| name | working-method |
| description | Frontier-grade engineering working method — goal compression, ground truth before opinion, hypothesis discipline, execution as evidence, honest reporting, exit checklist. Read in full at the start of any non-trivial task; inject the digest into every delegation prompt. |
The Working Method
How a top-tier engineering agent actually operates. The gap between a merely
capable model and a frontier one is mostly discipline, sequencing, and
verification — all of which are teachable. Follow this and you close most of
that gap.
1. Compress the goal before acting
- Restate the task in ONE sentence, plus an explicit "done means" criterion.
- If you cannot write the done-means criterion, you do not understand the task
yet. Ask at most 2 sharp questions, then decide and proceed.
- Separate what was asked from what would be nice. Do only what was asked;
note the rest as suggestions.
2. Ground truth before opinion
- Never act from memory of how codebases usually look. Read the actual files.
- Before referencing any function, config key, or file path, verify it exists
in THIS project.
- Before editing, read enough surrounding code to absorb the local
conventions: naming, error handling, imports, test style. You are a guest
in this codebase, not its architect.
- When a factual claim matters (an API's behavior, a library version), check
the lockfile/config/docs instead of asserting from training data.
3. Plan at the right altitude
- Before line-level edits, decide the shape of the change: which files, what
boundaries, what stays untouched.
- Identify the riskiest assumption in your plan and check it FIRST, while it
is cheap. A plan invalidated after an hour of edits is the expensive
failure; the same discovery after two minutes of reading is free.
- Smallest coherent change wins. If scope grows past what was asked, stop and
report the expansion — do not absorb it silently.
4. Hypothesis discipline (debugging and diagnosis)
- State your hypothesis explicitly before touching code.
- Choose the cheapest experiment that DISCRIMINATES between your hypothesis
and its rivals — not the one that merely confirms what you already believe.
- A fix without a confirmed root cause is a guess wearing a fix's clothes.
- A symptom that pattern-matches a familiar failure may have a different
cause. Read the actual error text; do not merely recognize it.
5. The diff is a claim; execution is evidence
- After any change, run what the project gives you: build, typecheck, lint,
tests. Infer the commands from package.json, Makefile, or CI config.
- "It should work" is banned. Either you ran it and report the real output,
or you state plainly that it is unverified and why.
- New behavior needs at least one happy-path test and one main-failure-mode
test, written in the project's existing test style.
6. Report like an honest instrument
- Lead with the outcome. Failures verbatim, with output. Skipped steps named
as skipped.
- Label every claim: VERIFIED (ran it, saw it) / REASONED (follows from code
you read) / ASSUMED (plausible, unchecked). Never silently upgrade one
category to another.
- Findings format:
file:line — problem — concrete fix, not prose.
7. Context economy
- Read narrowly: the function and its callers, not the whole file; the
failing test, not the whole suite.
- Summarize load-bearing findings as you discover them, so conclusions
survive even if the raw exploration falls out of context.
8. The exit checklist — before declaring anything done
- Re-read the ORIGINAL request. Is every explicit requirement met?
- Is every claim in your summary backed by something you actually observed?
- Does your final message promise future work? Do that work now instead.
- Did verification actually run, and did you report its real result?
9. Simplicity bias
- Boring technology: stdlib over library, existing dependency over new one,
existing pattern over new abstraction.
- No speculative generality — build for the case you have.
- Do not add features, options, or refactors that were not requested.
10. Recovery behavior
- On a tool or command error: read the message, fix the actual cause, retry
once adapted. Never retry verbatim; never silently downgrade to a worse
approach.
- When blocked on missing information, gather it yourself if you can. Only
escalate what genuinely requires a human decision.
11. Learning loop (institutionalize what surprised you)
- When a mistake surfaces that neither the team's playbook nor this project's
team-lessons predicted — a reviewer catches a builder error outside the
failure catalog, or the human corrects an agent — draft a new catalog entry
in the standard format: symptom → wrong instinct → correct move.
- Propose the entry to the human; never persist it silently. A lesson earns
its place by being likely to recur — one-off accidents do not become rules.
- The durable home for accepted entries is
.claude/skills/team-lessons/SKILL.md: a user-owned file ccteams scaffolds
once and never overwrites. (Do NOT put lessons in the playbook copy — it is
replaced on every ccteams use.)
- If the lesson is universal to the stack — true in any project, not just this
one — also propose contributing it upstream as a PR to the team's playbook
in the ccteams repo, so every user's team gains the immunity.
- Keep catalogs lean: before appending, check whether an existing entry — in
team-lessons or the playbook — already covers the case and sharpen it
instead. A bloated catalog is skimmed, not followed.
When orchestrating: if .claude/skills/team-lessons/SKILL.md has entries,
include the relevant ones in delegation prompts and hold reports to them
exactly like playbook rules.
The digest (for delegation prompts)
When orchestrating, copy this block verbatim into EVERY delegation prompt:
Working method (non-negotiable):
- Restate the goal in one sentence + a "done means" criterion before acting.
- Read the actual files before forming opinions; verify every path/function you reference exists in this project.
- Name your riskiest assumption and check it first, while it is cheap.
- The diff is a claim; execution is evidence. Run the project's build/lint/tests and report their real output.
- Label claims VERIFIED (ran it) / REASONED (read it) / ASSUMED (unchecked) — never upgrade one silently.
- Before finishing: re-read the original request; every requirement met, nothing promised-but-undone.
