Goal Formatter

Goal Formatter

When to use

Invoke this skill whenever you want to convert a verbal description into a structured, testable goal that any model or agent can execute without your context.

Trigger phrases:

• "Write me a goal"
• "Create a goal for X"
• "Generate a goal"
• "Frame this as a goal"
• "Set a goal for [task]"

Best for:

• Multi-step tasks with unclear scope
• Work that needs both execution clarity and test coverage planning
• Decisions about what "done" means before starting

What the skill does

1. Analyze repo state

• Read git history (recent commits, branches, tags)
• Parse memory files (MEMORY.md, project memory)
• Check current code state (test pass rate, build status)
• Surface technical debt and blockers from context

2. Synthesize a testable goal

The skill generates a draft goal in this structure (it is not presented yet — step 3 reviews it first):

## Goal
[Clear, measurable end state]

## Success Criteria
- Test coverage: [high-level percentage paths]
- Build status: [specific condition]
- Artifact validation: [what must be true]
- Constraints: [what must NOT change]

## Verification Method
[How the executing agent will prove it's done — measurable, observable]

## Reference Points
[Known-good examples, related prior work, baseline state]

3. Self-review the draft (MANDATORY — do not skip)

Before presenting the goal, review your own draft against the rubric below and revise. A goal is an instruction set another agent (or model) will execute without your context — a vague or untestable goal silently fails downstream. This gate catches that.

This is a rigid step: always run it, even when the draft "looks fine."

1. Hold the drafted goal as draft.
2. Score it against every item in the Self-Review Rubric (below). For each FAIL, note the specific defect.
3. If any rubric item FAILs → revise draft to fix it, then re-score. Max 2 revision rounds.
4. After 2 rounds, if an item still can't pass, present the goal anyway but flag the residual gap explicitly to the user (e.g. "⚠️ Verification for criterion X is manual-only — no automated check exists"). Never silently ship a known-weak goal.
5. Only present the goal once it passes the rubric (or carries explicit flagged residuals).

Do this review in the same turn, inline — it does not require a subagent. For a large or high-stakes goal, you MAY dispatch a separate critic/verifier subagent for an independent pass, but the inline rubric check is always required.

4. Apply model-agnostic framing principles

The goal is written so any model or agent can execute it without your context. Apply these universal prompting principles to the goal text:

• Front-load the critical constraint. State the single most important "do this / don't do that" at the very top, before the detail.
• Use imperative, unambiguous language. MUST / MUST NOT / ONLY — not "should" or "try to."
• Decompose multi-step work into ordered, independent sub-steps with explicit handoffs and dependencies. Name what can run in parallel vs. what is strictly sequential.
• State the reasoning shape when it matters. For analysis-heavy work, say "analyze X and report findings before changing anything"; for mechanical work, keep it terse.
• Make every instruction self-contained. Assume zero conversation history — spell out file paths, IDs, and baselines inline.

These principles are model-independent: they make a goal robust whether it's executed by a frontier model, a smaller local model, or an autonomous agent loop. If you ever need to tune for a specific model's quirks, do that as a thin layer on top — but the core goal stays portable.

Spec mode (OPTIONAL — additive; default behavior is unchanged)

By default this skill emits only the goal. In spec mode, it additionally emits a draft engineering specification alongside the goal, as the first step of a longer spec-authoring chain:

goal-formatter (spec mode)  →  goal-spec  →  spec-review
 (goal + DRAFT spec)           (full spec     (autonomous
                                via Q&A)        hardening)

When to enter spec mode:

• The user explicitly asks for a spec: "write me a goal and a spec", "goal plus spec", "spec this out", "I need the detailed design too".
• The request is clearly a build/design effort (new feature, system, component) where a goal alone won't be enough downstream.

When the ask is a plain "write me a goal" with no spec signal, stay in default mode — emit the goal only. Do not force a spec on a user who didn't ask for one.

What spec mode adds: after producing and self-reviewing the goal as normal, also produce a draft of the 8-section engineering spec (below). It is a draft — deliberately incomplete where the verbal description doesn't determine an answer. Mark every such gap with an explicit [GAP: …] placeholder rather than inventing a value; the downstream goal-spec skill resolves gaps by asking the user.

The 8-section engineering spec template (shared verbatim with goal-spec and spec-review so the three hand off cleanly — same section names, no renaming):

# Specification: <title>

## 1. Problem & Context        — what/why + the goal this spec satisfies
## 2. Scope                    — in scope / out of scope (both explicit)
## 3. Requirements             — Functional (FR1…) + Non-functional (NFR1…)
## 4. Constraints              — hard limits bounding the solution
## 5. Design Approach          — components, data flow, key decisions + rationale
## 6. Success Criteria         — observable, binary, mapped to FRs (inherits the goal's)
## 7. Risks & Mitigations      — what could go wrong + planned response
## 8. Open Questions           — every [GAP] collected here for goal-spec to resolve

Spec-mode rules:

• The goal still comes first and still passes the step-3 self-review gate unchanged.
• The draft spec's Success Criteria (§6) must be consistent with the goal's success criteria — they describe the same end state at different granularity.
• Don't over-fill. A good draft spec confidently states what the description determines and honestly marks the rest [GAP]. Resolving gaps is goal-spec's job, not this skill's.
• After presenting, point the user at the next step: "Run goal-spec to complete this draft into a full spec (it'll ask clarifying questions), then spec-review to harden it."

Goal Template

The baseline goal works across all models and agents:

## Goal
[Clear, measurable end state]

## Success Criteria
- Test coverage: [high-level percentage paths]
- Build/validation status: [specific condition]
- Artifact quality: [what must be true]
- Constraints: [what must NOT change]

## Verification Method
[How the goal is proven complete — measurable, observable, checkable]

## Test Plan
[Rough outline of test coverage needed to validate all paths]

## Reference Points
[Known-good examples, baseline state, related prior work]

This template works with any model. The framing principles in step 4 are baked into how the goal is written, so no model-specific rewrite is needed before use.

How to invoke

Say your verbal description naturally:

"I want to close the v1 gap closure work. We've been focusing on v2 for months but there are still 22 expected-failing tests in v1. I know we have some progress on safety negatives but I'm not sure what the actual blocker is."

The skill will:

1. Automatically read:

   • Recent v1/v2 commit history
   • Memory entries on v1 gap closure state
   • Current test status from your repo

2. Generate a goal like:

## Goal
Achieve 100% v1 safety negative pass rate (57/57 tests green) 
while maintaining v2 test suite at ≥90% pass rate.

## Success Criteria
- v1 test suite: 57/57 safety negatives pass
- v2 contract tests: ≥90% pass rate (baseline protection)
- No regression in v1/v2 bridge coverage
- Git history shows clear atomic commits per fix

## Verification
$ pytest qa/v1/ -k safety_negative  # must exit 0
$ pytest qa/v2/ --tb=short         # must show ≥90% pass

## Execution Notes
MUST front-load: "Fix v1 safety negatives. Do NOT touch v2.
Each fix must be atomic and tested separately."
Two-phase: (1) analyze v1 failures and report root causes
BEFORE changing anything; (2) apply fixes with v2 regression guards.
Context to load: [recent commits, current blockers from memory].

## Reference
- Memory: v1_gap_closure_work_package.md (33 items, 6 spikes)
- Baseline: commit abc1234 (88.5% v1 verified)
- Known blocker: BRIDGE-NNN-013 pending restart

Test-based success criteria

Every goal includes test coverage suggestions at a high level:

• Unit path coverage: "All code paths in X must be covered"

  • Example: "safety_negative.py → 5 assertion paths → each must have a test"

• Integration coverage: "End-to-end flows must validate"

  • Example: "v1 bridge accept_action → groovy execution → test assertion"

• Constraint validation: "No regression in [baseline area]"

  • Example: "v2 contract tests must stay ≥90%"

The goal describes the coverage shape, not individual test names. You then write tests to fill that shape.

Self-Review Rubric

Run this against every drafted goal in step 3 before presenting it. Each row is PASS/FAIL. Any FAIL → revise (max 2 rounds), then present with explicit flags for anything still failing.

Revision discipline: fix the specific failing row — don't rewrite wholesale. Re-score only after editing. If two rounds don't clear a row (commonly #2/#3 when no automated check exists, or #6 when the repo can't confirm a fact), present the goal with that row called out as a residual risk so the user decides.

Worked self-review (compressed)

Draft criterion: "Success: the diagrams look good."

• Rubric #1 FAIL (not measurable), #2/#3 FAIL (unverifiable).
• Revision: "Success: each exported diagram passes the visual self-check defect table (no label truncation, no overlapping symbols) — verified by the agent reading the exported JPEG and reporting a clean verdict, max 2 rounds." → #1 ✅ #2 ✅ #3 ✅.

Example flow

You say:

"Tidy up the main branch. We've merged a lot of feature work but docs are out of sync, old feature branches are cluttering things, and I want a clean state."

Skill generates:

## Goal
main branch is in "release-ready" state: docs synchronized, 
no stale branches, all PR checks green, CHANGELOG updated.

## Success Criteria
- docs/ARCHITECTURE.md ↔ actual codebase architecture (verified via code tour)
- Stale branches: 0 (anything not touched in 30d is archived)
- CI all green (no skipped checks)
- CHANGELOG.md ↔ recent commits (no gaps, all merges documented)

## Verification
1. Run code-tour tool against docs → zero architectural drift
2. Run: git branch -v | grep 30-days-old → empty
3. CI dashboard: all checks ✓
4. git log --grep="merge" | diff CHANGELOG.md → zero diff

## Execution Notes
MUST sequence in order: docs-sync → branch-cleanup → verify-CI.
Do NOT merge anything new during cleanup.
Phase 1 (compare docs to code) is reasoning-heavy — analyze and
report drift before editing. Phases 2-3 (stale branches, sign-off
summary) are mechanical.

What you do after

1. Copy the goal into your /goal invocation
2. Write tests to cover the coverage percentages suggested
3. Run the goal in Claude Code (or any model/agent of your choice)
4. Verify that the executing agent achieves the criteria you set

The skill gives you a head start on clarity; the self-review gate (step 3) means the goal you receive has already been checked against the rubric — but you still own final sign-off and testability.

Note on the self-review gate: the rubric pass in step 3 runs before the goal is shown to you. If you see a goal with a "⚠️ residual" flag, that's the gate being honest about a rubric item it couldn't fully satisfy in 2 rounds (usually an unverifiable criterion) — not an oversight. Decide whether that residual is acceptable before running the goal.

Guidance sources

• Claude Code /goal Documentation
• Prompt Engineering Best Practices