md (<topic> = short kebab-case

slug). <specDir> is preferences.specDir from config when present, else docs/plans.

Use the template, section-scaling rules, writing style, and Decision Log format

from absolute-work's references/spec-writing.md — that file is the single source of truth;

do not restate it here, load it. In particular:

• Pick the complexity tier (Simple / Medium / Complex) using the complexity

heuristic table, and scale sections accordingly. Remove sections that would only say

"N/A".

• Be concrete: real file paths, real endpoints, real schemas in code blocks — never

"an endpoint for X".

• Fill the Decision Log with the forks resolved in CLARIFY plus any you recommended,

each with a one-line rationale.

Step 4 — REVIEW (scored, independent)

Dispatch a separate reviewer subagent — the agent that wrote the spec does not grade

it (generator ≠ evaluator). Use the rubric and reviewer prompt template verbatim from

absolute-work's references/spec-writing.md (*Scored Spec Review Protocol*): graded on

Completeness, Consistency, Clarity, Scope, Testability (1–5 each, weighted).

| Weighted Score | Verdict | Action |

|---|---|---|

| 4.0 – 5.0 | Approved | Proceed to HANDOFF |

| 3.0 – 3.9 | Needs Work | Fix the flagged issues, re-dispatch (max 3 iterations) |

| < 3.0 | Major Gaps | Surface to the user immediately; do not silently iterate |

Reviewer approval is necessary but not the end — the user still has the final say at

HANDOFF.

Step 5 — HANDOFF

• Present — the spec path, the complexity tier, and the reviewer's weighted score +

verdict. Summarize the key decisions in 2–4 lines.

• Offer to build — chain into /absolute work, which will pick up the spec at

its decompose phase rather than re-running intake. Make this an explicit, optional

next step.

• Stop. No code, no board. Never run git commit — suggest a message if asked;

the user commits.

Gotchas

• Drifting into work. If you start decomposing into tasks or writing code, you've

left spec. Stop at the reviewed doc.

• Asking what the code answers. The bounded pass is for preferences and forks only —

scan first, ask second.

• Skipping the review to "stay light". The review is the cheap quality gate that

makes a standalone spec trustworthy. Keep it.

• Restating spec-writing.md. Load it; don't copy the template/rubric in here —

two copies drift.

• Over-asking. More than ~5 questions means you're running work's interview. Tighten.

Output / Response Style

Respond terse like smart caveman. All technical substance stay. Only fluff die. Drop

articles, filler, pleasantries, hedging. Fragments OK. Technical terms exact. Code blocks

and quoted errors unchanged. Drop caveman for security warnings, irreversible-action

confirmations, and any multi-step sequence where fragment order risks misread; resume

after. Spec prose itself is written normally (it's a deliverable, not chat).

Companion commands

Sibling commands in this skill chain naturally around spec:

• /absolute work — build what the spec describes (picks the spec up at decompose).
• /absolute ui — design the interface for a feature the spec defines.
• /absolute docs — once it's built and shipped, document it for readers.

Suggest them where relevant; they're always available (same skill, no extra install).