hacking-workflow
Skill1 starsChanged 3 months agoMITView on GitHub
Coordinates a non-trivial git-hosted change from intent through verified integration. Use at the START when the user says "let's build X", "hack on X", "start a hacking session", "run the loop", "new project", or "implement this". Not for a quick direct artifact — that's builder
In this repository23 other files it gives its agents.
- AGENTS.mdAGENTS.md
- CLAUDE.mdCLAUDE.md
- .cursor/rules/beads.mdcCursor rule
- beadsSkill
- builderSkill
- consulting-codexSkill
- dispatching-subagentsSkill
- evidence-before-claimsSkill
- finishingSkill
- naming-as-designSkill
- planningSkill
- reading-like-proseSkill
- reviewing-workSkill
- shapingSkill
- spike-and-stabilizeSkill
- subtraction-firstSkill
- test-firstSkill
- using-beadsSkill
- using-hackinatorSkill
- using-worktreesSkill
- walking-skeletonSkill
- whittlerSkill
- writing-skillsSkill
The file
We split this skill into the rules it gives an agent
1 rule. About 0 are common, 1 is distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈1
Distinctive · estimate
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 14
This is the front door: When a build starts, run the loop.
≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Hacking Workflow
Overview
Core principle: Real work runs a loop with decision and evidence gates. Resolve choices that materially affect the result, then keep moving until fresh evidence supports the outcome. This skill is the spine; each phase is its own skill it hands off to.
This is the front door. When a build starts, run the loop. Right-size the ceremony to the work (a spike skips most of it), the host's capabilities, and the active authority boundaries.
Skills guide technique; they do not grant permission. System, user, host, and repository policy decide which actions and capabilities are available.
Classify First (right-size the loop)
Classify the path and state it briefly when that helps the user follow the work. Complexity can only upgrade the path, never downgrade it.
| Path | What it is | Loop |
| Spike | "Can this even be done?" throwaway probe | spike-and-stabilize only, report back. Skip the rest. |
| Bounded | A clear, contained change | Shape briefly → build → verify → finish. Skip the plan doc. |
| Architectural | New system, many parts, real design choices | The full loop below. |
The Loop
Track the whole thing in beads from the first phase (using-beads) — file the work, close as you go.
Phase 0 — SHAPE resolve unclear intent and consequential design choices
→ shaping when needed. GATE: no unresolved choice that materially changes the result.
Phase 1 — ISOLATE protect concurrent or risky edits when the host permits it
→ using-worktrees when useful. Otherwise work serially in the current workspace.
GATE: understand the baseline and preserve unrelated changes.
Phase 2 — PLAN (architectural only) map files + right-sized tasks, file as beads
→ planning. GATE: plan filed; pick a permitted execution mode.
Phase 3 — BUILD make it work, thinnest slice first
→ builder (walking-skeleton, spike-and-stabilize). Keeper code is test-first → test-first.
Phase 4 — CRAFT make it elegant, behavior-preserving
→ whittler.
Phase 5 — VERIFY prove it actually works, with fresh evidence
→ evidence-before-claims. GATE: fresh command output, scoped to the change's blast radius —
never a reflex full suite, never just the new test.
Phase 6 — REVIEW an independent pass before it lands
→ reviewing-work via the best available independent mechanism; use an adversarial self-review
when none exists. GATE: findings resolved by severity.
Phase 7 — FINISH follow the authorized integration choice; clean up
→ finishing. GATE: the branch's landing gate green on the final tree; request a choice only
when none was supplied.Build and craft interleave per slice; verify and review gate each meaningful chunk, not just the end.
Verification breadth escalates along the loop: a slice proves its own blast radius (Phase 5); the branch proves the union plus its seams (Phase 7). Re-running everything at every slice is not rigor — it's a tax that makes nobody read the output.
Checkpoint Green Slices When Authorized
GREEN EVIDENCE = SAFE CHECKPOINT OPPORTUNITYThe moment a slice verifies — fresh output in hand (evidence-before-claims) — checkpoint it if the active profile or user authorizes commits. A commit is not a publication, but this workflow is not standing authority to create one. Without commit authority, preserve the working tree and report the verified boundary clearly.
| The excuse | The answer |
| "I'll commit when it's all done" | If commits are authorized, a green checkpoint is safer and easier to review. |
| "It's not clean enough to commit" | Clean is Phase 4's job. An authorized working checkpoint can still be the block whittling carves from. |
| "It'll get squashed anyway" | Fine — squashing later is trivial. Recovering lost uncommitted work isn't. |
Decision Boundaries
Proceed autonomously on reversible, in-scope work when the result is clear. An explicit request to implement a clear, bounded change supplies intent; do not ask the user to approve it again.
Pause when:
- A missing product or design choice would materially change the result.
- An action is destructive, irreversible, security-sensitive, externally visible, or incurs real spend and has not already been authorized.
- The next requested integration, commit, sync, or push action lacks authority or conflicts with active policy.
Everywhere else, decide and record the ruling (in a bead) rather than stalling — momentum over paralysis.
Common Mistakes
- Re-asking after a clear implementation request. Resolve real ambiguity; don't manufacture it.
- Forcing isolation the host forbids. Preserve unrelated changes and work serially in place.
- Claiming done from Phase 3. "It compiles" is not "it works." Phase 5 exists for a reason.
- Defaulting every gate to the full suite. Scope Phase 5 to what the slice could break; save the branch-wide gate for Phase 7. Escalate on ambiguity, not on habit.
- Treating the workflow as commit authority. Checkpoint green slices only when active policy or the user permits commits.
- Running the full loop for a spike. Ceremony must match the work. Classify first.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
Reviews can't be read right now.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

