Codifying Tasks

Codifying Tasks

Any task done twice will be done a third time. The second occurrence is the signal to codify it: the procedure is proven, the gotchas are known, and the cost of writing it down is lower than the cost of re-deriving it.

When to fire

• The task you just completed resembles one done earlier in this project (similar prompt, similar steps, similar output)
• The user signals it: "make this repeatable", "turn this into a skill", "we'll do this again", "same as last time"
• A multi-step workflow just succeeded after trial and error — the corrected path is fresh and worth preserving

When NOT to fire

• First occurrence of a task — wait for the second; premature codification bakes in guesses
• One-off work that won't recur (a migration, a single deliverable)
• The task is already covered by an existing skill — improve that skill instead (see the correcting-mistakes skill)
• The "procedure" is a single prompt with no steps, decisions, or gotchas — a skill adds nothing over just asking again
• The procedure carries secrets or one-account specifics (API keys, tokens, internal hostnames, a particular customer's data) — either skip it, or codify only the generalized shape with those values as arguments; never bake a secret into a saved skill

Workflow

Step 1: Offer, Briefly

One sentence, not a ceremony: "This is the second time we've done X — want me to codify it as a skill so next time it's one command?" If the user declines, drop it without argument.

Step 2: Extract the Procedure from the Session

Reconstruct what actually worked — not an idealized version:

• The steps in the order they actually succeeded
• Decisions made along the way and what drove them
• Gotchas hit and the working resolution (state the correct approach directly; don't narrate the failure)
• Validation used to confirm the output was right
• Inputs that varied between the two occurrences — these become the skill's arguments
• Any secrets or one-account specifics — redact these from the procedure now; they become arguments, never literals

Step 3: Decide Scope

Pick where the skill lives before authoring, so the file is written in the right place:

• Task specific to this project → .claude/skills//
• Task that recurs across the user's projects → ~/.claude/skills//

When unsure, default to the project; promotion to user scope is easy later.

Step 4: Author via creating-skills

Invoke the creating-skills skill (/compound:creating-skills) with the extracted procedure as raw material and the Step 3 scope/save-location. It handles naming, frontmatter, structure, and quality validation. Keep the scoping dialog short — most answers (procedure, arguments, save location) are already known from the session.

Step 5: Verify and Hand Off

1. Tell the user the invocation (/) and what arguments it takes
2. A brand-new skill may not appear until the session reloads — if it doesn't show in the skills list on the next message, tell the user to restart the session (or run /reload-plugins for a plugin skill). Don't report failure on the strength of it not showing immediately.
3. Suggest running it on the next real occurrence rather than a synthetic test — and fixing anything it gets wrong via the correcting-mistakes skill

Principles

• Codify the proven path. The skill records what worked, not what should have worked.
• Gotchas become instructions. "Use the v2 endpoint" — not "Warning: the v1 endpoint fails."
• Modular over monolithic. If the session contained two separable procedures, write two skills.
• The loop closes. correcting-mistakes maintains what codifying-tasks creates. Together they compound.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: Sawyer-Middeleer
• Source: Sawyer-Middeleer/dot-claude
• License: MIT

Install and usage instructions live in the source repository linked above.